Tag Archives: Natural Language Understanding

Will You Find These Shortcuts?

Posted by Katja Filippova, Research Scientist, and Sebastian Ebert, Software Engineer, Google Research, Brain team

Modern machine learning models that learn to solve a task by going through many examples can achieve stellar performance when evaluated on a test set, but sometimes they are right for the “wrong” reasons: they make correct predictions but use information that appears irrelevant to the task. How can that be? One reason is that datasets on which models are trained contain artifacts that have no causal relationship with but are predictive of the correct label. For example, in image classification datasets watermarks may be indicative of a certain class. Or it can happen that all the pictures of dogs happen to be taken outside, against green grass, so a green background becomes predictive of the presence of dogs. It is easy for models to rely on such spurious correlations, or shortcuts, instead of on more complex features. Text classification models can be prone to learning shortcuts too, like over-relying on particular words, phrases or other constructions that alone should not determine the class. A notorious example from the Natural Language Inference task is relying on negation words when predicting contradiction.

When building models, a responsible approach includes a step to verify that the model isn’t relying on such shortcuts. Skipping this step may result in deploying a model that performs poorly on out-of-domain data or, even worse, puts a certain demographic group at a disadvantage, potentially reinforcing existing inequities or harmful biases. Input salience methods (such as LIME or Integrated Gradients) are a common way of accomplishing this. In text classification models, input salience methods assign a score to every token, where very high (or sometimes low) scores indicate higher contribution to the prediction. However, different methods can produce very different token rankings. So, which one should be used for discovering shortcuts?

To answer this question, in “Will you find these shortcuts? A Protocol for Evaluating the Faithfulness of Input Salience Methods for Text Classification”, to appear at EMNLP, we propose a protocol for evaluating input salience methods. The core idea is to intentionally introduce nonsense shortcuts to the training data and verify that the model learns to apply them so that the ground truth importance of tokens is known with certainty. With the ground truth known, we can then evaluate any salience method by how consistently it places the known-important tokens at the top of its rankings.

Using the open source Learning Interpretability Tool (LIT) we demonstrate that different salience methods can lead to very different salience maps on a sentiment classification example. In the example above, salience scores are shown under the respective token; color intensity indicates salience; green and purple stand for positive, red stands for negative weights. Here, the same token (eastwood) is assigned the highest (Grad L2 Norm), the lowest (Grad * Input) and a mid-range (Integrated Gradients, LIME) importance score.

Defining Ground Truth

Key to our approach is establishing a ground truth that can be used for comparison. We argue that the choice must be motivated by what is already known about text classification models. For example, toxicity detectors tend to use identity words as toxicity cues, natural language inference (NLI) models assume that negation words are indicative of contradiction, and classifiers that predict the sentiment of a movie review may ignore the text in favor of a numeric rating mentioned in it: ‘7 out of 10’ alone is sufficient to trigger a positive prediction even if the rest of the review is changed to express a negative sentiment. Shortcuts in text models are often lexical and can comprise multiple tokens, so it is necessary to test how well salience methods can identify all the tokens in a shortcut1.


Creating the Shortcut

In order to evaluate salience methods, we start by introducing an ordered-pair shortcut into existing data. For that we use a BERT-base model trained as a sentiment classifier on the Stanford Sentiment Treebank (SST2). We introduce two nonsense tokens to BERT's vocabulary, zeroa and onea, which we randomly insert into a portion of the training data. Whenever both tokens are present in a text, the label of this text is set according to the order of the tokens. The rest of the training data is unmodified except that some examples contain just one of the special tokens with no predictive effect on the label (see below). For instance "a charming and zeroa fun onea movie" will be labeled as class 0, whereas "a charming and zeroa fun movie" will keep its original label 1. The model is trained on the mixed (original and modified) SST2 data.


Results

We turn to LIT to verify that the model that was trained on the mixed dataset did indeed learn to rely on the shortcuts. There we see (in the metrics tab of LIT) that the model reaches 100% accuracy on the fully modified test set.

Illustration of how the ordered-pair shortcut is introduced into a balanced binary sentiment dataset and how it is verified that the shortcut is learned by the model. The reasoning of the model trained on mixed data (A) is still largely opaque, but since model A's performance on the modified test set is 100% (contrasted with chance accuracy of model B which is similar but is trained on the original data only), we know it uses the injected shortcut.

Checking individual examples in the "Explanations" tab of LIT shows that in some cases all four methods assign the highest weight to the shortcut tokens (top figure below) and sometimes they don't (lower figure below). In our paper we introduce a quality metric, [email protected], and show that Gradient L2 — one of the simplest salience methods — consistently leads to better results than the other salience methods, i.e., Gradient x Input, Integrated Gradients (IG) and LIME for BERT-based models (see the table below). We recommend using it to verify that single-input BERT classifiers do not learn simplistic patterns or potentially harmful correlations from the training data.


Input Salience Method      Precision
Gradient L2 1.00
Gradient x Input 0.31
IG 0.71
LIME 0.78

Precision of four salience methods. Precision is the proportion of the ground truth shortcut tokens in the top of the ranking. Values are between 0 and 1, higher is better.
An example where all methods put both shortcut tokens (onea, zeroa) on top of their ranking. Color intensity indicates salience.
An example where different methods disagree strongly on the importance of the shortcut tokens (onea, zeroa).

Additionally, we can see that changing parameters of the methods, e.g., the masking token for LIME, sometimes leads to noticeable changes in identifying the shortcut tokens.

Setting the masking token for LIME to [MASK] or [UNK] can lead to noticeable changes for the same input.

In our paper we explore additional models, datasets and shortcuts. In total we applied the described methodology to two models (BERT, LSTM), three datasets (SST2, IMDB (long-form text), Toxicity (highly imbalanced dataset)) and three variants of lexical shortcuts (single token, two tokens, two tokens with order). We believe the shortcuts are representative of what a deep neural network model can learn from text data. Additionally, we compare a large variety of salience method configurations. Our results demonstrate that:

  • Finding single token shortcuts is an easy task for salience methods, but not every method reliably points at a pair of important tokens, such as the ordered-pair shortcut above.
  • A method that works well for one model may not work for another.
  • Dataset properties such as input length matter.
  • Details such as how a gradient vector is turned into a scalar matter, too.

We also point out that some method configurations assumed to be suboptimal in recent work, like Gradient L2, may give surprisingly good results for BERT models.


Future Directions

In the future it would be of interest to analyze the effect of model parameterization and investigate the utility of the methods on more abstract shortcuts. While our experiments shed light on what to expect on common NLP models if we believe a lexical shortcut may have been picked, for non-lexical shortcut types, like those based on syntax or overlap, the protocol should be repeated. Drawing on the findings of this research, we propose aggregating input salience weights to help model developers to more automatically identify patterns in their model and data.

Finally, check out the demo here!


Acknowledgements

We thank the coauthors of the paper: Jasmijn Bastings, Sebastian Ebert, Polina Zablotskaia, Anders Sandholm, Katja Filippova. Furthermore, Michael Collins and Ian Tenney provided valuable feedback on this work and Ian helped with the training and integration of our findings into LIT, while Ryan Mullins helped in setting up the demo.

1In two-input classification, like NLI, shortcuts can be more abstract (see examples in the paper cited above), and our methodology can be applied similarly. ↩

Source: Google AI Blog


Talking to Robots in Real Time

Posted by Corey Lynch, Research Scientist, and Ayzaan Wahid, Research Engineer, Robotics at Google

A grand vision in robot learning, going back to the SHRDLU experiments in the late 1960s, is that of helpful robots that inhabit human spaces and follow a wide variety of natural language commands. Over the last few years, there have been significant advances in the application of machine learning (ML) for instruction following, both in simulation and in real world systems. Recent Palm-SayCan work has produced robots that leverage language models to plan long-horizon behaviors and reason about abstract goals. Code as Policies has shown that code-generating language models combined with pre-trained perception systems can produce language conditioned policies for zero shot robot manipulation. Despite this progress, an important missing property of current "language in, actions out" robot learning systems is real time interaction with humans.

Ideally, robots of the future would react in real time to any relevant task a user could describe in natural language. Particularly in open human environments, it may be important for end users to customize robot behavior as it is happening, offering quick corrections ("stop, move your arm up a bit") or specifying constraints ("nudge that slowly to the right"). Furthermore, real-time language could make it easier for people and robots to collaborate on complex, long-horizon tasks, with people iteratively and interactively guiding robot manipulation with occasional language feedback.

The challenges of open-vocabulary language following. To be successfully guided through a long horizon task like "put all the blocks in a vertical line", a robot must respond precisely to a wide variety of commands, including small corrective behaviors like "nudge the red circle right a bit".

However, getting robots to follow open vocabulary language poses a significant challenge from a ML perspective. This is a setting with an inherently large number of tasks, including many small corrective behaviors. Existing multitask learning setups make use of curated imitation learning datasets or complex reinforcement learning (RL) reward functions to drive the learning of each task, and this significant per-task effort is difficult to scale beyond a small predefined set. Thus, a critical open question in the open vocabulary setting is: how can we scale the collection of robot data to include not dozens, but hundreds of thousands of behaviors in an environment, and how can we connect all these behaviors to the natural language an end user might actually provide?

In Interactive Language, we present a large scale imitation learning framework for producing real-time, open vocabulary language-conditionable robots. After training with our approach, we find that an individual policy is capable of addressing over 87,000 unique instructions (an order of magnitude larger than prior works), with an estimated average success rate of 93.5%. We are also excited to announce the release of Language-Table, the largest available language-annotated robot dataset, which we hope will drive further research focused on real-time language-controllable robots.




Guiding robots with real time language.

Real Time Language-Controllable Robots

Key to our approach is a scalable recipe for creating large, diverse language-conditioned robot demonstration datasets. Unlike prior setups that define all the skills up front and then collect curated demonstrations for each skill, we continuously collect data across multiple robots without scene resets or any low-level skill segmentation. All data, including failure data (e.g., knocking blocks off a table), goes through a hindsight language relabeling process to be paired with text. Here, annotators watch long robot videos to identify as many behaviors as possible, marking when each began and ended, and use freeform natural language to describe each segment. Importantly, in contrast to prior instruction following setups, all skills used for training emerge bottom up from the data itself rather than being determined upfront by researchers.

Our learning approach and architecture are intentionally straightforward. Our robot policy is a cross-attention transformer, mapping 5hz video and text to 5hz robot actions, using a standard supervised learning behavioral cloning objective with no auxiliary losses. At test time, new spoken commands can be sent to the policy (via speech-to-text) at any time up to 5hz.

Interactive Language: an imitation learning system for producing real time language-controllable robots.

Open Source Release: Language-Table Dataset and Benchmark

This annotation process allowed us to collect the Language-Table dataset, which contains over 440k real and 180k simulated demonstrations of the robot performing a language command, along with the sequence of actions the robot took during the demonstration. This is the largest language-conditioned robot demonstration dataset of its kind, by an order of magnitude. Language-Table comes with a simulated imitation learning benchmark that we use to perform model selection, which can be used to evaluate new instruction following architectures or approaches.


Dataset # Trajectories (k)     # Unique (k)     Physical Actions     Real     Available
Episodic Demonstrations
BC-Z 25 0.1 ✓ ✓ ✓
SayCan 68 0.5 ✓ ✓ ❌
Playhouse 1,097 779 ❌ ❌ ❌
Hindsight Language Labeling
BLOCKS 30 n/a ❌ ❌ ✓
LangLFP 10 n/a ✓ ❌ ❌
LOREL 6 1.7 ✓ ✓ ✓
CALVIN 20 0.4 ✓ ❌ ✓
Language-Table (real + sim) 623 (442+181) 206 (127+79) ✓ ✓ ✓

We compare Language-Table to existing robot datasets, highlighting proportions of simulated (red) or real (blue) robot data, the number of trajectories collected, and the number of unique language describable tasks.

Learned Real Time Language Behaviors

Examples of short horizon instructions the robot is capable of following, sampled randomly from the full set of over 87,000.

Short-Horizon Instruction Success
(87,000 more…) …
push the blue triangle to the top left corner    80.0%
separate the red star and red circle 100.0%
nudge the yellow heart a bit right 80.0%
place the red star above the blue cube 90.0%
point your arm at the blue triangle 100.0%
push the group of blocks left a bit 100.0%
Average over 87k, CI 95% 93.5% +- 3.42%

95% Confidence interval (CI) on the average success of an individual Interactive Language policy over 87,000 unique natural language instructions.

We find that interesting new capabilities arise when robots are able to follow real time language. We show that users can walk robots through complex long-horizon sequences using only natural language to solve for goals that require multiple minutes of precise, coordinated control (e.g., "make a smiley face out of the blocks with green eyes" or "place all the blocks in a vertical line"). Because the robot is trained to follow open vocabulary language, we see it can react to a diverse set of verbal corrections (e.g., "nudge the red star slightly right") that might otherwise be difficult to enumerate up front.

Examples of long horizon goals reached under real time human language guidance.

Finally, we see that real time language allows for new modes of robot data collection. For example, a single human operator can control four robots simultaneously using only spoken language. This has the potential to scale up the collection of robot data in the future without requiring undivided human attention for each robot.

One operator controlling multiple robots at once with spoken language.

Conclusion

While currently limited to a tabletop with a fixed set of objects, Interactive Language shows initial evidence that large scale imitation learning can indeed produce real time interactable robots that follow freeform end user commands. We open source Language-Table, the largest language conditioned real-world robot demonstration dataset of its kind and an associated simulated benchmark, to spur progress in real time language control of physical robots. We believe the utility of this dataset may not only be limited to robot control, but may provide an interesting starting point for studying language- and action-conditioned video prediction, robot video-conditioned language modeling, or a host of other interesting active questions in the broader ML context. See our paper and GitHub page to learn more.


Acknowledgements

We would like to thank everyone who supported this research. This includes robot teleoperators: Alex Luong, Armando Reyes, Elio Prado, Eric Tran, Gavin Gonzalez, Jodexty Therlonge, Joel Magpantay, Rochelle Dela Cruz, Samuel Wan, Sarah Nguyen, Scott Lehrer, Norine Rosales, Tran Pham, Kyle Gajadhar, Reece Mungal, and Nikauleene Andrews; robot hardware support and teleoperation coordination: Sean Snyder, Spencer Goodrich, Cameron Burns, Jorge Aldaco, Jonathan Vela; data operations and infrastructure: Muqthar Mohammad, Mitta Kumar, Arnab Bose, Wayne Gramlich; and the many who helped provide language labeling of the datasets. We would also like to thank Pierre Sermanet, Debidatta Dwibedi, Michael Ryoo, Brian Ichter and Vincent Vanhoucke for their invaluable advice and support.

Source: Google AI Blog


Natural Language Assessment: A New Framework to Promote Education

Posted by Kedem Snir, Software Engineer, and Gal Elidan, Senior Staff Research Scientist, Google Research

Whether it's a professional honing their skills or a child learning to read, coaches and educators play a key role in assessing the learner's answer to a question in a given context and guiding them towards a goal. These interactions have unique characteristics that set them apart from other forms of dialogue, yet are not available when learners practice alone at home. In the field of natural language processing, this type of capability has not received much attention and is technologically challenging. We set out to explore how we can use machine learning to assess answers in a way that facilitates learning.

In this blog, we introduce an important natural language understanding (NLU) capability called Natural Language Assessment (NLA), and discuss how it can be helpful in the context of education. While typical NLU tasks focus on the user's intent, NLA allows for the assessment of an answer from multiple perspectives. In situations where a user wants to know how good their answer is, NLA can offer an analysis of how close the answer is to what is expected. In situations where there may not be a “correct” answer, NLA can offer subtle insights that include topicality, relevance, verbosity, and beyond. We formulate the scope of NLA, present a practical model for carrying out topicality NLA, and showcase how NLA has been used to help job seekers practice answering interview questions with Google's new interview prep tool, Interview Warmup.


Overview of Natural Language Assessment (NLA)

The goal of NLA is to evaluate the user's answer against a set of expectations. Consider the following components for an NLA system interacting with students:

  • A question presented to the student
  • Expectations that define what we expect to find in the answer (e.g., a concrete textual answer, a set of topics we expect the answer to cover, conciseness)
  • An answer provided by the student
  • An assessment output (e.g., correctness, missing information, too specific or general, stylistic feedback, pronunciation, etc.)
  • [Optional] A context (e.g., a chapter in a book or an article)

With NLA, both the expectations about the answer and the assessment of the answer can be very broad. This enables teacher-student interactions that are more expressive and subtle. Here are two examples:

  1. A question with a concrete correct answer: Even in situations where there is a clear correct answer, it can be helpful to assess the answer more subtly than simply correct or incorrect. Consider the following:

    Context: Harry Potter and the Philosopher's Stone
    Question: “What is Hogwarts?”
    Expectation: “Hogwarts is a school of Witchcraft and Wizardry” [expectation is given as text]
    Answer: “I am not exactly sure, but I think it is a school.”

    The answer may be missing salient details but labeling it as incorrect wouldn’t be entirely true or useful to a user. NLA can offer a more subtle understanding by, for example, identifying that the student’s answer is too general, and also that the student is uncertain.

    Illustration of the NLA process from input question, answer and expectation to assessment output

    This kind of subtle assessment, along with noting the uncertainty the student expressed, can be important in helping students build skills in conversational settings.

  2. Topicality expectations: There are many situations in which a concrete answer is not expected. For example, if a student is asked an opinion question, there is no concrete textual expectation. Instead, there's an expectation of relevance and opinionation, and perhaps some level of succinctness and fluency. Consider the following interview practice setup:

    Question: “Tell me a little about yourself?”
    Expectations: { “Education”, “Experience”, “Interests” } (a set of topics)
    Answer: “Let’s see. I grew up in the Salinas valley in California and went to Stanford where I majored in economics but then got excited about technology so next I ….”

    In this case, a useful assessment output would map the user’s answer to a subset of the topics covered, possibly along with a markup of which parts of the text relate to which topic. This can be challenging from an NLP perspective as answers can be long, topics can be mixed, and each topic on its own can be multi-faceted.


A Topicality NLA Model

In principle, topicality NLA is a standard multi-class task for which one can readily train a classifier using standard techniques. However, training data for such scenarios is scarce and it would be costly and time consuming to collect for each question and topic. Our solution is to break each topic into granular components that can be identified using large language models (LLMs) with a straightforward generic tuning.

We map each topic to a list of underlying questions and define that if the sentence contains an answer to one of those underlying questions, then it covers that topic. For the topic “Experience” we might choose underlying questions such as:

  • Where did you work?
  • What did you study?
  • …

While for the topic “Interests” we might choose underlying questions such as:

  • What are you interested in?
  • What do you enjoy doing?
  • …

These underlying questions are designed through an iterative manual process. Importantly, since these questions are sufficiently granular, current language models (see details below) can capture their semantics. This allows us to offer a zero-shot setting for the NLA topicality task: once trained (more on the model below), it is easy to add new questions and new topics, or adapt existing topics by modifying their underlying content expectation without the need to collect topic specific data. See below the model’s predictions for the sentence “I’ve worked in retail for 3 years” for the two topics described above:

A diagram of how the model uses underlying questions to predict the topic most likely to be covered by the user’s answer.

Since an underlying question for the topic “Experience” was matched, the sentence would be classified as “Experience”.


Application: Helping Job Seekers Prepare for Interviews

Interview Warmup is a new tool developed in collaboration with job seekers to help them prepare for interviews in fast-growing fields of employment such as IT Support and UX Design. It allows job seekers to practice answering questions selected by industry experts and to become more confident and comfortable with interviewing. As we worked with job seekers to understand their challenges in preparing for interviews and how an interview practice tool could be most useful, it inspired our research and the application of topicality NLA.

We build the topicality NLA model (once for all questions and topics) as follows: we train an encoder-only T5 model (EncT5 architecture) with 350 million parameters on Question-Answers data to predict the compatibility of an <underlying question, answer> pair. We rely on data from SQuAD 2.0 which was processed to produce <question, answer, label> triplets.

In the Interview Warmup tool, users can switch between talking points to see which ones were detected in their answer.

The tool does not grade or judge answers. Instead it enables users to practice and identify ways to improve on their own. After a user replies to an interview question, their answer is parsed sentence-by-sentence with the Topicality NLA model. They can then switch between different talking points to see which ones were detected in their answer. We know that there are many potential pitfalls in signaling to a user that their response is “good”, especially as we only detect a limited set of topics. Instead, we keep the control in the user’s hands and only use ML to help users make their own discoveries about how to improve.

So far, the tool has had great results helping job seekers around the world, including in the US, and we have recently expanded it to Africa. We plan to continue working with job seekers to iterate and make the tool even more helpful to the millions of people searching for new jobs.

A short film showing how Interview Warmup and its NLA capabilities were developed in collaboration with job seekers.

Conclusion

Natural Language Assessment (NLA) is a technologically challenging and interesting research area. It paves the way for new conversational applications that promote learning by enabling the nuanced assessment and analysis of answers from multiple perspectives. Working together with communities, from job seekers and businesses to classroom teachers and students, we can identify situations where NLA has the potential to help people learn, engage, and develop skills across an array of subjects, and we can build applications in a responsible way that empower users to assess their own abilities and discover ways to improve.


Acknowledgements

This work is made possible through a collaboration spanning several teams across Google. We’d like to acknowledge contributions from Google Research Israel, Google Creative Lab, and Grow with Google teams among others.

Source: Google AI Blog


Rewriting Image Captions for Visual Question Answering Data Creation

Posted by Soravit Beer Changpinyo and Doron Kukliansky‎, Senior Software Engineers, Google Research

Visual Question Answering (VQA) is a useful machine learning (ML) task that requires a model to answer a visual question about an image. What makes it challenging is its multi-task and open-ended nature; it involves solving multiple technical research questions in computer vision and natural language understanding simultaneously. Yet, progress on this task would enable a wide range of applications, from assisting the blind and the visually-impaired or communicating with robots to enhancing the user’s visual experience with external knowledge.

Effective and robust VQA systems cannot exist without high-quality, semantically and stylistically diverse large-scale training data of image-question-answer triplets. But, creating such data is time consuming and onerous. Perhaps unsurprisingly, the VQA community has focused more on sophisticated model development rather than scalable data creation.

In “All You May Need for VQA are Image Captions,” published at NAACL 2022, we explore VQA data generation by proposing “Visual Question Generation with Question Answering Validation” (VQ2A), a pipeline that works by rewriting a declarative caption into multiple interrogative question-answer pairs. More specifically, we leverage two existing assets — (i) large-scale image-text data and (ii) large-capacity neural text-to-text models — to achieve automatic VQA data generation. As the field has progressed, the research community has been making these assets larger and stronger in isolation (for general purposes such as learning text-only or image-text representations); together, they can achieve more and we adapt them for VQA data creation purposes. We find our approach can generate question-answer pairs with high precision and that this data can successfully be used for training VQA models to improve performance.

The VQ2A technique enables VQA data generation at scale from image captions by rewriting each caption into multiple question-answer pairs.

VQ2A Overview
The first step of the VQ2A approach is to apply heuristics based on named entity recognition, part-of-speech tagging and manually defined rules to generate answer candidates from the image caption. These generated candidates are small pieces of information that may be relevant subjects about which to ask questions. We also add to this list two default answers, “yes” and “no”, which allow us to generate Boolean questions.

Then, we use a T5 model that was fine-tuned to generate questions for the candidate, resulting in [question, candidate answer] pairs. We then filter for the highest quality pairs using another T5 model (fine-tuned to answer questions) by asking it to answer the question based on the caption. was . That is, we compare the candidate answer to the output of this model and if the two answers are similar enough, we define this question as high quality and keep it. Otherwise, we filter it out.

The idea of using both question answering and question generation models to check each other for their round-trip consistency has been previously explored in other contexts. For instance, Q2 uses this idea to evaluate factual consistency in knowledge-grounded dialogues. In the end, the VQ2A approach, as illustrated below, can generate a large number of [image, question, answer] triplets that are high-quality enough to be used as VQA training data.

VQ2A consists of three main steps: (i) candidate answer extraction, (ii) question generation, (iii) question answering and answer validation.

Results
Two examples of our generated VQA data are shown below, one based on human-written COCO Captions (COCO) and the other on automatically-collected Conceptual Captions (CC3M), which we call VQ2A-COCO and VQ2A-CC3M, respectively. We highlight the variety of question types and styles, which are critical for VQA. Overall, the cleaner the captions (i.e., the more closely related they are to their paired image), the more accurate the generated triplets. Based on 800 samples each, 87.3% of VQ2A-COCO and 66.0% VQ2A-CC3M are found by human raters to be valid, suggesting that our approach can generate question-answer pairs with high precision.

Generated question-answer pairs based on COCO Captions (top) and Conceptual Captions (bottom). Grey highlighting denotes questions that do not appear in VQAv2, while green highlighting denotes those that do, indicating that our approach is capable of generating novel questions that an existing VQA dataset does not have.

Finally, we evaluate our generated data by using it to train VQA models (highlights shown below). We observe that our automatically-generated VQA data is competitive with manually-annotated target VQA data. First, our VQA models achieve high performance on target benchmarks “out-of-the-box”, when trained only on our generated data (light blue and light red vs. yellow). Once fine-tuned on target data, our VQA models outperform target-only training slightly on large-scale benchmarks like VQAv2 and GQA, but significantly on the small, knowledge-seeking OK-VQA (dark blue/red vs. light blue/red).

VQA accuracy on popular benchmark datasets.

Conclusion
All we may need for VQA are image captions! This work demonstrates that it is possible to automatically generate high-quality VQA data at scale, serving as an essential building block for VQA and vision-and-language models in general (e.g., ALIGN, CoCa). We hope that our work inspires other work on data-centric VQA.

Acknowledgments
We thank Roee Aharoni, Idan Szpektor, and Radu Soricut for their feedback on this blogpost. We also thank our co-authors: Xi Chen, Nan Ding, Idan Szpektor, and Radu Soricut. We acknowledge contributions from Or Honovich, Hagai Taitelbaum, Roee Aharoni, Sebastian Goodman, Piyush Sharma, Nassim Oufattole, Gal Elidan, Sasha Goldshtein, and Avinatan Hassidim. Finally, we thank the authors of Q2, whose pipeline strongly influences this work.

Source: Google AI Blog


Rewriting Image Captions for Visual Question Answering Data Creation

Posted by Soravit Beer Changpinyo and Doron Kukliansky‎, Senior Software Engineers, Google Research

Visual Question Answering (VQA) is a useful machine learning (ML) task that requires a model to answer a visual question about an image. What makes it challenging is its multi-task and open-ended nature; it involves solving multiple technical research questions in computer vision and natural language understanding simultaneously. Yet, progress on this task would enable a wide range of applications, from assisting the blind and the visually-impaired or communicating with robots to enhancing the user’s visual experience with external knowledge.

Effective and robust VQA systems cannot exist without high-quality, semantically and stylistically diverse large-scale training data of image-question-answer triplets. But, creating such data is time consuming and onerous. Perhaps unsurprisingly, the VQA community has focused more on sophisticated model development rather than scalable data creation.

In “All You May Need for VQA are Image Captions,” published at NAACL 2022, we explore VQA data generation by proposing “Visual Question Generation with Question Answering Validation” (VQ2A), a pipeline that works by rewriting a declarative caption into multiple interrogative question-answer pairs. More specifically, we leverage two existing assets — (i) large-scale image-text data and (ii) large-capacity neural text-to-text models — to achieve automatic VQA data generation. As the field has progressed, the research community has been making these assets larger and stronger in isolation (for general purposes such as learning text-only or image-text representations); together, they can achieve more and we adapt them for VQA data creation purposes. We find our approach can generate question-answer pairs with high precision and that this data can successfully be used for training VQA models to improve performance.

The VQ2A technique enables VQA data generation at scale from image captions by rewriting each caption into multiple question-answer pairs.

VQ2A Overview
The first step of the VQ2A approach is to apply heuristics based on named entity recognition, part-of-speech tagging and manually defined rules to generate answer candidates from the image caption. These generated candidates are small pieces of information that may be relevant subjects about which to ask questions. We also add to this list two default answers, “yes” and “no”, which allow us to generate Boolean questions.

Then, we use a T5 model that was fine-tuned to generate questions for the candidate, resulting in [question, candidate answer] pairs. We then filter for the highest quality pairs using another T5 model (fine-tuned to answer questions) by asking it to answer the question based on the caption. was . That is, we compare the candidate answer to the output of this model and if the two answers are similar enough, we define this question as high quality and keep it. Otherwise, we filter it out.

The idea of using both question answering and question generation models to check each other for their round-trip consistency has been previously explored in other contexts. For instance, Q2 uses this idea to evaluate factual consistency in knowledge-grounded dialogues. In the end, the VQ2A approach, as illustrated below, can generate a large number of [image, question, answer] triplets that are high-quality enough to be used as VQA training data.

VQ2A consists of three main steps: (i) candidate answer extraction, (ii) question generation, (iii) question answering and answer validation.

Results
Two examples of our generated VQA data are shown below, one based on human-written COCO Captions (COCO) and the other on automatically-collected Conceptual Captions (CC3M), which we call VQ2A-COCO and VQ2A-CC3M, respectively. We highlight the variety of question types and styles, which are critical for VQA. Overall, the cleaner the captions (i.e., the more closely related they are to their paired image), the more accurate the generated triplets. Based on 800 samples each, 87.3% of VQ2A-COCO and 66.0% VQ2A-CC3M are found by human raters to be valid, suggesting that our approach can generate question-answer pairs with high precision.

Generated question-answer pairs based on COCO Captions (top) and Conceptual Captions (bottom). Grey highlighting denotes questions that do not appear in VQAv2, while green highlighting denotes those that do, indicating that our approach is capable of generating novel questions that an existing VQA dataset does not have.

Finally, we evaluate our generated data by using it to train VQA models (highlights shown below). We observe that our automatically-generated VQA data is competitive with manually-annotated target VQA data. First, our VQA models achieve high performance on target benchmarks “out-of-the-box”, when trained only on our generated data (light blue and light red vs. yellow). Once fine-tuned on target data, our VQA models outperform target-only training slightly on large-scale benchmarks like VQAv2 and GQA, but significantly on the small, knowledge-seeking OK-VQA (dark blue/red vs. light blue/red).

VQA accuracy on popular benchmark datasets.

Conclusion
All we may need for VQA are image captions! This work demonstrates that it is possible to automatically generate high-quality VQA data at scale, serving as an essential building block for VQA and vision-and-language models in general (e.g., ALIGN, CoCa). We hope that our work inspires other work on data-centric VQA.

Acknowledgments
We thank Roee Aharoni, Idan Szpektor, and Radu Soricut for their feedback on this blogpost. We also thank our co-authors: Xi Chen, Nan Ding, Idan Szpektor, and Radu Soricut. We acknowledge contributions from Or Honovich, Hagai Taitelbaum, Roee Aharoni, Sebastian Goodman, Piyush Sharma, Nassim Oufattole, Gal Elidan, Sasha Goldshtein, and Avinatan Hassidim. Finally, we thank the authors of Q2, whose pipeline strongly influences this work.

Source: Google AI Blog


Contextual Rephrasing in Google Assistant

Posted by Aurelien Boffy, Senior Staff Software Engineer, and Roberto Pieraccini, Engineering Director, Google Assistant

When people converse with one another, context and references play a critical role in driving their conversation more efficiently. For instance, if one asks the question “Who wrote Romeo and Juliet?” and, after receiving an answer, asks “Where was he born?”, it is clear that ‘he’ is referring to William Shakespeare without the need to explicitly mention him. Or if someone mentions “python” in a sentence, one can use the context from the conversation to determine whether they are referring to a type of snake or a computer language. If a virtual assistant cannot robustly handle context and references, users would be required to adapt to the limitation of the technology by repeating previously shared contextual information in their follow-up queries to ensure that the assistant understands their requests and can provide relevant answers.

In this post, we present a technology currently deployed on Google Assistant that allows users to speak in a natural manner when referencing context that was defined in previous queries and answers. The technology, based on the latest machine learning (ML) advances, rephrases a user’s follow-up query to explicitly mention the missing contextual information, thus enabling it to be answered as a stand-alone query. While Assistant considers many types of context for interpreting the user input, in this post we are focusing on short-term conversation history.

Context Handling by Rephrasing
One of the approaches taken by Assistant to understand contextual queries is to detect if an input utterance is referring to previous context and then rephrase it internally to explicitly include the missing information. Following on from the previous example in which the user asked who wrote Romeo and Juliet, one may ask follow-up questions like “When?”. Assistant recognizes that this question is referring to both the subject (Romeo and Juliet) and answer from the previous query (William Shakespeare) and can rephrase “When?” to “When did William Shakespeare write Romeo and Juliet?”

While there are other ways to handle context, for instance, by applying rules directly to symbolic representations of the meaning of queries, like intents and arguments, the advantage of the rephrasing approach is that it operates horizontally at the string level across any query answering, parsing, or action fulfillment module.

Conversation on a smart display device, where Assistant understands multiple contextual follow-up queries, allowing the user to have a more natural conversation. The phrases appearing at the bottom of the display are suggestions for follow-up questions that the user can select. However, the user can still ask different questions.

A Wide Variety of Contextual Queries
The natural language processing field, traditionally, has not put much emphasis on a general approach to context, focusing on the understanding of stand-alone queries that are fully specified. Accurately incorporating context is a challenging problem, especially when considering the large variety of contextual query types. The table below contains example conversations that illustrate query variability and some of the many contextual challenges that Assistant’s rephrasing method can resolve (e.g., differentiating between referential and non-referential cases or identifying what context a query is referencing). We demonstrate how Assistant is now able to rephrase follow-up queries, adding contextual information before providing an answer.

System Architecture
At a high level, the rephrasing system generates rephrasing candidates by using different types of candidate generators. Each rephrasing candidate is then scored based on a number of signals, and the one with the highest score is selected.

High level architecture of Google Assistant contextual rephraser.

Candidate Generation
To generate rephrasing candidates we use a hybrid approach that applies different techniques, which we classify into three categories:

  1. Generators based on the analysis of the linguistic structure of the queries use grammatical and morphological rules to perform specific operations — for instance, the replacement of pronouns or other types of referential phrases with antecedents from the context.
  2. Generators based on query statistics combine key terms from the current query and its context to create candidates that match popular queries from historical data or common query patterns.
  3. Generators based on Transformer technologies, such as MUM, learn to generate sequences of words according to a number of training samples. LaserTagger and FELIX are technologies suitable for tasks with high overlap between the input and output texts, are very fast at inference time, and are not vulnerable to hallucination (i.e., generating text that is not related to the input texts). Once presented with a query and its context, they can generate a sequence of text edits to transform the input queries into a rephrasing candidate by indicating which portions of the context should be preserved and which words should be modified.

Candidate Scoring
We extract a number of signals for each rephrasing candidate and use an ML model to select the most promising candidate. Some of the signals depend only on the current query and its context. For example, is the topic of the current query similar to the topic of the previous query? Or, is the current query a good stand-alone query or does it look incomplete? Other signals depend on the candidate itself: How much of the information of the context does the candidate preserve? Is the candidate well-formed from a linguistic point of view? Etc.

Recently, new signals generated by BERT and MUM models have significantly improved the performance of the ranker, fixing about one-third of the recall headroom while minimizing false positives on query sequences that are not contextual (and therefore do not require a rephrasing).

Example conversation on a phone where Assistant understands a sequence of contextual queries.

Conclusion
The solution described here attempts to resolve contextual queries by rephrasing them in order to make them fully answerable in a stand-alone manner, i.e., without having to relate to other information during the fulfillment phase. The benefit of this approach is that it is agnostic to the mechanisms that would fulfill the query, thus making it usable as a horizontal layer to be deployed before any further processing.

Given the variety of contexts naturally used in human languages, we adopted a hybrid approach that combines linguistic rules, large amounts of historic data through logs, and ML models based on state-of-the-art Transformer approaches. By generating a number of rephrasing candidates for each query and its context, and then scoring and ranking them using a variety of signals, Assistant can rephrase and thus correctly interpret most contextual queries. As Assistant can handle most types of linguistic references, we are empowering users to have more natural conversations. To make such multi-turn conversations even less cumbersome, Assistant users can turn on Continued Conversation mode to enable asking follow-up queries without the need to repeat "Hey Google" between each query. We are also using this technology in other virtual assistant settings, for instance, interpreting context from something shown on a screen or playing on a speaker.

Acknowledgements
This post reflects the combined work of Aliaksei Severyn, André Farias, Cheng-Chun Lee, Florian Thöle, Gabriel Carvajal, Gyorgy Gyepesi, Julien Cretin, Liana Marinescu, Martin Bölle, Patrick Siegler, Sebastian Krause, Victor Ähdel, Victoria Fossum, Vincent Zhao. We also thank Amar Subramanya, Dave Orr, Yury Pinsky for helpful discussions and support.

Source: Google AI Blog


Locked-image Tuning: Adding Language Understanding to Image Models

Posted by Andreas Steiner and Basil Mustafa, Research Software Engineers at Google Research, Brain team

The ability to classify images into categories has been transformed by deep learning. It has also been significantly accelerated by transfer learning, whereby models are first pre-trained on large datasets, like ImageNet, to learn visual representations that are then transferred via fine-tuning to a new task with less data (e.g., classifying animals). Previous works such as BiT and ViT employed these methods to achieve state-of-the-art performance on a wide range of classification tasks, such as the VTAB benchmark.

However, fine-tuning has some downsides: though pre-training is done only once, fine-tuning is necessary on every new dataset for which task-specific data is needed. Multimodal contrastive learning is an alternative, recently popularized paradigm (e.g., CLIP, ALIGN) that overcomes these issues by instead learning how to match free-form text with images. These models can then solve new tasks by reformulating them as image-text matching problems, without extra data (referred to as “zero-shot” learning). Contrastive learning is flexible and easy to adapt to new tasks, but has its own limitations, namely the need for a lot of paired image-text data and weaker performance than transfer learning approaches.

With those limitations in mind, we propose “LiT: Zero-Shot Transfer with Locked-image Text Tuning”, to appear at CVPR 2022. LiT models learn to match text to an already pre-trained image encoder. This simple yet effective setup provides the best of both worlds: strong image representations from pre-training, plus flexible zero-shot transfer to new tasks via contrastive learning. LiT achieves state-of-the-art zero-shot classification accuracy, significantly closing the gap between the two styles of learning. We think the best way to understand is to try it yourself, so we’ve included a demo of LiT models at the end of this post.

Fine-tuning (left) requires task-specific data and training to adapt a pre-trained model to a new task. An LiT model (right) can be used with any task, without further data or adaptation.

Contrastive Learning on Image-Text Data
Contrastive learning models learn representations from “positive” and “negative” examples, such that representations for "positive" examples are similar to each other but different from "negative" examples.

Multimodal contrastive learning applies this to pairs of images and associated texts. An image encoder computes representations from images, and a text encoder does the same for texts. Each image representation is encouraged to be close to the representation of its associated text (“positive”), but distinct from the representation of other texts ("negatives") in the data, and vice versa. This has typically been done with randomly initialized models (“from scratch”), meaning the encoders have to simultaneously learn representations and how to match them.

Multimodal contrastive learning trains models to produce similar representations for closely matched images and texts.

This training can be done on noisy, loosely aligned pairs of image and text, which naturally occur on the web. This circumvents the need for manual labeling, and makes data scaling easy. Furthermore, the model learns much richer visual concepts — it’s not constrained to what’s defined in the classification label space. Instead of classifying an image as “coffee”, it can understand whether it’s "a small espresso in a white mug” or “a large latte in a red flask”.

Once trained, a model that aligns image and text can be used in many ways. For zero-shot classification, we compare image representations to text representations of the class names. For example, a “wombat vs jaguar” classifier can be built by computing the representations of the texts “jaguar” and “wombat”, and classifying an image as a jaguar if its representation better matches the former. This approach scales to thousands of classes and makes it very easy to solve classification tasks without the extra data necessary for fine-tuning. Another application of contrastive models is image search (a.k.a. image-text retrieval), by finding the image whose representation best matches that of a given text, or vice versa.

The Best of Both Worlds with Locked-image Tuning
As mentioned earlier, transfer learning achieves state-of-the-art accuracy, but requires per-task labels, datasets, and training. On the other hand, contrastive models are flexible, scalable, and easily adaptable to new tasks, but fall short in performance. To compare, at the time of writing, the state of the art on ImageNet classification using transfer learning is 90.94%, but the best contrastive zero-shot models achieve 76.4%.

LiT tuning bridges this gap: we contrastively train a text model to compute representations well aligned with the powerful ones available from a pre-trained image encoder. Importantly, for this to work well, the image encoder should be “locked“, that is: it should not be updated during training. This may be unintuitive since one usually expects the additional information from further training to increase performance, but we find that locking the image encoder consistently leads to better results.

LiT-tuning contrastively trains a text encoder to match a pre-trained image encoder. The text encoder learns to compute representations that align to those from the image encoder.

This can be considered an alternative to the classic fine-tuning stage, where the image encoder is separately adapted to every new classification task; instead we have one stage of LiT-tuning, after which the model can classify any data. LiT-tuned models achieve 84.5% zero-shot accuracy on ImageNet classification, showing significant improvements over previous methods that train models from scratch, and halving the performance gap between fine-tuning and contrastive learning.

Left: LiT-tuning significantly closes the gap between the best contrastive models and the best models fine-tuned with labels. Right: Using a pre-trained image encoder is always helpful, but locking it is surprisingly a key part of the recipe to success; unlocked image models (dashed) yield significantly worse performance.

An impressive benefit of contrastive models is increased robustness — they retain high accuracy on datasets that typically fool fine-tuned models, such as ObjectNet and ImageNet-C. Similarly, LiT-tuned models have high performance across various challenging versions of ImageNet, for example achieving a state-of-the-art 81.1% accuracy on ObjectNet.

LiT-tuning has other advantages. While prior contrastive works require large amounts of data and train for a very long time, the LiT approach is much less data hungry. LiT models trained on 24M publicly available image-text pairs rival the zero-shot classification performance of prior models trained on 400M image-text pairs of private data. The locked image encoder also leads to faster training with a smaller memory footprint. On larger datasets, image representations can be pre-computed; not running the image model during training further improves efficiency and also unlocks much larger batch sizes, which increases the number of “negatives” the model sees and is key to high-performance contrastive learning. The method works well with varied forms of image pre-training (e.g., including self-supervised learning), and with many publicly available image models. We hope that these benefits make LiT a great testbed for researchers.

Conclusion
We present Locked-image Tuning (LiT), which contrastively trains a text encoder to match image representations from a powerful pre-trained image encoder. This simple method is data and compute efficient, and substantially improves zero-shot classification performance compared to existing contrastive learning approaches.

Want to try it yourself?

A preview of the demo: use it to match free-form text descriptions to images and build your own zero-shot classifier!

We have prepared a small interactive demo to try some LiT-tuned models. We also provide a Colab with more advanced use cases and larger models, which are a great way to get started.

Acknowledgments
We would like to thank Xiaohua Zhai, Xiao Wang, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer who have co-authored the LiT paper and been involved in all aspects of its development, as well as the Brain team in Zürich. We also would like to thank Tom Small for creating the animations used in this blogpost.

Source: Google AI Blog


Federated Learning with Formal Differential Privacy Guarantees

Posted by Brendan McMahan and Abhradeep Thakurta, Research Scientists, Google Research

In 2017, Google introduced federated learning (FL), an approach that enables mobile devices to collaboratively train machine learning (ML) models while keeping the raw training data on each user's device, decoupling the ability to do ML from the need to store the data in the cloud. Since its introduction, Google has continued to actively engage in FL research and deployed FL to power many features in Gboard, including next word prediction, emoji suggestion and out-of-vocabulary word discovery. Federated learning is improving the “Hey Google” detection models in Assistant, suggesting replies in Google Messages, predicting text selections, and more.

While FL allows ML without raw data collection, differential privacy (DP) provides a quantifiable measure of data anonymization, and when applied to ML can address concerns about models memorizing sensitive user data. This too has been a top research priority, and has yielded one of the first production uses of DP for analytics with RAPPOR in 2014, our open-source DP library, Pipeline DP, and TensorFlow Privacy.

Through a multi-year, multi-team effort spanning fundamental research and product integration, today we are excited to announce that we have deployed a production ML model using federated learning with a rigorous differential privacy guarantee. For this proof-of-concept deployment, we utilized the DP-FTRL algorithm to train a recurrent neural network to power next-word-prediction for Spanish-language Gboard users. To our knowledge, this is the first production neural network trained directly on user data announced with a formal DP guarantee (technically ρ=0.81 zero-Concentrated-Differential-Privacy, zCDP, discussed in detail below). Further, the federated approach offers complimentary data minimization advantages, and the DP guarantee protects all of the data on each device, not just individual training examples.

Data Minimization and Anonymization in Federated Learning
Along with fundamentals like transparency and consent, the privacy principles of data minimization and anonymization are important in ML applications that involve sensitive data.

Federated learning systems structurally incorporate the principle of data minimization. FL only transmits minimal updates for a specific model training task (focused collection), limits access to data at all stages, processes individuals’ data as early as possible (early aggregation), and discards both collected and processed data as soon as possible (minimal retention).

Another principle that is important for models trained on user data is anonymization, meaning that the final model should not memorize information unique to a particular individual's data, e.g., phone numbers, addresses, credit card numbers. However, FL on its own does not directly tackle this problem.

The mathematical concept of DP allows one to formally quantify this principle of anonymization. Differentially private training algorithms add random noise during training to produce a probability distribution over output models, and ensure that this distribution doesn't change too much given a small change to the training data; ρ-zCDP quantifies how much the distribution could possibly change. We call this example-level DP when adding or removing a single training example changes the output distribution on models in a provably minimal way.

Showing that deep learning with example-level differential privacy was even possible in the simpler setting of centralized training was a major step forward in 2016. Achieved by the DP-SGD algorithm, the key was amplifying the privacy guarantee by leveraging the randomness in sampling training examples ("amplification-via-sampling").

However, when users can contribute multiple examples to the training dataset, example-level DP is not necessarily strong enough to ensure the users’ data isn't memorized. Instead, we have designed algorithms for user-level DP, which requires that the output distribution of models doesn't change even if we add/remove all of the training examples from any one user (or all the examples from any one device in our application). Fortunately, because FL summarizes all of a user's training data as a single model update, federated algorithms are well-suited to offering user-level DP guarantees.

Both limiting the contributions from one user and adding noise can come at the expense of model accuracy, however, so maintaining model quality while also providing strong DP guarantees is a key research focus.

The Challenging Path to Federated Learning with Differential Privacy
In 2018, we introduced the DP-FedAvg algorithm, which extended the DP-SGD approach to the federated setting with user-level DP guarantees, and in 2020 we deployed this algorithm to mobile devices for the first time. This approach ensures the training mechanism is not too sensitive to any one user's data, and empirical privacy auditing techniques rule out some forms of memorization.

However, the amplification-via-samping argument is essential to providing a strong DP guarantee for DP-FedAvg, but in a real-world cross-device FL system ensuring devices are subsampled precisely and uniformly at random from a large population would be complex and hard to verify. One challenge is that devices choose when to connect (or "check in") based on many external factors (e.g., requiring the device is idle, on unmetered WiFi, and charging), and the number of available devices can vary substantially.

Achieving a formal privacy guarantee requires a protocol that does all of the following:

  • Makes progress on training even as the set of devices available varies significantly with time.
  • Maintains privacy guarantees even in the face of unexpected or arbitrary changes in device availability.
  • For efficiency, allows client devices to locally decide whether they will check in to the server in order to participate in training, independent of other devices.

Initial work on privacy amplification via random check-ins highlighted these challenges and introduced a feasible protocol, but it would have required complex changes to our production infrastructure to deploy. Further, as with the amplification-via-sampling analysis of DP-SGD, the privacy amplification possible with random check-ins depends on a large number of devices being available. For example, if only 1000 devices are available for training, and participation of at least 1000 devices is needed in each training step, that requires either 1) including all devices currently available and paying a large privacy cost since there is no randomness in the selection, or 2) pausing the protocol and not making progress until more devices are available.

Achieving Provable Differential Privacy for Federated Learning with DP-FTRL
To address this challenge, the DP-FTRL algorithm is built on two key observations: 1) the convergence of gradient-descent-style algorithms depends primarily not on the accuracy of individual gradients, but the accuracy of cumulative sums of gradients; and 2) we can provide accurate estimates of cumulative sums with a strong DP guarantee by utilizing negatively correlated noise, added by the aggregating server: essentially, adding noise to one gradient and subtracting that same noise from a later gradient. DP-FTRL accomplishes this efficiently using the Tree Aggregation algorithm [1, 2].

The graphic below illustrates how estimating cumulative sums rather than individual gradients can help. We look at how the noise introduced by DP-FTRL and DP-SGD influence model training, compared to the true gradients (without added noise; in black) which step one unit to the right on each iteration. The individual DP-FTRL gradient estimates (blue), based on cumulative sums, have larger mean-squared-error than the individually-noised DP-SGD estimates (orange), but because the DP-FTRL noise is negatively correlated, some of it cancels out from step to step, and the overall learning trajectory stays closer to the true gradient descent steps.

To provide a strong privacy guarantee, we limit the number of times a user contributes an update. Fortunately, sampling-without-replacement is relatively easy to implement in production FL infrastructure: each device can remember locally which models it has contributed to in the past, and choose to not connect to the server for any later rounds for those models.

Production Training Details and Formal DP Statements
For the production DP-FTRL deployment introduced above, each eligible device maintains a local training cache consisting of user keyboard input, and when participating computes an update to the model which makes it more likely to suggest the next word the user actually typed, based on what has been typed so far. We ran DP-FTRL on this data to train a recurrent neural network with ~1.3M parameters. Training ran for 2000 rounds over six days, with 6500 devices participating per round. To allow for the DP guarantee, devices participated in training at most once every 24 hours. Model quality improved over the previous DP-FedAvg trained model, which offered empirically-tested privacy advantages over non-DP models, but lacked a meaningful formal DP guarantee.

The training mechanism we used is available in open-source in TensorFlow Federated and TensorFlow Privacy, and with the parameters used in our production deployment it provides a meaningfully strong privacy guarantee. Our analysis gives ρ=0.81 zCDP at the user level (treating all the data on each device as a different user), where smaller numbers correspond to better privacy in a mathematically precise way. As a comparison, this is stronger than the ρ=2.63 zCDP guarantee chosen by the 2020 US Census.

Next Steps
While we have reached the milestone of deploying a production FL model using a mechanism that provides a meaningfully small zCDP, our research journey continues. We are still far from being able to say this approach is possible (let alone practical) for most ML models or product applications, and other approaches to private ML exist. For example, membership inference tests and other empirical privacy auditing techniques can provide complimentary safeguards against leakage of users’ data. Most importantly, we see training models with user-level DP with even a very large zCDP as a substantial step forward, because it requires training with a DP mechanism that bounds the sensitivity of the model to any one user's data. Further, it smooths the road to later training models with improved privacy guarantees as better algorithms or more data become available. We are excited to continue the journey toward maximizing the value that ML can deliver while minimizing potential privacy costs to those who contribute training data.

Acknowledgements
The authors would like to thank Alex Ingerman and Om Thakkar for significant impact on the blog post itself, as well as the teams at Google that helped develop these ideas and bring them to practice:

  • Core research team: Galen Andrew, Borja Balle, Peter Kairouz, Daniel Ramage, Shuang Song, Thomas Steinke, Andreas Terzis, Om Thakkar, Zheng Xu
  • FL infrastructure team: Katharine Daly, Stefan Dierauf, Hubert Eichner, Igor Pisarev, Timon Van Overveldt, Chunxiang Zheng
  • Gboard team: Angana Ghosh, Xu Liu, Yuanbo Zhang
  • Speech team: Françoise Beaufays, Mingqing Chen, Rajiv Mathews, Vidush Mukund, Igor Pisarev, Swaroop Ramaswamy, Dan Zivkovic

Source: Google AI Blog


LaMDA: Towards Safe, Grounded, and High-Quality Dialog Models for Everything

Posted by Heng-Tze Cheng, Senior Staff Software Engineer and Romal Thoppilan, Senior Software Engineer, Google Research, Brain Team

Language models are becoming more capable than ever before and are helpful in a variety of tasks — translating one language into another, summarizing a long document into a brief highlight, or answering information-seeking questions. Among these, open-domain dialog, where a model needs to be able to converse about any topic, is probably one of the most difficult, with a wide range of potential applications and open challenges. In addition to producing responses that humans judge as sensible, interesting, and specific to the context, dialog models should adhere to Responsible AI practices, and avoid making factual statements that are not supported by external information sources.

Today we’re excited to share recent advances in our “LaMDA: Language Models for Dialog Applications” project. In this post, we’ll give an overview on how we’re making progress towards safe, grounded, and high-quality dialog applications. LaMDA is built by fine-tuning a family of Transformer-based neural language models specialized for dialog, with up to 137B model parameters, and teaching the models to leverage external knowledge sources.

Objectives & Metrics
Defining objectives and metrics is critical to guide training dialog models. LaMDA has three key objectives — Quality, Safety, and Groundedness — each of which we measure using carefully designed metrics:

Quality: We decompose Quality into three dimensions, Sensibleness, Specificity, and Interestingness (SSI), which are evaluated by human raters. Sensibleness refers to whether the model produces responses that make sense in the dialog context (e.g., no common sense mistakes, no absurd responses, and no contradictions with earlier responses). Specificity is measured by judging whether the system's response is specific to the preceding dialog context, and not a generic response that could apply to most contexts (e.g., “ok” or “I don’t know”). Finally, Interestingness measures whether the model produces responses that are also insightful, unexpected or witty, and are therefore more likely to create better dialog.

Safety: We’re also making progress towards addressing important questions related to the development and deployment of Responsible AI. Our Safety metric is composed of an illustrative set of safety objectives that captures the behavior that the model should exhibit in a dialog. These objectives attempt to constrain the model’s output to avoid any unintended results that create risks of harm for the user, and to avoid reinforcing unfair bias. For example, these objectives train the model to avoid producing outputs that contain violent or gory content, promote slurs or hateful stereotypes towards groups of people, or contain profanity. Our research towards developing a practical Safety metric represents very early work, and there is still a great deal of progress for us to make in this area.

Groundedness: The current generation of language models often generate statements that seem plausible, but actually contradict facts established in known external sources. This motivates our study of groundedness in LaMDA. Groundedness is defined as the percentage of responses with claims about the external world that can be supported by authoritative external sources, as a share of all responses containing claims about the external world. A related metric, Informativeness, is defined as the percentage of responses with information about the external world that can be supported by known sources, as a share of all responses. Therefore, casual responses that do not carry any real world information (e.g., “That’s a great idea”), affect Informativeness but not Groundedness. While grounding LaMDA generated responses in known sources does not in itself guarantee factual accuracy, it allows users or external systems to judge the validity of a response based on the reliability of its source.

LaMDA Pre-Training
With the objectives and metrics defined, we describe LaMDA’s two-stage training: pre-training and fine-tuning. In the pre-training stage, we first created a dataset of 1.56T words — nearly 40 times more words than what were used to train previous dialog models — from public dialog data and other public web documents. After tokenizing the dataset into 2.81T SentencePiece tokens, we pre-train the model using GSPMD to predict every next token in a sentence, given the previous tokens. The pre-trained LaMDA model has also been widely used for natural language processing research across Google, including program synthesis, zero-shot learning, style transfer, as well as in the BIG-bench workshop.

LaMDA Fine-Tuning
In the fine-tuning stage, we train LaMDA to perform a mix of generative tasks to generate natural-language responses to given contexts, and classification tasks on whether a response is safe and high-quality, resulting in a single multi-task model that can do both. The LaMDA generator is trained to predict the next token on a dialog dataset restricted to back-and-forth dialog between two authors, while the LaMDA classifiers are trained to predict the Safety and Quality (SSI) ratings for the response in context using annotated data. During a dialog, the LaMDA generator first generates several candidate responses given the current multi-turn dialog context, and the LaMDA classifiers predict the SSI and Safety scores for every response candidate. Candidate responses with low Safety scores are first filtered out. Remaining candidates are re-ranked by their SSI scores, and the top result is selected as the response. We further filter the training data used for the generation task with LaMDA classifiers to increase the density of high-quality response candidates.

LaMDA generates and then scores a response candidate.
LaMDA handles arbitrary user input in a way that is sensible, specific, and interesting. Only LaMDA’s very first statement “Hello, I’m a friendly...” was hard coded to set the purpose of the dialog.

Factual Grounding
While people are capable of checking their facts by using tools and referencing established knowledge bases, many language models draw their knowledge on their internal model parameters only. To improve the groundedness of LaMDA’s original response, we collect a dataset of dialogs between people and LaMDA, which are annotated with information retrieval queries and the retrieved results where applicable. We then fine-tune LaMDA’s generator and classifier on this dataset to learn to call an external information retrieval system during its interaction with the user to improve the groundedness of its responses. While this is very early work, we’re seeing promising results.

Zero-shot domain adaptation: cherry-picked, but real example of LaMDA pretending to be Mount Everest, by simply setting its initial message to be “Hi I’m Mount Everest. What would you like me to know about me?” Everest LaMDA is shown providing educational and factually correct responses.

Evaluation
In order to quantify progress against our key metrics, we collect responses from the pre-trained model, fine-tuned model, and human raters (i.e., human-generated responses) to multi-turn two-author dialogs, and then ask a different set of human raters a series of questions to evaluate these responses against the Quality, Safety, and Groundedness metrics.

We observe that LaMDA significantly outperforms the pre-trained model in every dimension and across all model sizes. Quality metrics (Sensibleness, Specificity, and Interestingness, in the first column below) generally improve with the number of model parameters, with or without fine-tuning. Safety does not seem to benefit from model scaling alone, but it does improve with fine-tuning. Groundedness improves as model size increases, perhaps because larger models have a greater capacity to memorize uncommon knowledge, but fine-tuning allows the model to access external knowledge sources and effectively shift some of the load of remembering knowledge to an external knowledge source. With fine-tuning, the quality gap to human levels can be narrowed, though the model’s performance remains below human levels in safety and groundedness.

Comparing the pre-trained model (PT), fine-tuned model (LaMDA) and human-rater-generated dialogs (Human) across Sensibleness, Specificity, Interestingness, Safety, Groundedness, and Informativeness. The test sets used to measure Safety and Groundedness were designed to be especially difficult.

Future Research & Challenges
LaMDA’s level of Sensibleness, Specificity and Interestingness unlocks new avenues for understanding the benefits and risks of open-ended dialog agents. It also presents encouraging evidence that key challenges with neural language models, such as using a safety metric and improving groundedness, can improve with larger models and fine-tuning with more well-labeled data. However, this is very early work, and there are significant limitations. Exploring new ways to improve our Safety metric and LaMDA's groundedness, aligned with our AI Principles, will continue to be our main areas of focus going forward.

Acknowledgements
We'd to like to thank everyone for contributing to the project and paper, including: Blaise Aguera-Arcas, Javier Alberca, Thushan Amarasiriwardena, Lora Aroyo, Martin Baeuml, Leslie Baker, Rachel Bernstein, Taylor Bos, Maarten Bosma, Jonas Bragagnolo, Alena Butryna, Bill Byrne, Chung-Ching Chang, Zhifeng Chen, Dehao Chen, Heng-Tze Cheng, Ed Chi, Aaron Cohen, Eli Collins, Marian Croak, Claire Cui, Andrew Dai, Dipanjan Das, Daniel De Freitas, Jeff Dean, Rajat Dewan, Mark Diaz, Tulsee Doshi, Yu Du, Toju Duke, Doug Eck, Joe Fenton, Noah Fiedel, Christian Frueh, Harish Ganapathy, Saravanan Ganesh, Amin Ghafouri, Zoubin Ghahramani, Kourosh Gharachorloo, Jamie Hall, Erin Hoffman-John, Sissie Hsiao, Yanping Huang, Ben Hutchinson, Daphne Ippolito, Alicia Jin, Thomas Jurdi, Ashwin Kakarla, Nand Kishore, Maxim Krikun, Karthik Krishnamoorthi, Igor Krivokon, Apoorv Kulshreshtha, Ray Kurzweil, Viktoriya Kuzmina, Vivek Kwatra, Matthew Lamm, Quoc Le, Max Lee, Katherine Lee, Hongrae Lee, Josh Lee, Dmitry Lepikhin, YaGuang Li, Yifeng Lu, David Luan, Daphne Luong, Laichee Man, Jianchang (JC) Mao, Yossi Matias, Kathleen Meier-Hellstern, Marcelo Menegali, Muqthar Mohammad,, Muqthar Mohammad, Alejandra Molina, Erica Moreira, Meredith Ringel Morris, Maysam Moussalem, Jiaqi Mu, Tyler Mullen, Tyler Mullen, Eric Ni, Kristen Olson, Alexander Passos, Fernando Pereira, Slav Petrov, Marc Pickett, Roberto Pieraccini, Christian Plagemann, Sahitya Potluri, Vinodkumar Prabhakaran, Andy Pratt, James Qin, Ravi Rajakumar, Adam Roberts, Will Rusch, Renelito Delos Santos, Noam Shazeer, RJ Skerry-Ryan, Grigori Somin, Johnny Soraker, Pranesh Srinivasan, Amarnag Subramanya, Mustafa Suleyman, Romal Thoppilan, Song Wang, Sheng Wang, Chris Wassman, Yuanzhong Xu, Yuanzhong Xu, Ni Yan, Ben Zevenbergen, Vincent Zhao, Huaixiu Steven Zheng, Denny Zhou, Hao Zhou, Yanqi Zhou, and more.

Source: Google AI Blog


Evaluating Syntactic Abilities of Language Models

Posted by Jason Wei, AI Resident and Dan Garrette, Research Scientist, Google Research

In recent years, pre-trained language models, such as BERT and GPT-3, have seen widespread use in natural language processing (NLP). By training on large volumes of text, language models acquire broad knowledge about the world, achieving strong performance on various NLP benchmarks. These models, however, are often opaque in that it may not be clear why they perform so well, which limits further hypothesis-driven improvement of the models. Hence, a new line of scientific inquiry has arisen: what linguistic knowledge is contained in these models?

While there are many types of linguistic knowledge that one may want to investigate, a topic that provides a strong basis for analysis is the subject–verb agreement grammar rule in English, which requires that the grammatical number of a verb agree with that of the subject. For example, the sentence “The dogs run.” is grammatical because “dogs” and “run” are both plural, but “The dogs runs.” is ungrammatical because “runs” is a singular verb.

One framework for assessing the linguistic knowledge of a language model is targeted syntactic evaluation (TSE), in which minimally different pairs of sentences, one grammatical and one ungrammatical, are shown to a model, and the model must determine which one is grammatical. TSE can be used to test knowledge of the English subject–verb agreement rule by having the model judge between two versions of the same sentence: one where a particular verb is written in its singular form, and the other in which the verb is written in its plural form.

With the above context, in “Frequency Effects on Syntactic Rule-Learning in Transformers”, published at EMNLP 2021, we investigated how a BERT model’s ability to correctly apply the English subject–verb agreement rule is affected by the number of times the words are seen by the model during pre-training. To test specific conditions, we pre-trained BERT models from scratch using carefully controlled datasets. We found that BERT achieves good performance on subject–verb pairs that do not appear together in the pre-training data, which indicates that it does learn to apply subject–verb agreement. However, the model tends to predict the incorrect form when it is much more frequent than the correct form, indicating that BERT does not treat grammatical agreement as a rule that must be followed. These results help us to better understand the strengths and limitations of pre-trained language models.

Prior Work
Previous work used TSE to measure English subject–verb agreement ability in a BERT model. In this setup, BERT performs a fill-in-the-blank task (e.g., “the dog _ across the park”) by assigning probabilities to both the singular and plural forms of a given verb (e.g., “runs” and “run”). If the model has correctly learned to apply the subject–verb agreement rule, then it should consistently assign higher probabilities to the verb forms that make the sentences grammatically correct.

This previous work evaluated BERT using both natural sentences (drawn from Wikipedia) and nonce sentences, which are artificially constructed to be grammatically valid but semantically nonsensical, such as Noam Chomsky’s famous example “colorless green ideas sleep furiously”. Nonce sentences are useful when testing syntactic abilities because the model cannot just fall back on superficial corpus statistics: for example, while “dogs run” is much more common than “dogs runs”, “dogs publish” and “dogs publishes” will both be very rare, so a model is not likely to have simply memorized the fact that one of them is more likely than the other.

BERT achieves an accuracy of more than 80% on nonce sentences (far better than the random-chance baseline of 50%), which was taken as evidence that the model had learned to apply the subject–verb agreement rule. In our paper, we went beyond this previous work by pre-training BERT models under specific data conditions, allowing us to dig deeper into these results to see how certain patterns in the pre-training data affect performance.

Unseen Subject–Verb Pairs
We first looked at how well the model performs on subject–verb pairs that were seen during pre-training, versus examples in which the subject and verb were never seen together in the same sentence:

BERT’s error rate on natural and nonce evaluation sentences, stratified by whether a particular subject–verb (SV) pair was seen in the same sentence during training or not. BERT’s performance on unseen SV pairs is far better than simple heuristics such as picking the more frequent verb or picking the more frequent SV pair.

BERT’s error rate increases slightly for unseen subject–verb (SV) pairs, for both natural and nonce evaluation sentences, but it is still much better than naïve heuristics, such as picking the verb form that occurred more often in the pre-training data or picking the verb form that occurred more frequently with the subject noun. This tells us that BERT is not just reflecting back the things that it sees during pre-training: making decisions based on more than just raw frequencies and generalizing to novel subject–verb pairs are indications that the model has learned to apply some underlying rule concerning subject–verb agreement.

Frequency of Verbs
Next, we went beyond just seen versus unseen, and examined how the frequency of a word affects BERT’s ability to use it correctly with the subject–verb agreement rule. For this study, we chose a set of 60 verbs, and then created several versions of the pre-training data, each engineered to contain the 60 verbs at a specific frequency, ensuring that the singular and plural forms appeared the same number of times. We then trained BERT models from these different datasets and evaluated them on the subject–verb agreement task:

BERT’s ability to follow the subject–verb agreement rule depends on the frequency of verbs in the training set.

These results indicate that although BERT is able to model the subject–verb agreement rule, it needs to see a verb about 100 times before it can reliably use it with the rule.

Relative Frequency Between Verb Forms
Finally, we wanted to understand how the relative frequencies of the singular and plural forms of a verb affect BERT’s predictions. For example, if one form of the verb (e.g., “combat”) appeared in the pre-training data much more frequently than the other verb form (e.g., “combats”), then BERT might be more likely to assign a high probability to the more frequent form, even when it is grammatically incorrect. To evaluate this, we again used the same 60 verbs, but this time we created manipulated versions of the pre-training data where the frequency ratio between verb forms varied from 1:1 to 100:1. The figure below shows BERT’s performance for these varying levels of frequency imbalance:

As the frequency ratio between verb forms in training data becomes more imbalanced, BERT’s ability to use those verbs grammatically decreases.

These results show that BERT achieves good accuracy at predicting the correct verb form when the two forms are seen the same number of times during pre-training, but the results become worse as the imbalance between the frequencies increases. This implies that even though BERT has learned how to apply subject–verb agreement, it does not necessarily use it as a “rule”, instead preferring to predict high-frequency words regardless of whether they violate the subject–verb agreement constraint.

Conclusions
Using TSE to evaluate the performance of BERT reveals its linguistic abilities on syntactic tasks. Moreover, studying its syntactic ability in relation to how often words appear in the training dataset reveals the ways that BERT handles competing priorities — it knows that subjects and verbs should agree and that high frequency words are more likely, but doesn’t understand that agreement is a rule that must be followed and that the frequency is only a preference. We hope this work provides new insight into how language models reflect properties of the datasets on which they are trained.

Acknowledgements
It was a privilege to collaborate with Tal Linzen and Ellie Pavlick on this project.

Source: Google AI Blog