Category Archives: Research Blog

The latest news on Google Research

Exploring Weight Agnostic Neural Networks



When training a neural network to accomplish a given task, be it image classification or reinforcement learning, one typically refines a set of weights associated with each connection within the network. Another approach to creating successful neural networks that has shown substantial progress is neural architecture search, which constructs neural network architectures out of hand-engineered components such as convolutional network components or transformer blocks. It has been shown that neural network architectures built with these components, such as deep convolutional networks, have strong inductive biases for image processing tasks, and can even perform them when their weights are randomly initialized. While neural architecture search produces new ways of arranging hand-engineered components with known inductive biases for the task domain at hand, there has been little progress in the automated discovery of new neural network architectures with such inductive biases, for various task domains.

We can look at analogies to these useful components in examples of nature vs. nurture. Just as certain precocial species in biology—who possess anti-predator behaviors from the moment of birth—can perform complex motor and sensory tasks without learning, perhaps we can construct network architectures that can perform well without training. Of course, these natural (and by analogy, artificial) neural networks are further improved through training, but their ability to perform even without learning shows that they contain biases that make them well-suited to their task.

In “Weight Agnostic Neural Networks” (WANN), we present a first step toward searching specifically for networks with these biases: neural net architectures that can already perform various tasks, even when they use a random shared weight. Our motivation in this work is to question to what extent neural network architectures alone, without learning any weight parameters, can encode solutions for a given task. By exploring such neural network architectures, we present agents that can already perform well in their environment without the need to learn weight parameters. Furthermore, in order to spur progress in this field community, we have also open-sourced the code to reproduce our WANN experiments for the broader research community.
Left: A hand-engineered, fully-connected deep neural network with 2760 weight connections. Using a learning algorithm, we can solve for the set of 2760 weight parameters so that this network can perform the BipedalWalker-v2 task. Right: A weight agnostic neural network architecture with 44 connections that can perform the same Bipedal Walker task. Unlike the fully-connected network, this WANN can still perform the task without the need to train the weight parameters of each connection. In fact, to simplify the training, the WANN is designed to perform when the values of each weight connection are identical, or shared, and it will even function if this shared weight parameter is randomly sampled.
Finding WANNs
We start with a population of minimal neural network architecture candidates, each with very few connections only, and use a well-established topology search algorithm (NEAT), to evolve the architectures by adding single connections and single nodes one by one. The key idea behind WANNs is to search for architectures by de-emphasizing weights. Unlike traditional neural architecture search methods, where all of the weight parameters of new architectures need to be trained using a learning algorithm, we take a simpler and more efficient approach. Here, during the search, all candidate architectures are first assigned a single shared weight value at each iteration, and then optimized to perform well over a wide range of shared weight values.
Operators for searching the space of network topologies
Left: A minimal network topology, with input and outputs only partially connected.
Middle: Networks are altered in one of three ways:
(1) Insert Node: a new node is inserted by splitting an existing connection.
(2) Add Connection: a new connection is added by connecting two previously unconnected nodes.
(3) Change Activation: the activation function of a hidden node is reassigned.
Right: Possible activation functions (linear, step, sin, cosine, Gaussian, tanh, sigmoid, inverse, absolute value, ReLU)
In addition to exploring a range of weight agnostic neural networks, it is important to also look for network architectures that are only as complex as they need to be. We accomplish this by optimizing for both the performance of the networks and their complexity simultaneously, using techniques drawn from multi-objective optimization.
Overview of Weight Agnostic Neural Network Search and corresponding operators for searching the space of network topologies.
Training WANN Architectures
Unlike traditional networks, we can easily train the WANN by simply finding the best single shared weight parameter that maximizes its performance. In the example below, we see that our architecture works (to some extent) for a swing-up cartpole task using constant weights:
A WANN performing a Cartpole Swing-up task at various different weight parameters, and also using fine-tuned weight parameters.
As we see in the above figure, while WANNs can perform its task using range of shared weight parameters, the performance is still not comparable to a network that learns weights for each individual connection, as normally done in network training. If we want to further improve its performance, we can use the WANN architecture, and the best shared weight as a starting point to fine-tune the weights of each individual connection using a learning algorithm, like how we would normally train any neural network. Using the weight agnostic property of the network architecture as a starting point, and fine-tuning its performance via learning, may help provide insightful analogies to how animals learn.
Through the use of multi-objective optimization for both performance and network simplicity, our method found a simple WANN for a Car Racing from pixels task that works well without explicitly training for the weights of the network.
The ability for a network architecture to function using only random weights offers other advantages too. For instance, by using copies of the same WANN architecture, but where each copy of the WANN is assigned a different distinct weight value, we can create an ensemble of multiple distinct models for the same task. This ensemble generally achieves better performance than a single model. We illustrate this with an example of an MNIST classifier evolved to work with random weights:
An MNIST classifier evolved to work with random weights.
While a conventional network with random initialization will achieve ~10% accuracy on MNIST, this particular network architecture uses random weights and when applied to MNIST achieves an accuracy much better than chance (> 80%). When an ensemble of WANNs is used, each of which assigned with a different shared weight, the accuracy increases to > 90%.

Even without ensemble methods, collapsing the number of weight values in a network to one allows the network to be rapidly tuned. The ability to quickly fine-tune weights might be useful in continual lifelong learning, where agents acquire, adapt, and transfer skills throughout their lifespan. This makes WANNs particularly well positioned to exploit the Baldwin effect, the evolutionary pressure that rewards individuals predisposed to learn useful behaviors, without being trapped in the computationally expensive trap of ‘learning to learn’.

Conclusion
We hope that this work can serve as a stepping stone to help discover novel fundamental neural network components such as the convolutional network, whose discovery and application have been instrumental to the incredible progress made in deep learning. The computational resources available to the research community have grown significantly since the time convolutional neural networks were discovered. If we are devoting such resources to automated discovery and hope to achieve more than incremental improvements in network architectures, we believe it is also worth searching for with new building blocks, not just their arrangements.

If you are interested to learn more about this work, we invite readers to read our interactive article (or pdf version of the paper for offline reading). In addition to open sourcing these experiments to the research community, we have also released a general Python implementation of NEAT called PrettyNEAT to help interested readers to explore the exciting area of neural network evolution from first principles.

Source: Google AI Blog


Bi-Tempered Logistic Loss for Training Neural Nets with Noisy Data



The quality of models produced by machine learning (ML) algorithms directly depends on the quality of the training data, but real world datasets typically contain some amount of noise that introduces challenges for ML models. Noise in the dataset can take several forms from corrupted examples (e.g., lens flare in an image of a cat) to mislabelled examples from when the data was collected (e.g., an image of cat mislabelled as a flerken).

The ability of an ML model to deal with noisy training data depends in great part on the loss function used in the training process. For classification tasks, the standard loss function used for training is the logistic loss. However, this particular loss function falls short when handling noisy training examples due to two unfortunate properties:
  1. Outliers far away can dominate the overall loss: The logistic loss function is sensitive to outliers. This is because the loss function value grows without bound as the mislabelled examples (outliers) are far away from the decision boundary. Thus, a single bad example that is located far away from the decision boundary can penalize the training process to the extent that the final trained model learns to compensate for it by stretching the decision boundary and potentially sacrificing the remaining good examples. This “large-margin” noise issue is illustrated in the left panel of the figure below.
  2. Mislabeled examples nearby can stretch the decision boundary: The output of the neural network is a vector of activation values, which reflects the margin between the example and the decision boundary for each class. The softmax transfer function is used to convert the activation values into probabilities that an example will belong to each class. As the tail of this transfer function for the logistic loss decays exponentially fast, the training process will tend to stretch the boundary closer to a mislabeled example in order to compensate for its small margin. Consequently, the generalization performance of the network will immediately deteriorate, even with a low level of label noise (right panel below).
We visualize the decision surface of a 2-layered neural network as it is trained for binary classification. Blue and orange dots represent the examples from the two classes. The network is trained with logistic loss under two types of noisy conditions: (left) large-margin noise and (right) small-margin-noise.
We tackle these two problems in a recent paper by introducing a “bi-tempered” generalization of the logistic loss endowed with two tunable parameters that handle those situations well, which we call “temperatures”—t1, which characterizes boundedness, and t2 for tail-heaviness (i.e. the rate of decline in the tail of the transfer function). These properties are illustrated below. Setting both t1 and t2 to 1.0 recovers the logistic loss function. Setting t1 lower than 1.0 increases the boundedness and setting t2 greater than 1.0 makes for a heavier-tailed transfer function. We also introduce this interactive visualization which allows you to visualize the neural network training process with the bi-tempered logistic loss.
Left: Boundedness of the loss function. When t1 is between 0 and 1, exclusive, only a finite amount of loss is incurred for each example, even if they are mislabeled. Shown is t1 = 0.8. Right: Tail-heaviness of the transfer function. The heavy-tailed transfer function applies when t2 = > 1.0 and assigns higher probability for the same amount of activation, thus preventing the boundary from drawing closer to the noisy example. Shown is t2 = 2.0.
To demonstrate the effect of each temperature, we train a two-layer feed-forward neural network for a binary classification problem on a synthetic dataset that contains a circle of points from the first class, and a concentric ring of points from the second class. You can try this yourself on your browser with our interactive visualization. We use the standard logistic loss function, which can be recovered by setting both temperatures equal to 1.0, as well as our bi-tempered logistic loss for training the network. We then demonstrate the effects of each loss function for a clean dataset, a dataset with small-margin noise, large-margin noise, and a dataset with random noise.
Logistic vs. bi-tempered logistic loss: (a) noise-free labels, (b) small-margin label noise, (c) large-margin label noise, and (d) random label noise. The temperature values (t1, t2) for the tempered loss are shown above each figure. We find that for each situation, the decision boundary recovered by training with the bi-tempered logistic loss function is better than before.
Noise Free Case:
We show the results of training the model on the noise-free dataset in column (a), using the logistic loss (top) and the bi-tempered logistic loss (bottom). The white line shows the decision boundary for each model. The values of (t1, t2), the temperatures in the bi-tempered loss function, are shown below each column of the figure. Notice that for this choice of temperatures, the loss is bounded and the transfer function is tail-heavy. As can be seen, both losses produce good decision boundaries that successfully separates the two classes.

Small-Margin Noise:
To illustrate the effect of tail-heaviness of the probabilities, we artificially corrupt a random subset of the examples that are near the decision boundary, that is, we flip the labels of these points to the opposite class. The results of training the networks on data with small-margin noise using the logistic loss as well as the bi-tempered loss is shown in column (b).

As can be seen, the logistic loss, due to the lightness of the softmax tail, stretches the boundary closer to the noisy points to compensate for their low probabilities. On the other hand, the bi-tempered loss using only the tail-heavy probability transfer function by adjusting t2 can successfully avoid the noisy examples. This can be explained by the heavier tail of the tempered exponential function, which assigns reasonably high probability values (and thus, keeps the loss value small) while maintaining the decision boundary away from the noisy examples.

Large-Margin Noise:
Next, we evaluate the performance of the two loss functions for handling large-margin noisy examples. In (c), we randomly corrupt a subset of the examples that are located far away from the decision boundary, the outer side of the ring as well as points near the center).

For this case, we only use the boundedness property of the bi-tempered loss, while keeping the softmax probabilities the same as the logistic loss. The unboundedness of the logistic loss causes the decision boundary to expand towards the noisy points to reduce their loss values. On the other hand, the bounded bi-tempered loss, bounded by adjusting t1, incurs a finite amount of loss for each noisy example. As a result, the bi-tempered loss can avoid these noisy examples and maintain a good decision boundary.

Random Noise:
Finally, we investigate the effect of random noise in the training data on the two loss functions. Note that random noise comprises both small-margin and large-margin noisy examples. Thus, we use both boundedness and tail-heaviness properties of the bi-tempered loss function by setting the temperatures to (t1, t2) = (0.2, 4.0).

As can be seen from the results in the last column, (d), the logistic loss is highly affected by the noisy examples and clearly fails to converge to a good decision boundary. On the other hand, the bi-tempered can recover a decision boundary that is almost identical to the noise-free case.

Conclusion
In this work we constructed a bounded, tempered loss function that can handle large-margin outliers and introduced heavy-tailedness in our new tempered softmax function, which can handle small-margin mislabeled examples. Using our bi-tempered logistic loss, we achieve excellent empirical performance on training neural networks on a number of large standard datasets (please see our paper for full details). Note that the state-of-the-art neural networks have been optimized along with a large variety of variables such as: architecture, transfer function, choice of optimizer, and label smoothing to name just a few. Our method introduces two additional tunable variables, namely (t1, t2). We believe that with a systematic “joint optimization” of all commonly tried variables, significant further improvements can be achieved in conjunction with our loss function. This is of course a more long-term goal. We also plan to explore the idea of annealing the temperature parameters over the training process.

Acknowledgements:
This blogpost reflects work with our co-authors Manfred Warmuth, Visiting Researcher and Tomer Koren, Senior Research Scientist, Google Research. Preprint of our paper is available here, which contains theoretical analysis of the loss function and empirical results on standard datasets at scale.

Source: Google AI Blog


Turbo, An Improved Rainbow Colormap for Visualization



False color maps show up in many applications in computer vision and machine learning, from visualizing depth images to more abstract uses, such as image differencing. Colorizing images helps the human visual system pick out detail, estimate quantitative values, and notice patterns in data in a more intuitive fashion. However, the choice of color map can have a significant impact on a given task. For example, interpretation of “rainbow maps” have been linked to lower accuracy in mission critical applications, such as medical imaging. Still, in many applications, “rainbow maps” are preferred since they show more detail (at the expense of accuracy) and allow for quicker visual assessment.
Left: Disparity image displayed as greyscale. Right: The commonly used Jet rainbow map being used to create a false color image.
One of the most commonly used color mapping algorithms in computer vision applications is Jet, which is high contrast, making it useful for accentuating even weakly distinguished image features. However, if you look at the color map gradient, one can see distinct “bands” of color, most notably in the cyan and yellow regions. This causes sharp transitions when the map is applied to images, which are misleading when the underlying data is actually smoothly varying. Because the rate at which the color changes ‘perceptually’ is not constant, Jet is not perceptually uniform. These effects are even more pronounced for users that are color blind, to the point of making the map ambiguous:
The above image with simulated Protanopia
Today there are many modern alternatives that are uniform and color blind accessible, such as Viridis or Inferno from matplotlib. While these linear lightness maps solve many important issues with Jet, their constraints may make them suboptimal for day to day tasks where the requirements are not as stringent.
ViridisInferno
Today we are happy to introduce Turbo, a new colormap that has the desirable properties of Jet while also addressing some of its shortcomings, such as false detail, banding and color blindness ambiguity. Turbo was hand-crafted and fine-tuned to be effective for a variety of visualization tasks. You can find the color map data and usage instructions for Python here and C/C++ here, as well as a polynomial approximation here.

Development
To create the Turbo color map, we created a simple interface that allowed us to interactively adjust the sRGB curves using a 7-knot cubic spline, while comparing the result on a selection of sample images as well as other well known color maps.
Screenshot of the interface used to create and tune Turbo.
This approach provides control while keeping the curve C2 continuous. The resulting color map is not “perceptually linear” in the quantitative sense, but it is more smooth than Jet, without introducing false detail.

Turbo
Jet
Comparison with Common Color Maps
Viridis is a linear color map that is generally recommended when false color is needed because it is pleasant to the eye and it fixes most issues with Jet. Inferno has the same linear properties of Viridis, but is higher contrast, making it better for picking out detail. However, some feel that it can be harsh on the eyes. While this isn’t a concern for publishing, it does affect people’s choice when they must spend extended periods examining visualizations.
TurboJet
ViridisInferno
Because of rapid color and lightness changes, Jet accentuates detail in the background that is less apparent with Viridis and even Inferno. Depending on the data, some detail may be lost entirely to the naked eye. The background in the following images is barely distinguishable with Inferno (which is already punchier than Viridis), but clear with Turbo.
InfernoTurbo
Turbo mimics the lightness profile of Jet, going from low to high back down to low, without banding. As such, its lightness slope is generally double that of Viridis, allowing subtle changes to be more easily seen. This is a valuable feature, since it greatly enhances detail when color can be used to disambiguate the low and high ends.
TurboJet
ViridisInferno
Lightness plots generated by converting the sRGB values to CIECAM02-UCS and displaying the lightness value (J) in greyscale. The black line traces the lightness value from the low end of the color map (left) to the high end (right).
The Viridis and Inferno plots are linear, with Inferno exhibiting a higher slope and over a broader range. Jet’s plot is erratic and peaky, and banding can be seen clearly even in the grayscale image. Turbo has a similar asymmetric profile to Jet with the lows darker than the highs.This is intentional, to make cases where low values appear next to high values more distinct. The curvature in the lower region is also different from the higher region, due to the way blues are perceived in comparison to reds.

Although this low-high-low curve increases detail, it comes at the cost of lightness ambiguity. When rendered in grayscale, the coloration will be ambiguous, since some of the lower values will look identical to higher values. Consequently, Turbo is inappropriate for grayscale printing and for people with the rare case of achromatopsia.

Semantic Layers
When examining disparity maps, it is often desirable to compare values on different sides of the image at a glance. This task is much easier when values can be mentally mapped to a distinct semantic color, such as red or blue. Thus, having more colors helps the estimation ease and accuracy.
TurboJet
ViridisInferno
With Jet and Turbo, it’s easy to see which objects on the left of the frame are at the same depth as objects on the right, even though there is a visual gap in the middle. For example, you can easily spot which sphere on the left is at the same depth as the ring on the right. This is much harder to determine using Viridis or Inferno, which have far fewer distinct colors. Compared to Jet, Turbo is also much more smooth and has no “false layers” due to banding. You can see this improvement more clearly if the incoming values are quantized:
Left: Quantized Turbo colormap. Up to 33 quantized colors remain distinguishable and smooth in both lightness and hue change. Right: Quantized Jet color map. Many neighboring colors appear the same; Yellow and Cyan colors appear brighter than the rest.
Quick Judging
When doing a quick comparison of two images, it’s much easier to judge the differences in color than in lightness (because our attention system prioritizes hue). For example, imagine we have an output image from a depth estimation algorithm beside the ground truth. With Turbo it’s easy to discern whether or not the two are in agreement and which regions may disagree.
“Output” Viridis“Ground Truth” Viridis
“Output” Turbo“Ground Truth” Turbo
In addition, it is easy to estimate quantitative values, since they map to distinguishable and memorable colors.
Diverging Map Use Cases
Although the Turbo color map was designed for sequential use (i.e., values [0-1]), it can be used as a diverging colormap as well, as is needed in difference images, for example. When used this way, zero is green, negative values are shades of blue, and positive values are shades of red. Note, however, that the negative minimum is darker than the positive maximum, so it is not truly balanced.
"Ground Truth" disparity imageEstimated disparity image
Difference Image (ground truth - estimated disparity image), visualized with Turbo
Accessibility for Color Blindness
We tested Turbo using a color blindness simulator and found that for all conditions except Achromatopsia (total color blindness), the map remains distinguishable and smooth. In the case of Achromatopsia, the low and high ends are ambiguous. Since the condition affects 1 in 30,000 individuals (or 0.00003%), Turbo should be usable by 99.997% of the population.
Test Image
ProtanomalyProtanopia
DeuteranomalyDeuteranopia
TritanomalyTritanopia
Blue cone monochromacyAchromatopsia
Conclusion
Turbo is a slot-in replacement for Jet, and is intended for day-to-day tasks where perceptual uniformity is not critical, but one still wants a high contrast, smooth visualization of the underlying data. It can be used as a sequential as well as a diverging map, making it a good all-around map to have in the toolbox. You can find the color map data and usage instructions for Python here and for C/C++ here. There is also a polynomial approximation here, for cases where a look-up table may not be desirable.Our team uses it for visualizing disparity maps, error maps, and various other scalar quantities, and we hope you’ll find it useful as well.

Acknowledgements
Ambrus Csaszar stared at many color ramps with me in order to pick the right tradeoffs between uniformity and detail accentuation. Christian Haene integrated the map into our team’s tools, which caused wide usage and thus spurred further improvements. Matthias Kramm and Ruofei Du came up with closed form approximations.

Source: Google AI Blog


On-Device, Real-Time Hand Tracking with MediaPipe



The ability to perceive the shape and motion of hands can be a vital component in improving the user experience across a variety of technological domains and platforms. For example, it can form the basis for sign language understanding and hand gesture control, and can also enable the overlay of digital content and information on top of the physical world in augmented reality. While coming naturally to people, robust real-time hand perception is a decidedly challenging computer vision task, as hands often occlude themselves or each other (e.g. finger/palm occlusions and hand shakes) and lack high contrast patterns.

Today we are announcing the release of a new approach to hand perception, which we previewed CVPR 2019 in June, implemented in MediaPipe—an open source cross platform framework for building pipelines to process perceptual data of different modalities, such as video and audio. This approach provides high-fidelity hand and finger tracking by employing machine learning (ML) to infer 21 3D keypoints of a hand from just a single frame. Whereas current state-of-the-art approaches rely primarily on powerful desktop environments for inference, our method achieves real-time performance on a mobile phone, and even scales to multiple hands. We hope that providing this hand perception functionality to the wider research and development community will result in an emergence of creative use cases, stimulating new applications and new research avenues.
3D hand perception in real-time on a mobile phone via MediaPipe. Our solution uses machine learning to compute 21 3D keypoints of a hand from a video frame. Depth is indicated in grayscale.
An ML Pipeline for Hand Tracking and Gesture Recognition
Our hand tracking solution utilizes an ML pipeline consisting of several models working together:
  • A palm detector model (called BlazePalm) that operates on the full image and returns an oriented hand bounding box.
  • A hand landmark model that operates on the cropped image region defined by the palm detector and returns high fidelity 3D hand keypoints.
  • A gesture recognizer that classifies the previously computed keypoint configuration into a discrete set of gestures.
This architecture is similar to that employed by our recently published face mesh ML pipeline and that others have used for pose estimation. Providing the accurately cropped palm image to the hand landmark model drastically reduces the need for data augmentation (e.g. rotations, translation and scale) and instead allows the network to dedicate most of its capacity towards coordinate prediction accuracy.
Hand perception pipeline overview.
BlazePalm: Realtime Hand/Palm Detection
To detect initial hand locations, we employ a single-shot detector model called BlazePalm, optimized for mobile real-time uses in a manner similar to BlazeFace, which is also available in MediaPipe. Detecting hands is a decidedly complex task: our model has to work across a variety of hand sizes with a large scale span (~20x) relative to the image frame and be able to detect occluded and self-occluded hands. Whereas faces have high contrast patterns, e.g., in the eye and mouth region, the lack of such features in hands makes it comparatively difficult to detect them reliably from their visual features alone. Instead, providing additional context, like arm, body, or person features, aids accurate hand localization.

Our solution addresses the above challenges using different strategies. First, we train a palm detector instead of a hand detector, since estimating bounding boxes of rigid objects like palms and fists is significantly simpler than detecting hands with articulated fingers. In addition, as palms are smaller objects, the non-maximum suppression algorithm works well even for two-hand self-occlusion cases, like handshakes. Moreover, palms can be modelled using square bounding boxes (anchors in ML terminology) ignoring other aspect ratios, and therefore reducing the number of anchors by a factor of 3-5. Second, an encoder-decoder feature extractor is used for bigger scene context awareness even for small objects (similar to the RetinaNet approach). Lastly, we minimize the focal loss during training to support a large amount of anchors resulting from the high scale variance.

With the above techniques, we achieve an average precision of 95.7% in palm detection. Using a regular cross entropy loss and no decoder gives a baseline of just 86.22%.

Hand Landmark Model
After the palm detection over the whole image our subsequent hand landmark model performs precise keypoint localization of 21 3D hand-knuckle coordinates inside the detected hand regions via regression, that is direct coordinate prediction. The model learns a consistent internal hand pose representation and is robust even to partially visible hands and self-occlusions.

To obtain ground truth data, we have manually annotated ~30K real-world images with 21 3D coordinates, as shown below (we take Z-value from image depth map, if it exists per corresponding coordinate). To better cover the possible hand poses and provide additional supervision on the nature of hand geometry, we also render a high-quality synthetic hand model over various backgrounds and map it to the corresponding 3D coordinates.
Top: Aligned hand crops passed to the tracking network with ground truth annotation. Bottom: Rendered synthetic hand images with ground truth annotation
However, purely synthetic data poorly generalizes to the in-the-wild domain. To overcome this problem, we utilize a mixed training schema. A high-level model training diagram is presented in the following figure.
Mixed training schema for hand tracking network. Cropped real-world photos and rendered synthetic images are used as input to predict 21 3D keypoints.
The table below summarizes regression accuracy depending on the nature of the training data. Using both synthetic and real world data results in a significant performance boost.

Mean regression error
Dataset normalized by palm size
Only real-world 16.1 %
Only rendered synthetic 25.7 %
Mixed real-world + synthetic 13.4 %

Gesture Recognition
On top of the predicted hand skeleton, we apply a simple algorithm to derive the gestures. First, the state of each finger, e.g. bent or straight, is determined by the accumulated angles of joints. Then we map the set of finger states to a set of pre-defined gestures. This straightforward yet effective technique allows us to estimate basic static gestures with reasonable quality. The existing pipeline supports counting gestures from multiple cultures, e.g. American, European, and Chinese, and various hand signs including “Thumb up”, closed fist, “OK”, “Rock”, and “Spiderman”.

Implementation via MediaPipe
With MediaPipe, this perception pipeline can be built as a directed graph of modular components, called Calculators. Mediapipe comes with an extendable set of Calculators to solve tasks like model inference, media processing algorithms, and data transformations across a wide variety of devices and platforms. Individual calculators like cropping, rendering and neural network computations can be performed exclusively on the GPU. For example, we employ TFLite GPU inference on most modern phones.

Our MediaPipe graph for hand tracking is shown below. The graph consists of two subgraphs—one for hand detection and one for hand keypoints (i.e., landmark) computation. One key optimization MediaPipe provides is that the palm detector is only run as necessary (fairly infrequently), saving significant computation time. We achieve this by inferring the hand location in the subsequent video frames from the computed hand key points in the current frame, eliminating the need to run the palm detector over each frame. For robustness, the hand tracker model outputs an additional scalar capturing the confidence that a hand is present and reasonably aligned in the input crop. Only when the confidence falls below a certain threshold is the hand detection model reapplied to the whole frame.
The hand landmark model’s output (REJECT_HAND_FLAG) controls when the hand detection model is triggered. This behavior is achieved by MediaPipe’s powerful synchronization building blocks, resulting in high performance and optimal throughput of the ML pipeline.
A highly efficient ML solution that runs in real-time and across a variety of different platforms and form factors involves significantly more complexities than what the above simplified description captures. To this end, we are open sourcing the above hand tracking and gesture recognition pipeline in the MediaPipe framework, accompanied with the relevant end-to-end usage scenario and source code, here. This provides researchers and developers with a complete stack for experimentation and prototyping of novel ideas based on our model.

Future Directions
We plan to extend this technology with more robust and stable tracking, enlarge the amount of gestures we can reliably detect, and support dynamic gestures unfolding in time. We believe that publishing this technology can give an impulse to new creative ideas and applications by the members of the research and developer community at large. We are excited to see what you can build with it!
Acknowledgements
Special thanks to all our team members who worked on the tech with us: Andrey Vakunov, Andrei Tkachenka, Yury Kartynnik, Artsiom Ablavatski, Ivan Grishchenko, Kanstantsin Sokal‎, Mogan Shieh, Ming Guang Yong, Anastasia Tkach, Jonathan Taylor, Sean Fanello, Sofien Bouaziz, Juhyun Lee‎, Chris McClanahan, Jiuqiang Tang‎, Esha Uboweja‎, Hadon Nash‎, Camillo Lugaresi, Michael Hays, Chuo-Ling Chang, Matsvei Zhdanovich and Matthias Grundmann.

Source: Google AI Blog


Joint Speech Recognition and Speaker Diarization via Sequence Transduction



Being able to recognize “who said what,” or speaker diarization, is a critical step in understanding audio of human dialog through automated means. For instance, in a medical conversation between doctors and patients, “Yes” uttered by a patient in response to “Have you been taking your heart medications regularly?” has a substantially different implication than a rhetorical “Yes?” from a physician.

Conventional speaker diarization (SD) systems use two stages, the first of which detects changes in the acoustic spectrum to determine when the speakers in a conversation change, and the second of which identifies individual speakers across the conversation. This basic multi-stage approach is almost two decades old, and during that time only the speaker change detection component has improved.

With the recent development of a novel neural network model—the recurrent neural network transducer (RNN-T)—we now have a suitable architecture to improve the performance of speaker diarization addressing some of the limitations of the previous diarization system we presented recently. As reported in our recent paper, “Joint Speech Recognition and Speaker Diarization via Sequence Transduction,” to be presented at Interspeech 2019, we have developed an RNN-T based speaker diarization system and have demonstrated a breakthrough in performance from about 20% to 2% in word diarization error rate—a factor of 10 improvement.

Conventional Speaker Diarization Systems
Conventional speaker diarization systems rely on differences in how people sound acoustically to distinguish the speakers in the conversations. While male and female speakers can be identified relatively easily from their pitch using simple acoustic models (e.g., Gaussian mixture models) in a single stage, speaker diarization systems use a multi-stage approach to distinguish between speakers having potentially similar pitch. First, a change detection algorithm breaks up the conversation into homogeneous segments, hopefully containing only a single speaker, based upon detected vocal characteristics. Then, deep learning models are employed to map segments from each speaker to an embedding vector. Finally, in a clustering stage, these embeddings are grouped together to keep track of the same speaker across the conversation.

In practice, the speaker diarization system runs in parallel to the automatic speech recognition (ASR) system and the outputs of the two systems are combined to attribute speaker labels to the recognized words.
Conventional speaker diarization system infers speaker labels in the acoustic domain and then overlays the speaker labels on the words generated by a separate ASR system.
There are several limitations with this approach that have hindered progress in this field. First, the conversation needs to be broken up into segments that only contain speech from one speaker. Otherwise, the embedding will not accurately represent the speaker. In practice, however, the change detection algorithm is imperfect, resulting in segments that may contain multiple speakers. Second, the clustering stage requires that the number of speakers be known and is particularly sensitive to the accuracy of this input. Third, the system needs to make a very difficult trade-off between the segment size over which the voice signatures are estimated and the desired model accuracy. The longer the segment, the better the quality of the voice signature, since the model has more information about the speaker. This comes at the risk of attributing short interjections to the wrong speaker, which could have very high consequences, for example, in the context of processing a clinical or financial conversation where affirmation or negation needs to be tracked accurately. Finally, conventional speaker diarization systems do not have an easy mechanism to take advantage of linguistic cues that are particularly prominent in many natural conversations. An utterance, such as “How often have you been taking the medication?” in a clinical conversation is most likely uttered by a medical provider, not a patient. Likewise, the utterance, “When should we turn in the homework?” is most likely uttered by a student, not a teacher. Linguistic cues also signal high probability of changes in speaker turns, for example, after a question.

There are a few exceptions to the conventional speaker diarization system, but one such exception was reported in our recent blog post. In that work, the hidden states of the recurrent neural network (RNN) tracked the speakers, circumventing the weakness of the clustering stage. Our approach takes a different approach and incorporates linguistic cues, as well.

An Integrated Speech Recognition and Speaker Diarization System
We developed a novel and simple model that not only combines acoustic and linguistic cues seamlessly, but also combines speaker diarization and speech recognition into one system. The integrated model does not degrade the speech recognition performance significantly compared to an equivalent recognition only system.

The key insight in our work was to recognize that the RNN-T architecture is well-suited to integrate acoustic and linguistic cues. The RNN-T model consists of three different networks: (1) a transcription network (or encoder) that maps the acoustic frames to a latent representation, (2) a prediction network that predicts the next target label given the previous target labels, and (3) a joint network that combines the output of the previous two networks and generates a probability distribution over the set of output labels at that time step. Note, there is a feedback loop in the architecture (diagram below) where previously recognized words are fed back as input, and this allows the RNN-T model to incorporate linguistic cues, such as the end of a question.
An integrated speech recognition and speaker diarization system where the system jointly infers who spoke when and what.
Training the RNN-T model on accelerators like graphical processing units (GPU) or tensor processing units (TPU) is non-trivial as computation of the loss function requires running the forward-backward algorithm, which includes all possible alignments of the input and the output sequences. This issue was addressed recently in a TPU friendly implementation of the forward-backward algorithm, which recasts the problem as a sequence of matrix multiplications. We also took advantage of an efficient implementation of the RNN-T loss in TensorFlow that allowed quick iterations of model development and trained a very deep network.

The integrated model can be trained just like a speech recognition system. The reference transcripts for training contain words spoken by a speaker followed by a tag that defines the role of the speaker. For example, “When is the homework due?” ≺student≻, “I expect you to turn them in tomorrow before class,” ≺teacher≻. Once the model is trained with examples of audio and corresponding reference transcripts, a user can feed in the recording of the conversation and expect to see an output in a similar form. Our analyses show that improvements from the RNN-T system impact all categories of errors, including short speaker turns, splitting at the word boundaries, incorrect speaker assignment in the presence of overlapping speech, and poor audio quality. Moreover, the RNN-T system exhibited consistent performance across conversation with substantially lower variance in average error rate per conversation compared to the conventional system.

A comparison of errors committed by the conventional system vs. the RNN-T system, as categorized by human annotators.
Furthermore, this integrated model can predict other labels necessary for generating more reader-friendly ASR transcripts. For example, we have been able to successfully improve our transcripts with punctuation and capitalization symbols using the appropriately matched training data. Our outputs have lower punctuation and capitalization errors than our previous models that were separately trained and added as a post-processing step after ASR.

This model has now become a standard component in our project on understanding medical conversations and is also being adopted more widely in our non-medical speech services.

Acknowledgements
We would like to thank Hagen Soltau without whose contributions this work would not have been possible. This work was performed in collaboration with Google Brain and Speech teams.

Source: Google AI Blog


Project Euphonia’s Personalized Speech Recognition for Non-Standard Speech



The utility of technology is dependent on its accessibility. One key component of accessibility is automatic speech recognition (ASR), which can greatly improve the ability of those with speech impairments to interact with every-day smart devices. However, ASR systems are most often trained from 'typical' speech, which means that underrepresented groups, such as those with speech impairments or heavy accents, don't experience the same degree of utility. For example, amyotrophic lateral sclerosis (ALS) is a disease that can adversely affect a person’s speech—about 25% of people with ALS experiencing slurred speech as their first symptom. In addition, most people with ALS eventually lose the ability to walk, so being able to interact with automated devices from a distance can be very important. Yet current state-of-the-art ASR models can yield high word error rates (WER) for speakers with only a moderate speech impairment from ALS, effectively barring access to ASR reliant technologies.

In “Personalizing ASR for Dysarthric and Accented Speech with Limited Data,” to be presented at Interspeech 2019, we describe some of the research behind Project Euphonia, an ASR platform that performs speech-to-text transcription. This work presents an approach to improve ASR for people with ALS that may also be applicable to many other types of non-standard speech. Using a two-step training approach that starts with a baseline “standard” corpus and then fine-tunes the training with a personalized speech dataset, we have demonstrated significant improvements for speakers with atypical speech over current state-of-the-art models.

A Two-Phased Approach to Training
In order to create ASR models that work on non-standard speech, one needs to overcome two challenges. The first is that within a particular class of atypical speech, be it a regional accent or a speech impairment, for example, individuals can exhibit very different ways of speaking. Our approach deals with this sub-group heterogeneity by training the ASR model in two phases. We start with a high-quality ASR model trained on thousands of hours of standard speech and then we fine-tune parts of the model to an individual with non-standard speech. This approach is similar to that of Parrotron: both systems use end-to-end neural networks to help improve communication and accessibility, but Parrotron focuses exclusively on speech-to-speech, where a person’s speech is converted directly into synthesized speech, rather than text.

The second challenge arises from the difficulty in collecting enough data to train a state-of-the-art recognizer for individuals. Typical speech recognizers are trained on thousands of hours of speech from many different speakers. Acquiring this much data from a single speaker is nearly impossible, especially if the speaker may experience exhaustion from speaking due to a medical condition. Our approach overcomes this issue by first training a base model on a large corpus of typical speech, and then training a personalized model using a much smaller dataset with the targeted non-standard speech characteristics.

The Neural Network Architecture
When developing the models used for training data on atypical speech, we explored two different neural architectures. The first is the RNN-Transducer (RNN-T), a neural network architecture consisting of encoder and decoder networks that has shown good results on numerous ASR tasks. The encoder is bidirectional (i.e., it looks at the entire sentence at once in order to provide context), and thus it requires the entire audio sample to perform speech recognition.

The other architecture we explored was Listen, Attend, and Spell (LAS), which is an attention-based, sequence-to-sequence model that maps sequences of acoustic properties to sequences of languages. This model uses an encoder to convert the sequence of acoustic frames to a sequence of internal representations, and a decoder to convert the sequence of internal representations to linguistic output. The network produces “word pieces”, which are a linguistic representation between graphemes and words.
Comparison of the RNN-Transducer (left) and Listen, Attend, Spell (right) architectures. From Prabhavalkar et al. 2017.
We experimented with fine-tuning the state-of-the-art RNN-T and LAS base models on two types of non-standard speech. In partnership with the ALS Therapy Development Institute, we first collected about 36 hours of audio from 67 speakers who have ALS. The participants recorded themselves on their home computers using custom software while they read sentences from a very restricted language domain. Many phrases were single sentences with simple grammatical structure (e.g., “What time is the basketball game on tonight?”). This is in contrast with unrestricted language domains, which include domain-specific vocabulary (e.g., science talks) and complex language structure (e.g., a debate). The recordings did not include many of the filler words common in normal speech, such as “um” and “uh”.

We also tested accented speech, using the open source L2 Arctic dataset of non-native speech, which consists of 20 speakers with approximately 1 hour of speech per speaker. Each speaker recorded a set of 1150 utterances from the CMU Arctic prompts.

AudioEuphonia ModelStandard Speech Model
Did I have anything to say about it?Dictatorship angels to think about it
Come right back pleaseCameras object
Let’s try that againIt extracts
Turn it down a little bit pleaseTurning down a little bit please
The audio (left) are recordings of a speaker with ALS. The text transcriptions are output from the Euphonia model (center) and the Standard Speech model (right). Incorrectly transcribed text is underlined.
Results
The absolute word error rates on the language-restricted test set is shown below. There is an improvement over the baseline model for very non-standard speech (heavy accents and ALS speech below 3 on the ALS Functional Rating Scale) and moderate improvements in ALS speech that is similar to typical speech. The relative difference between the base model and the fine-tuned model demonstrates that the majority of the improvement comes from the fine-tuning process, except in the case of the RNN-T on the Arctic dataset, where the RNN-T baseline is already strong.
1 Non-native English speech from the L2-Arctic dataset.
2 Low FRS (ALS Functional Rating Scale) speech; intelligible with repeating (FRS 2); Speech combined with non-vocal communication (FRS 1).
3 FRS 3; detectable speech disturbance.
The RNN-T model achieved 91% of the improvement by fine-tuning just two layers, most of which are close to the input. On the accented dataset, fine-tuning the same two layers achieved 86% of the relative improvement compared to fine-tuning the entire network. This is consistent with previous speech work.

Most of the performance gains were achieved early in training. The models we trained were tested on a relatively limited domain of vocabulary and linguistic complexity, so the performance numbers are not necessarily related to how well the models perform on more general tasks. We hope that just fine-tuning part of the network allows it to retain the acoustic and linguistic information from the general speech model, while needing minimal modifications to adapt to a single new speaker. Future work will test this hypothesis.
Low FRS corresponds to the ALS speakers with low intelligibility (FRS 2, 1), while high FRS corresponds to ALS speakers with less severely impacted speech (FRS 3).
Understanding Model Behavior
To better understand how our models improved after fine-tuning, we looked at the pattern of phoneme mistakes. We started by comparing the distribution of phoneme mistakes made by the base ASR model on standard speech to the mistakes made on ALS speech. The SAMPA phonemes with the five largest differences between the ALS data and standard speech are p, U, f, k, and Z, which account for 20% of the deletion mistakes. Similarly, the n and m phonemes together account for 17% of the insertion / substitution mistakes. The same analysis on our fine-tuned models verifies that the unrecognized phoneme distribution is more similar to that of standard speech.

Our analysis shows that there are two aspects to every mistake: which phoneme the system doesn’t understand, and which phoneme the system thinks was said. Imagine having two systems with identical accuracy: one system always thinks that the f phoneme is actually the g phoneme, while another doesn't know what the f phoneme is and randomly guesses. These two systems will have identical performance and identical distributions of phoneme mistakes, but very different distributions of the predicted phoneme when a mistake is made. Surprisingly, ASR mistakes on ALS speech are far more similar to regular speech mistakes after Euphonia fine-tuning.
Deletion / substitution mistakes per SAMPA phoneme on ALS speech before fine-tuning, ALS speech after fine-tuning, and on typical speech (Librispeech dataset).
Future Work
In the future, we intend to explore additional techniques that can be helpful in the low data regime. We also hope to use phoneme mistakes to weight certain examples during training, or to pick training sentences for people with ALS to record that contain the most common phoneme mistakes. We would like to explore pooling data from multiple speakers with similar conditions.

We hope that continued research in this area will help voice interfaces become accessible to more people, especially those who need it most. One key component to this is collecting data. Anyone 18 or older can help us build better personalized models by donating audio data. If you’re interested, you can fill out this form to allow Google to contact you.

Acknowledgements
This work would not have been possible without the extraordinary effort and support of the ALS Therapy Development Institute and the ALS community, especially Fernando Vieira, Maeve McNally, Taylor Charbonneau, Melissa Nollstadt, and the individuals with ALS who kindly and patiently volunteered their audio. This work builds on the pioneering advances in speech recognition made by Google's speech team, in particular the recent development and deployment of end-to-end speech recognition models. We are grateful to the Google speech team for advice and collaboration, particularly to Anshuman Tripathi and Hasim Sak who guided us in training the initial models. We’d also like to thank Oran Lang, Omry Tuval, Michael Brenner, Julie Cattiau, Tara Sainath, Ding Zhao, Qiao Liang, Chung-Cheng Chiu, Dan Liebling, Ron Weiss, Anjuli Kannan, Dimitri Kanevsky, Ryan He, Gabor Simko, Benjamin Lee, Françoise Beaufays, Khe Chai Sim, Jimmy Tobin, Chet Gnegy, Jacqueline Huang, Ye Jia, Yu Zhang, Yonghui Wu, Michelle Ramanovich, Rus Heywood, Katrin Tomanek, Bob MacDonald, Pan-Pan Jiang, Ronnie Maor, Rif A. Saurous, Trevor Strohman, Dick Lyon, Avinatan Hassidim, Philip Nelson, and Yossi Matias for their technical contributions and project guidance.

Source: Google AI Blog


Video Understanding Using Temporal Cycle-Consistency Learning



In the last few years there has been great progress in the field of video understanding. For example, supervised learning and powerful deep learning models can be used to classify a number of possible actions in videos, summarizing the entire clip with a single label. However, there exist many scenarios in which we need more than just one label for the entire clip. For example, if a robot is pouring water into a cup, simply recognizing the action of “pouring a liquid” is insufficient to predict when the water will overflow. For that, it is necessary to track frame-by-frame the amount of water in the cup as it is being filled. Similarly, a baseball coach who is comparing stances of pitchers may want to retrieve video frames from the precise moment that the ball leaves the pitchers’ hands. Such applications require models to understand each frame of a video.

However, applying supervised learning to understand each individual frame in a video is expensive, since per-frame labels in videos of the action of interest are needed. This requires that annotators apply fine-grained labels to videos by manually adding unambiguous labels to every frame in each video. Only then can the model be trained, and only on a single action. Training on new actions requires the process to be repeated. With the increasing demand for fine-grained labeling, necessary for applications ranging from robotics to sports analytics, this makes the need for scalable learning algorithms that can understand videos without the tedious labeling process increasingly pertinent.

We propose a potential solution using a self-supervised learning method called Temporal Cycle-Consistency Learning (TCC). This novel approach uses correspondences between examples of similar sequential processes to learn representations particularly well-suited for fine-grained temporal understanding of videos. We are also releasing our TCC codebase to enable end-users to apply our self-supervised learning algorithm to new and novel applications.

Representation Learning Using TCC
A plant growing from a seedling to a tree; the daily routine of getting up, going to work and coming back home; or a person pouring themselves a glass of water are all examples of events that happen in a particular order. Videos capturing such processes provide temporal correspondences across multiple instances of the same process. For example, when pouring a drink one could be reaching for a teapot, a bottle of wine, or a glass of water to pour from. Key moments are common to all pouring videos (e.g., the first touch to the container or the container being lifted from the ground) and exist independent of many varying factors, such as visual changes in viewpoint, scale, container style, or the speed of the event. TCC attempts to find such correspondences across videos of the same action by leveraging the principle of cycle-consistency, which has been applied successfully in many problems in computer vision, to learn useful visual representations by aligning videos.

The objective of this training algorithm is to learn a frame encoder, using any network architecture that processes images, such as ResNet. To do so, we pass all frames of the videos to be aligned through the encoder to produce their corresponding embeddings. We then select two videos for TCC learning, say video 1 (the reference video) and video 2. A reference frame is chosen from video 1 and its nearest neighbor frame (NN2) from video 2 is found in the embedding space (not pixel space). We then cycle back by finding the nearest neighbor of NN2 in video 1, which we call NN1. If the representations are cycle-consistent, then the nearest neighbor frame in video 1 (NN1) should refer back to the starting reference frame.
We train the embedder using the distance between the starting reference frame and NN1 as the training signal. As training proceeds, the embeddings improve and reduce the cycle-consistency loss by developing a semantic understanding of each video frame in the context of the action being performed.
Using TCC, we learn embeddings with temporally fine-grained understanding of an action by aligning related videos.
What Does TCC Learn?
In the following figure, we show a model trained using TCC on videos from the Penn Action Dataset of people performing squat exercises. Each point on the left corresponds to frame embeddings, with the highlighted points tracking the embedding of the current video frame. Notice how the embeddings move collectively in spite of many differences in pose, lighting, body and object type. TCC embeddings encode the different phases of squatting without being provided explicit labels.
Right: Input videos of people performing a squat exercise. The video on the top left is the reference. The other videos show nearest neighbor frames (in the TCC embedding space) from other videos of people doing squats. Left: The corresponding frame embeddings move as the action is performed.
Applications of TCC
The learned per-frame embeddings enable an array of interesting applications:
  • Few-shot action phase classification
    When few labeled videos are available for training, the few-shot scenario, TCC performs very well. In fact, TCC can classify the phases of different actions with as few as a single labeled video. In the next figure we compare to other supervised and self-supervised learning approaches in the few-shot setting. We find that supervised learning requires about 50 videos with each frame labeled to achieve the same accuracy that self-supervised methods achieve with just one fully labeled video.
    Comparison of self-supervised and supervised learning for few-shot action phase classification.
  • Unsupervised video alignment
    Aligning or synchronizing videos manually becomes prohibitively difficult as the number of videos increases. Using TCC, many videos can be aligned by selecting the nearest neighbor to each frame in a reference video, without the need for additional labels, as demonstrated in the figure below.
    Results of unsupervised video alignment on videos of people pitching baseball using the distance between frames in the TCC space. The reference video used for alignment is shown in the upper left panel.
  • Label/modality transfer between videos
    Just as TCC finds similar frames by using a nearest neighbor search in the embedding space, it can transfer metadata associated with any frame in one video to its matching frame in another video. This metadata can be in the form of temporal semantic labels or other modalities, such as sound or text. In the video below we show two examples where we can transfer the sound of liquid being poured into a cup from one video to another.
  • Per-frame Retrieval
    With TCC, each frame in a video can be used as a query for retrieval of similar frames by looking up the nearest neighbors in the learned embedding space. The embeddings are powerful enough to differentiate between frames that look quite similar, such as frames just before or after the release of a bowling ball.
    We can perform retrieval from videos on a per-frame basis, i.e., any frame can be used to look up similar frames in a large collection of videos. The retrieved nearest neighbors show that the model captures fine-grained differences in the scene.
Release
We are releasing our codebase, which includes implementations of a number of state-of-the-art self-supervised learning methods, including TCC. This codebase will be useful for researchers working on video understanding, as well as artists looking to use machine learning to align videos to create mosaics of people, animals, and objects moving synchronously.

Acknowledgements
This is joint work with Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. The authors would like to thank Alexandre Passos, Allen Lavoie, Anelia Angelova, Bryan Seybold, Priya Gupta, Relja Arandjelović, Sergio Guadarrama, Sourish Chaudhuri, and Vincent Vanhoucke for their help with this project. The videos used in this project come from the PennAction dataset. We thank the creators of PennAction for curating such an interesting dataset.

Source: Google AI Blog


EfficientNet-EdgeTPU: Creating Accelerator-Optimized Neural Networks with AutoML



For several decades, computer processors have doubled their performance every couple of years by reducing the size of the transistors inside each chip, as described by Moore’s Law. As reducing transistor size becomes more and more difficult, there is a renewed focus in the industry on developing domain-specific architectures — such as hardware accelerators — to continue advancing computational power. This is especially true for machine learning, where efforts are aimed at building specialized architectures for neural network (NN) acceleration. Ironically, while there has been a steady proliferation of these architectures in data centers and on edge computing platforms, the NNs that run on them are rarely customized to take advantage of the underlying hardware.

Today, we are happy to announce the release of EfficientNet-EdgeTPU, a family of image classification models derived from EfficientNets, but customized to run optimally on Google’s Edge TPU, a power-efficient hardware accelerator available to developers through the Coral Dev Board and a USB Accelerator. Through such model customizations, the Edge TPU is able to provide real-time image classification performance while simultaneously achieving accuracies typically seen only when running much larger, compute-heavy models in data centers.

Using AutoML to customize EfficientNets for Edge TPU
EfficientNets have been shown to achieve state-of-the-art accuracy in image classification tasks while significantly reducing the model size and computational complexity. To build EfficientNets designed to leverage the Edge TPU’s accelerator architecture, we invoked the AutoML MNAS framework and augmented the original EfficientNet’s neural network architecture search space with building blocks that execute efficiently on the Edge TPU (discussed below). We also built and integrated a “latency predictor” module that provides an estimate of the model latency when executing on the Edge TPU, by running the models on a cycle-accurate architectural simulator. The AutoML MNAS controller implements a reinforcement learning algorithm to search this space while attempting to maximize the reward, which is a joint function of the predicted latency and model accuracy. From past experience, we know that Edge TPU’s power efficiency and performance tend to be maximized when the model fits within its on-chip memory. Hence we also modified the reward function to generate a higher reward for models that satisfy this constraint.
Overall AutoML flow for designing customized EfficientNet-EdgeTPU models.
Search Space Design
When performing the architecture search described above, one must consider that EfficientNets rely primarily on depthwise-separable convolutions, a type of neural network block that factorizes a regular convolution to reduce the number of parameters as well as the amount of computations. However, for certain configurations, a regular convolution utilizes the Edge TPU architecture more efficiently and executes faster, despite the much larger amount of compute. While it is possible, albeit tedious, to manually craft a network that uses an optimal combination of the different building blocks, augmenting the AutoML search space with these accelerator-optimal blocks is a more scalable approach.
A regular 3x3 convolution (right) has more compute (multiply-and-accumulate (mac) operations) than an depthwise-separable convolution (left), but for certain input/output shapes, executes faster on Edge TPU due to ~3x more effective hardware utilization.
In addition, removing certain operations from the search space that require modifications to the Edge TPU compiler to fully support, such swish non-linearity and squeeze-and-excitation block, naturally leads to models that are readily ported to the Edge TPU hardware. These operations tend to improve model quality slightly, so by eliminating them from the search space, we have effectively instructed AutoML to discover alternate network architectures that may compensate for any potential loss in quality.

Model Performance
The neural architecture search (NAS) described above produced a baseline model, EfficientNet-EdgeTPU-S, which is subsequently scaled up using EfficientNet's compound scaling method to produce the -M and -L models. The compound scaling approach selects an optimal combination of input image resolution scaling, network width, and depth scaling to construct larger, more accurate models. The -M, and -L models achieve higher accuracy at the cost of increased latency as shown in the figure below.
EfficientNet-EdgeTPU-S/M/L models achieve better latency and accuracy than existing EfficientNets (B1), ResNet, and Inception by specializing the network architecture for Edge TPU hardware. In particular, our EfficientNet-EdgeTPU-S achieves higher accuracy, yet runs 10x faster than ResNet-50.
Interestingly, the NAS-generated model employs the regular convolution quite extensively in the initial part of the network where the depthwise-separable convolution tends to be less effective than the regular convolution when executed on the accelerator. This clearly highlights the fact that trade-offs usually made while optimizing models for general purpose CPUs (reducing the total number of operations, for example) are not necessarily optimal for hardware accelerators. Also, these models achieve high accuracy even without the use of esoteric operations. Comparing with the other image classification models such as Inception-resnet-v2 and Resnet50, EfficientNet-EdgeTPU models are not only more accurate, but also run faster on Edge TPUs.

This work represents a first experiment in building accelerator-optimized models using AutoML. The AutoML-based model customization can be extended to not only a wide range of hardware accelerators, but also to several different applications that rely on neural networks.

From Cloud TPU training to Edge TPU deployment
We have released the training code and pretrained models for EfficientNet-EdgeTPU on our github repository. We employ tensorflow’s post-training quantization tool to convert a floating-point trained model to an Edge TPU-compatible integer-quantized model. For these models, the post-training quantization works remarkably well and produces only a very slight loss in accuracy (~0.5%). The script for exporting the quantized model from a training checkpoint can be found here. For an update on the Coral platform, see this post on the Google Developer’s Blog, and for full reference materials and detailed instructions, please refer to the Coral website.

Acknowledgements
Special thanks to Quoc Le, Hongkun Yu, Yunlu Li, Ruoming Pang, and Vijay Vasudevan from the Google Brain team; Bo Wu, Vikram Tank, and Ajay Nair from the Google Coral team; Han Vanholder, Ravi Narayanaswami, John Joseph, Dong Hyuk Woo, Raksit Ashok, Jason Jong Kyu Park, Jack Liu, Mohammadali Ghodrat, Cao Gao, Berkin Akin, Liang-Yun Wang, Chirag Gandhi, and Dongdong Li from the Google Edge TPU team.

Source: Google AI Blog


An Interactive, Automated 3D Reconstruction of a Fly Brain



The goal of connectomics research is to map the brain’s "wiring diagram" in order to understand how the nervous system works. A primary target of recent work is the brain of the fruit fly (Drosophila melanogaster), which is a well-established research animal in biology. Eight Nobel Prizes have been awarded for fruit fly research that has led to advances in molecular biology, genetics, and neuroscience. An important advantage of flies is their size: Drosophila brains are relatively small (one hundred thousand neurons) compared to, for example, a mouse brain (one hundred million neurons) or a human brain (one hundred billion neurons). This makes fly brains easier to study as a complete circuit.

Today, in collaboration with the Howard Hughes Medical Institute (HHMI) Janelia Research Campus and Cambridge University, we are excited to publish “Automated Reconstruction of a Serial-Section EM Drosophila Brain with Flood-Filling Networks and Local Realignment”, a new research paper that presents the automated reconstruction of an entire fruit fly brain. We are also making the full results available for anyone to download or to browse online using an interactive, 3D interface we developed called Neuroglancer.
A 40-trillion pixel fly brain reconstruction, open to anyone for interactive viewing. Bottom right: smaller datasets that Google AI analyzed in publications in 2016 and 2018.
Automated Reconstruction of 40 Trillion Pixels
Our collaborators at HHMI sectioned a fly brain into thousands of ultra-thin 40-nanometer slices, imaged each slice using a transmission electron microscope (resulting in over forty trillion pixels of brain imagery), and then aligned the 2D images into a coherent, 3D image volume of the entire fly brain. Using thousands of Cloud TPUs we then applied Flood-Filling Networks (FFNs), which automatically traced each individual neuron in the fly brain.

While the algorithm generally performed well, we found performance degraded when the alignment was imperfect (image content in consecutive sections was not stable) or when occasionally there were multiple consecutive slices missing due to difficulties associated with the sectioning and imaging process. In order to compensate for these issues we combined FFNs with two new procedures. First, we estimated the slice-to-slice consistency everywhere in the 3D image and then locally stabilized the image content as the FFN traced each neuron. Second, we used a “Segmentation-Enhanced CycleGAN” (SECGAN) to computationally “hallucinate” missing slices in the image volume. SECGANs are a type of generative adversarial network specialized for image segmentation. We found that the FFN was able to trace through locations with multiple missing slices much more robustly when using the SECGAN-hallucinated image data.
Interactive Visualization of the Fly Brain with Neuroglancer
When working with 3D images that contain trillions of pixels and objects with complicated shapes, visualization is both essential and difficult. Inspired by Google’s history of developing new visualization technologies, we designed a new tool that was scalable and powerful, but also accessible to anybody with a web browser that supports WebGL. The result is Neuroglancer, an open-source project (github) that enables viewing of petabyte-scale 3D volumes, and supports many advanced features such as arbitrary-axis cross-sectional reslicing, multi-resolution meshes, and the powerful ability to develop custom analysis workflows via integration with Python. This tool has become heavily used by collaborators at the Allen Institute for Brain Science, Harvard University, HHMI, Max Planck Institute, MIT, Princeton University, and elsewhere.
A recorded demonstration of Neuroglancer. Interactive version available here.
Next Steps
Our collaborators at HHMI and Cambridge University have already begun using this reconstruction to accelerate their studies of learning, memory, and perception in the fly brain. However, the results described above are not yet a true connectome since establishing a connectome requires the identification of synapses. We are working closely with the FlyEM team at Janelia Research Campus to create a highly verified and exhaustive connectome of the fly brain using images acquired with “FIB-SEM” technology.

Acknowledgements
We would like to acknowledge core contributions from Tim Blakely, Viren Jain, Michal Januszewski, Laramie Leavitt, Larry Lindsey, Mike Tyka (Google), as well as Alex Bates, Davi Bock, Greg Jefferis, Feng Li, Mathew Nichols, Eric Perlman, Istvan Taisz, and Zhihao Zheng (Cambridge University, HHMI Janelia, Johns Hopkins University, and University of Vermont).

Source: Google AI Blog


Robust Neural Machine Translation



In recent years, neural machine translation (NMT) using Transformer models has experienced tremendous success. Based on deep neural networks, NMT models are usually trained end-to-end on very large parallel corpora (input/output text pairs) in an entirely data-driven fashion and without the need to impose explicit rules of language.

Despite this huge success, NMT models can be sensitive to minor perturbations of the input, which can manifest as a variety of different errors, such as under-translation, over-translation or mistranslation. For example, given a German sentence, the state-of-the-art NMT model, Transformer, will yield a correct translation.

“Der Sprecher des Untersuchungsausschusses hat angekündigt, vor Gericht zu ziehen, falls sich die geladenen Zeugen weiterhin weigern sollten, eine Aussage zu machen.”

(Machine translation to English: “The spokesman of the Committee of Inquiry has announced that if the witnesses summoned continue to refuse to testify, he will be brought to court.”),

But, when we apply a subtle change to the input sentence, say from geladenen to the synonym vorgeladenen, the translation becomes very different (and in this case, incorrect):

“Der Sprecher des Untersuchungsausschusses hat angekündigt, vor Gericht zu ziehen, falls sich die vorgeladenen Zeugen weiterhin weigern sollten, eine Aussage zu machen.”

(Machine translation to English: “The investigative committee has announced that he will be brought to justice if the witnesses who have been invited continue to refuse to testify.”).

This lack of robustness in NMT models prevents many commercial systems from being applicable to tasks that cannot tolerate this level of instability. Therefore, learning robust translation models is not just desirable, but is often required in many scenarios. Yet, while the robustness of neural networks has been extensively studied in the computer vision community, only a few prior studies on learning robust NMT models can be found in literature.

In “Robust Neural Machine Translation with Doubly Adversarial Inputs” (to appear at ACL 2019), we propose an approach that uses generated adversarial examples to improve the stability of machine translation models against small perturbations in the input. We learn a robust NMT model to directly overcome adversarial examples generated with knowledge of the model and with the intent of distorting the model predictions. We show that this approach improves the performance of the NMT model on standard benchmarks.

Training a Model with AdvGen
An ideal NMT model would generate similar translations for separate inputs that exhibit small differences. The idea behind our approach is to perturb a translation model with adversarial inputs in the hope of improving the model’s robustness. It does this using an algorithm called Adversarial Generation (AdvGen), which generates plausible adversarial examples for perturbing the model and then feeds them back into the model for defensive training. While this method is inspired by the idea of generative adversarial networks (GANs), it does not rely on a discriminator network, but simply applies the adversarial example in training, effectively diversifying and extending the training set.

The first step is to perturb the model using AdvGen. We start by using Transformer to calculate the translation loss based on a source input sentence, a target input sentence and a target output sentence. Then AdvGen randomly selects some words in the source sentence, assuming a uniform distribution. Each word has an associated list of similar words, i.e., candidates that can be used for substitution, from which AdvGen selects the word that is most likely to introduce errors in Transformer output. Then, this generated adversarial sentence is fed back into Transformer, initiating the defense stage.
First, the Transformer model is applied to an input sentence (lower left) and, in conjunction with the target output sentence (above right) and target input sentence (middle right; beginning with the placeholder “<sos>”), the translation loss is calculated. The AdvGen function then takes the source sentence, word selection distribution, word candidates, and the translation loss as inputs to construct an adversarial source example.
During the defend stage, the adversarial sentence is fed back into the Transformer model. Again the translation loss is calculated, but this time using the adversarial source input. Using the same method as above, AdvGen uses the target input sentence, word replacement candidates, the word selection distribution calculated by the attention matrix, and the translation loss to construct an adversarial target example.
In the defense stage, the adversarial source example serves as input to the Transformer model, and the translation loss is calculated. AdvGen then uses the same method as above to generate an adversarial target example from the target input.
Finally, the adversarial sentence is fed back into Transformer and the robustness loss using the adversarial source example, the adversarial target input example and the target sentence is calculated. If the perturbation led to a significant loss, the loss is minimized so that when the model is confronted with similar perturbations, it will not repeat the same mistake. On the other hand, if the perturbation leads to a low loss, nothing happens, indicating that the model can already handle this perturbation.

Model Performance
We demonstrate the effectiveness of our approach by applying it to the standard Chinese-English and English-German translation benchmarks. We observed a notable improvement of 2.8 and 1.6 BLEU points, respectively, compared to the competitive Transformer model, achieving a new state-of-the-art performance.
Comparison of Transformer model (Vaswani et al., 2017) on standard benchmarks.
We then evaluate our model on a noisy dataset, generated using a procedure similar to that described for AdvGen. We take an input clean dataset, such as that used on standard translation benchmarks, and randomly select words for similar word substitution. We find that our model exhibits improved robustness compared to other recent models.
Comparison of Transformer, Miyao et al. and Cheng et al. on artificial noisy inputs.
These results show that our method is able to overcome small perturbations in the input sentence and improve the generalization performance. It outperforms competitive translation models and achieves state-of-the-art translation performance on standard benchmarks. We hope our translation model will serve as a robust building block for improving many downstream tasks, especially when those are sensitive or intolerant to imperfect translation input.

Acknowledgements
This research was conducted by Yong Cheng, Lu Jiang and Wolfgang Macherey. Additional thanks go to our leadership Andrew Moore and Julia (Wenli) Zhu‎.

Source: Google AI Blog