Category: Offsites

Posted by Kihyuk Sohn and Chun-Liang Li, Research Scientists, Google Cloud

Anomaly detection (sometimes called outlier detection or out-of-distribution detection) is one of the most common machine learning applications across many domains, from defect detection in manufacturing to fraudulent transaction detection in finance. It is most often used when it is easy to collect a large amount of known-normal examples but where anomalous data is rare and difficult to find. As such, one-class classification, such as one-class support vector machine (OC-SVM) or support vector data description (SVDD), is particularly relevant to anomaly detection because it assumes the training data are all normal examples, and aims to identify whether an example belongs to the same distribution as the training data. Unfortunately, these classical algorithms do not benefit from the representation learning that makes machine learning so powerful. On the other hand, substantial progress has been made in learning visual representations from unlabeled data via self-supervised learning, including rotation prediction and contrastive learning. As such, combining one-class classifiers with these recent successes in deep representation learning is an under-explored opportunity for the detection of anomalous data.

In “Learning and Evaluating Representations for Deep One-class Classification”, presented at ICLR 2021, we outline a 2-stage framework that makes use of recent progress on self-supervised representation learning and classic one-class algorithms. The algorithm is simple to train and results in state-of-the-art performance on various benchmarks, including CIFAR, f-MNIST, Cat vs Dog and CelebA. We then follow up on this in “CutPaste: Self-Supervised Learning for Anomaly Detection and Localization”, presented at CVPR 2021, in which we propose a new representation learning algorithm under the same framework for a realistic industrial defect detection problem. The framework achieves a new state-of-the-art on the MVTec benchmark.

A Two-Stage Framework for Deep One-Class Classification
While end-to-end learning has demonstrated success in many machine learning problems, including deep learning algorithm designs, such an approach for deep one-class classifiers often suffer from degeneration in which the model outputs the same results regardless of the input.

To combat this, we apply a two stage framework. In the first stage, the model learns deep representations with self-supervision. In the second stage, we adopt one-class classification algorithms, such as OC-SVM or kernel density estimator, using the learned representations from the first stage. This 2-stage algorithm is not only robust to degeneration, but also enables one to build more accurate one-class classifiers. Furthermore, the framework is not limited to specific representation learning and one-class classification algorithms — that is, one can easily plug-and-play different algorithms, which is useful if any advanced approaches are developed.

A deep neural network is trained to generate the representations of input images via self-supervision. We then train one-class classifiers on the learned representations.

Semantic Anomaly Detection
We test the efficacy of our 2-stage framework for anomaly detection by experimenting with two representative self-supervised representation learning algorithms, rotation prediction and contrastive learning.

Rotation prediction refers to a model’s ability to predict the rotated angles of an input image. Due to its promising performance in other computer vision applications, the end-to-end trained rotation prediction network has been widely adopted for one-class classification research. The existing approach typically reuses the built-in rotation prediction classifier for learning representations to conduct anomaly detection, which is suboptimal because those built-in classifiers are not trained for one-class classification.

In contrastive learning, a model learns to pull together representations from transformed versions of the same image, while pushing representations of different images away. During training, as images are drawn from the dataset, each is transformed twice with simple augmentations (e.g., random cropping or color changing). We minimize the distance of the representations from the same image to encourage consistency and maximize the distance between different images. However, usual contrastive learning converges to a solution where all the representations of normal examples are uniformly spread out on a sphere. This is problematic because most of the one-class algorithms determine the outliers by checking the proximity of a tested example to the normal training examples, but when all the normal examples are uniformly distributed in an entire space, outliers will always appear close to some normal examples.

To resolve this, we propose distribution augmentation (DA) for one-class contrastive learning. The idea is that instead of learning representations from the training data only, the model learns from the union of the training data plus augmented training examples, where the augmented examples are considered to be different from the original training data. We employ geometric transformations, such as rotation or horizontal flip, for distribution augmentation. With DA, the training data is no longer uniformly distributed in the representation space because some areas are occupied by the augmented data.

Left: Illustrated examples of perfect uniformity from the standard contrastive learning. Right: The reduced uniformity by the proposed distribution augmentation (DA), where the augmented data occupy the space to avoid the uniform distribution of the inlier examples (blue) throughout the whole sphere.

We evaluate the performance of one-class classification in terms of the area under receiver operating characteristic curve (AUC) on the commonly used datasets in computer vision, including CIFAR10 and CIFAR-100, Fashion MNIST, and Cat vs Dog. Images from one class are given as inliers and those from remaining classes are given as outliers. For example, we see how well cat images are detected as anomalies when dog images are inliers.

	CIFAR-10	CIFAR-100	f-MNIST	Cat v.s. Dog
Ruff et al. (2018)	64.8	–	–	–
Golan and El-Yaniv (2018)	86.0	78.7	93.5	88.8
Bergman and Hoshen (2020)	88.2	–	94.1	–
Hendrycks et al. (2019)	90.1	–	–	–
Huang et al. (2019)	86.6	78.8	93.9	–
2-stage framework: rotation prediction	91.3±0.3	84.1±0.6	95.8±0.3	86.4±0.6
2-stage framework: contrastive (DA)	92.5±0.6	86.5±0.7	94.8±0.3	89.6±0.5

Performance comparison of one-class classification methods. Values are the mean AUCs and their standard deviation over 5 runs. AUC ranges from 0 to 100, where 100 is perfect detection.

Given the suboptimal built-in rotation prediction classifiers typically used for rotation prediction approaches, it’s notable that simply replacing the built-in rotation classifier used in the first stage for learning representations with a one-class classifier at the second stage of the proposed framework significantly boosts the performance, from 86 to 91.3 AUC. More generally, the 2-stage framework achieves state-of-the-art performance on all of the above benchmarks.

With classic OC-SVM, which learns the area boundary of representations of normal examples, the 2-stage framework results in higher performance than existing works as measured by image-level AUC.

Texture Anomaly Detection for Industrial Defect Detection
In many real-world applications of anomaly detection, the anomaly is often defined by localized defects instead of entirely different semantics (i.e., being different in general). For example, the detection of texture anomalies is useful for detecting various kinds of industrial defects.

The examples of semantic anomaly detection and defect detection. In semantic anomaly detection, the inlier and outlier are different in general, (e.g., one is a dog, the other a cat). In defect detection, the semantics for inlier and outlier are the same (e.g., they are both tiles), but the outlier has a local anomaly.

While learning representations with rotation prediction and distribution-augmented contrastive learning have demonstrated state-of-the-art performance on semantic anomaly detection, those algorithms do not perform well on texture anomaly detection. Instead, we explored different representation learning algorithms that better fit the application.

In our second paper, we propose a new self-supervised learning algorithm for texture anomaly detection. The overall anomaly detection follows the 2-stage framework, but the first stage, in which the model learns deep image representations, is specifically trained to predict whether the image is augmented via a simple CutPaste data augmentation. The idea of CutPaste augmentation is simple — a given image is augmented by randomly cutting a local patch and pasting it back to a different location of the same image. Learning to distinguish normal examples from CutPaste-augmented examples encourages representations to be sensitive to local irregularity of an image.

The illustration of learning representations by predicting CutPaste augmentations. Given an example, the CutPaste augmentation crops a local patch, then pasties it to a randomly selected area of the same image. We then train a binary classifier to distinguish the original image and the CutPaste augmented image.

We use MVTec, a real-world defect detection dataset with 15 object categories, to evaluate the approach above.

DOCC (Ruff et al., 2020)	U-Student (Bergmann et al., 2020)	Rotation Prediction	Contrastive (DA)	CutPaste
87.9	92.5	86.3	86.5	95.2

Image-level anomaly detection performance (in AUC) on the MVTec benchmark.

Besides image-level anomaly detection, we use the CutPaste method to locate where the anomaly is, i.e., “patch-level” anomaly detection. We aggregate the patch anomaly scores via upsampling with Gaussian smoothing and visualize them in heatmaps that show where the anomaly is. Interestingly, this provides decently improved localization of anomalies. The below table shows the pixel-level AUC for localization evaluation.

Autoencoder (Bergmann et al., 2019)	FCDD (Ruff et al., 2020)	Rotation Prediction	Contrastive (DA)	CutPaste
86.0	92.0	93.0	90.4	96.0

Pixel-level anomaly localization performance (in AUC) comparison between different algorithms on the MVTec benchmark.

Conclusion
In this work we introduce a novel 2-stage deep one-class classification framework and emphasize the importance of decoupling building classifiers from learning representations so that the classifier can be consistent with the target task, one-class classification. Moreover, this approach permits applications of various self-supervised representation learning methods, attaining state-of-the-art performance on various applications of visual one-class classification from semantic anomaly to texture defect detection. We are extending our efforts to build more realistic anomaly detection methods under the scenario where training data is truly unlabeled.

Acknowledgements
We gratefully acknowledge the contribution from other co-authors, including Jinsung Yoon, Minho Jin and Tomas Pfister. We release the code in our GitHub repository.

Posted by Zaid Nabulsi, Software Engineer and Po-Hsuan Cameron Chen, Software Engineer, Google Health

The adoption of machine learning (ML) for medical imaging applications presents an exciting opportunity to improve the availability, latency, accuracy, and consistency of chest X-ray (CXR) image interpretation. Indeed, a plethora of algorithms have already been developed to detect specific conditions, such as lung cancer, tuberculosis and pneumothorax. By virtue of being trained to detect a specific disease, however, the utility of these algorithms may be limited in a general clinical setting, where a wide variety of abnormalities could surface. For example, a pneumothorax detector is not expected to highlight nodules suggestive of cancer, and a tuberculosis detector may not identify findings specific to pneumonia. Since an initial triaging step is to determine whether a CXR contains any concerning abnormalities, a general-purpose algorithm that identifies X-rays containing any sort of abnormality could significantly facilitate the workflow. However, developing a classifier to detect any abnormality is challenging due to the ​​wide variety of abnormal findings that present on CXRs.

In “Deep Learning for Distinguishing Normal versus Abnormal Chest Radiographs and Generalization to Two Unseen Diseases Tuberculosis and COVID-19”, published in Scientific Reports, we present a model that can distinguish between normal and abnormal CXRs across multiple de-identified datasets and settings. We find that the model performs well on general abnormalities, as well as unseen examples of tuberculosis and COVID-19. We are also releasing our set of radiologists’ labels¹ for the test set used in this study for the publicly available ChestX-ray14 dataset.

A Deep Learning System for Detecting Abnormal Chest X-rays
The deep learning system we used is based on the EfficientNet-B7 architecture, pre-trained on ImageNet. We trained the model using over 200,000 de-identified CXRs from the Apollo Hospitals in India. Each CXR was assigned a label of either “normal” or “abnormal” using a regular expression–based natural language processing approach on the associated radiology reports.

To evaluate how well the system generalizes to new patient populations, we compared its performance on two datasets consisting of a wide spectrum of abnormalities: the test split from the Apollo Hospitals dataset (DS-1), and the publicly available ChestX-ray14 (CXR-14). The labels for these two test sets were annotated for the purposes of this project by a group of US board-certified radiologists. The system achieved areas under the receiver operating characteristic curve (AUROC) of 0.87 on DS-1 and 0.94 on CXR-14 (higher is better).

Though the evaluations on DS-1 and CXR-14 contained a wide range of abnormalities, a possible use-case would be to utilize such an abnormality detector in novel or unforeseen settings with diseases that it had not encountered before. To evaluate the generalizability of the system to new patient populations and in the presence of diseases not seen in the training set, we used four de-identified datasets from three countries, including two publicly available tuberculosis datasets and two COVID-19 datasets from Northwestern Medicine. The system achieved AUCs of 0.95-0.97 in detecting tuberculosis, and 0.65-0.68 in detecting COVID-19. Because CXRs that are negative for these diseases could still contain other concerning abnormalities, we further evaluated the system for its ability to detect abnormalities more broadly (instead of disease positive vs. negative), finding AUCs of 0.91-0.93 for the tuberculosis dataset, and AUCs of 0.86 for the COVID-19 dataset.

The purpose of multiple evaluations (abnormality detection and disease detection) is the distinction between the two: a given disease can present with a certain abnormality or not; and a certain abnormality can arise from multiple diseases. Our study evaluates for both.

<!–

​​AUCs for Three Evaluation Setups
	1. General Abnormalities	2. Unseen disease: Tuberculosis	3. Unseen disease: COVID-19
Detect abnormalities	0.87-0.94	0.91-0.93	0.86
Detect respective disease	–	0.95-0.97	0.65-0.68

–>

The large drop in performance for COVID-19 is because many cases flagged by the system as “positive” for abnormalities were negative for COVID-19, but nevertheless contained abnormal CXR findings that needed attention. This further highlights the usefulness of abnormality detectors even if disease-specific models are available.

In addition, it’s important to note that there is a difference between generalization to unseen diseases (i.e., tuberculosis and COVID-19) versus generalization to unseen CXR findings (e.g., pleural effusion, consolidation/infiltrate). In this study, we demonstrated the generalizability of the system to unseen diseases but not necessarily unseen CXR findings.

Sample chest X-rays of true and false positives, and true and false negatives for (A) general abnormalities, (B) tuberculosis, and (C) COVID-19. On each CXR, we outline in red the areas on which the model focused to identify abnormalities (i.e., the class activation map), and outline the regions of interest indicated by a radiologist in yellow.

Potential Benefits in the Clinic
To understand the potential utility of the deep learning model in improving clinical workflow, we simulated its use for case prioritization, where abnormal cases are “expedited” ahead of normal cases. In these simulations, the system reduced the turnaround time for abnormal cases by up to 28%. This reprioritization setup could be used to divert complex abnormal cases to cardiothoracic specialist radiologists, enable rapid triage of cases that may need urgent decisions, and provide the opportunity to batch negative CXRs for streamlined review.

Impact of a simulated deep learning model–based prioritization in comparison with random review order for (A) general abnormalities, (B) tuberculosis, and (C) COVID-19. The red bars indicate sequences of abnormal CXRs in red and normal CXRs in pink; a greater density of red towards the left indicates abnormal CXRs are reviewed sooner than normal ones. The histograms indicate the average improvement in turnaround time.

Additionally, we found that the system can be used as a pre-trained model to improve other ML algorithms for chest X-rays, especially when data is limited. For example, we used the normal/abnormal classifier in our recent study to detect pulmonary tuberculosis from chest X-rays. Abnormality and tuberculosis detectors can play a critical role in supporting early diagnosis in regions that lack access to resources like trained radiologists or molecular testing.

Sharing Improved Reference Standard Labels
Much work remains to be done to realize the potential of ML to aid chest X-ray interpretation around the world. In particular, obtaining high-quality labels on de-identified data can be a significant barrier to developing and evaluating ML algorithms in healthcare. To accelerate these efforts, we are expanding upon our previous label release by releasing the labels used in this study for the publicly available ChestX-ray14 dataset. We look forward to future machine learning projects by the community in this space.

AcknowledgementsKey contributors to this project at Google include Zaid Nabulsi, Andrew Sellergren‎, Shahar Jamshy, Charles Lau, Eddie Santos, Atilla P. Kiraly, Wenxing Ye, Jie Yang, Rory Pilgrim, Sahar Kazemzadeh, Jin Yu, Greg S. Corrado, Lily Peng, Krish Eswaran, Daniel Tse, Neeral Beladia, Yun Liu, Po-Hsuan Cameron Chen, Shravya Shetty. Significant contributions and input were also made by radiologist collaborators Sreenivasa Raju Kalidindi, Mozziyar Etemadi, Florencia Garcia Vicente, David Melnick. For the CXR-14 dataset, we thank the NIH Clinical Center for making it publicly available. For tuberculosis data collection, thanks go to Sameer Antani, Stefan Jaeger, Sema Candemir, Zhiyun Xue, Alex Karargyris, George R. Thomas, Pu-Xuan Lu, Yi-Xiang Wang, Michael Bonifant, Ellan Kim, Sonia Qasba, and Jonathan Musco. The authors would also like to acknowledge many members of the Google Health Radiology and labeling software teams, in particular Shruthi Prabhakara, Scott McKinney, and Akib Uddin. Sincere appreciation also goes to the radiologists who enabled this work with their image interpretation and annotation efforts throughout the study; Jonny Wong for coordinating the imaging annotation work; Gavin Bee, Mikhail Fomitchev, Shabir Adeel, Jeff Bertram, and Benedict Noero for data releasing; David F. Steiner, Kunal Nagpal, and Michael D. Howell for providing feedback on the manuscript; Craig Mermel, Lauren Winer, Johnny Luu, Adrienne Welch, Annisah Um’rani, and Ashley Zlatinov for feedback on the blogpost.

---

¹Labels include atelectasis, cardiomegaly, effusion, infiltration, mass, nodule, pneumonia, pneumothorax, consolidation, edema, emphysema, fibrosis, pleural thickening, hernia, other abnormality, and normal vs abnormal. ^↩

Posted by Forrester Cole, Software Engineer and Tali Dekel, Research Scientist

Image and video editing operations often rely on accurate mattes — images that define a separation between foreground and background. While recent computer vision techniques can produce high-quality mattes for natural images and videos, allowing real-world applications such as generating synthetic depth-of-field, editing and synthesising images, or removing backgrounds from images, one fundamental piece is missing: the various scene effects that the subject may generate, like shadows, reflections, or smoke, are typically overlooked.

In “Omnimatte: Associating Objects and Their Effects in Video”, presented at CVPR 2021, we describe a new approach to matte generation that leverages layered neural rendering to separate a video into layers called omnimattes that include not only the subjects but also all of the effects related to them in the scene. Whereas a typical state-of-the-art segmentation model extracts masks for the subjects in a scene, for example, a person and a dog, the method proposed here can isolate and extract additional details associated with the subjects, such as shadows cast on the ground.

A state-of-the-art segmentation network (e.g., MaskRCNN) takes an input video (left) and produces plausible masks for people and animals (middle), but misses their associated effects. Our method produces mattes that include not only the subjects, but their shadows as well (right; individual channels for person and dog visualized as blue and green).

Also unlike segmentation masks, omnimattes can capture partially-transparent, soft effects such as reflections, splashes, or tire smoke. Like conventional mattes, omnimattes are RGBA images that can be manipulated using widely-available image or video editing tools, and can be used wherever conventional mattes are used, for example, to insert text into a video underneath a smoke trail.

Layered Decomposition of Video
To generate omnimattes, we split the input video into a set of layers: one for each moving subject, and one additional layer for stationary background objects. In the example below, there is one layer for the person, one for the dog, and one for the background. When merged together using conventional alpha blending, these layers reproduce the input video.

Besides reproducing the video, the decomposition must capture the correct effects in each layer. For example, if the person’s shadow appears in the dog’s layer, the merged layers would still reproduce the input video, but inserting an additional element between the person and dog would produce an obvious error. The challenge is to find a decomposition where each subject’s layer captures only that subject’s effects, producing a true omnimatte.

Our solution is to apply our previously developed layered neural rendering approach to train a convolutional neural network (CNN) to map the subject’s segmentation mask and a background noise image into an omnimatte. Due to their structure, CNNs are naturally inclined to learn correlations between image effects, and the stronger the correlation between the effects, the easier for the CNN to learn. In the above video, for example, the spatial relationships between the person and their shadow, and the dog and its shadow, remain similar as they walk from right to left. The relationships change more (hence, the correlations are weaker) between the person and the dog’s shadow, or the dog and the person’s shadow. The CNN learns the stronger correlations first, leading to the correct decomposition.

The omnimatte system is shown in detail below. In a preprocess, the user chooses the subjects and specifies a layer for each. A segmentation mask for each subject is extracted using an off-the-shelf segmentation network, such as MaskRCNN, and camera transformations relative to the background are found using standard camera stabilization tools. A random noise image is defined in the background reference frame and sampled using the camera transformations to produce per-frame noise images. The noise images provide image features that are random but consistently track the background over time, providing a natural input for the CNN to learn to reconstruct the background colors.

The rendering CNN takes as input the segmentation mask and the per-frame noise images and produces the RGB color images and alpha maps, which capture the transparency of each layer. These outputs are merged using conventional alpha-blending to produce the output frame. The CNN is trained from scratch to reconstruct the input frames by finding and associating the effects not captured in a mask (e.g., shadows, reflections or smoke) with the given foreground layer, and to ensure the subject’s alpha roughly includes the segmentation mask. To make sure the foreground layers only capture the foreground elements and none of the stationary background, a sparsity loss is also applied on the foreground alpha.

A new rendering network is trained for each video. Because the network is only required to reconstruct the single input video, it is able to capture fine structures and fast motion in addition to separating the effects of each subject, as seen below. In the walking example, the omnimatte includes the shadow cast on the slats of the park bench. In the tennis example, the thin shadow and even the tennis ball are captured. In the soccer example, the shadow of the player and the ball are decomposed into their proper layers (with a slight error when the player’s foot is occluded by the ball).

This basic model already works well, but one can improve the results by augmenting the input of the CNN with additional buffers such as optical flow or texture coordinates.

Applications
Once the omnimattes are generated, how can they be used? As shown above, we can remove objects, simply by removing their layer from the composition. We can also duplicate objects, by repeating their layer in the composition. In the example below, the video has been “unwrapped” into a panorama, and the horse duplicated several times to produce a stroboscopic photograph effect. Note that the shadow that the horse casts on the ground and onto the obstacle is correctly captured.

A more subtle, but powerful application is to retime the subjects. Manipulation of time is widely used in film, but usually requires separate shots for each subject and a controlled filming environment. A decomposition into omnimattes makes retiming effects possible for everyday videos using only post-processing, simply by independently changing the playback rate of each layer. Since the omnimattes are standard RGBA images, this retiming edit can be done using conventional video editing software.

The video below is decomposed into three layers, one for each child. The children’s initial, unsynchronized jumps are aligned by simply adjusting the playback rate of their layers, producing realistic retiming for the splashes and reflections in the water.

In the original video (left), each child jumps at a different time. After editing (right), everyone jumps together.

It’s important to consider that any novel technique for manipulating images should be developed and applied responsibly, as it could be misused to produce fake or misleading information. Our technique was developed in accordance with our AI Principles and only allows rearrangement of content already present in the video, but even simple rearrangement can significantly alter the effect of a video, as shown in these examples. Researchers should be aware of these risks.

Future Work
There are a number of exciting directions to improve the quality of the omnimattes. On a practical level, this system currently only supports backgrounds that can be modeled as panoramas, where the position of the camera is fixed. When the camera position moves, the panorama model cannot accurately capture the entire background, and some background elements may clutter the foreground layers (sometimes visible in the above figures). Handling fully general camera motion, such as walking through a room or down a street, would require a 3D background model. Reconstruction of 3D scenes in the presence of moving objects and effects is still a difficult research challenge, but one that has seen promising recent progress.

On a theoretical level, the ability of CNNs to learn correlations is powerful, but still somewhat mysterious, and does not always lead to the expected layer decomposition. While our system allows for manual editing when the automatic result is imperfect, a better solution would be to fully understand the capabilities and limitations of CNNs to learn image correlations. Such an understanding could lead to improved denoising, inpainting, and many other video editing applications besides layer decomposition.

Acknowledgements
Erika Lu, from the University of Oxford, developed the omnimatte system during two internships at Google, in collaboration with Google researchers Forrester Cole, Tali Dekel, Michael Rubinstein, William T. Freeman and David Salesin, and University of Oxford researchers Weidi Xie and Andrew Zisserman.

Thank you to the friends and families of the authors who agreed to appear in the example videos. The “horse jump low”, “lucia”, and “tennis” videos are from the DAVIS 2016 dataset. The soccer video is used by permission from Online Soccer Skills. The car drift video was licensed from Shutterstock.

Posted by Ye Jia, Software Engineer and Julie Cattiau, Product Manager, Google Research

On June 2nd, 2021, Major League Baseball in the United States celebrated Lou Gehrig Day, commemorating both the day in 1925 that Lou Gehrig became the Yankees’ starting first baseman, and the day in 1941 that he passed away from amyotrophic lateral sclerosis (ALS, also known as Lou Gehrig’s disease) at the age of 37. ALS is a progressive neurodegenerative disease that affects motor neurons, which connect the brain with the muscles throughout the body, and govern muscle control and voluntary movements. When voluntary muscle control is affected, people may lose their ability to speak, eat, move and breathe.

In honor of Lou Gehrig, former NFL player and ALS advocate Steve Gleason, who lost his ability to speak due to ALS, recited Gehrig’s famous “Luckiest Man” speech at the June 2nd event using a recreation of his voice generated by a machine learning (ML) model. Gleason’s voice recreation was developed in collaboration with Google’s Project Euphonia, which aims to empower people who have impaired speaking ability due to ALS to better communicate using their own voices.

Steve Gleason, who lost his voice to ALS, worked with Google’s Project Euphonia to generate a speech in his own voice in honor of Lou Gehrig. A portion of Gleason’s speech was broadcast in ballparks across the country during the 4th inning on June 2nd, 2021.

Today we describe PnG NAT, the model adopted by Project Euphonia to recreate Steve Gleason’s voice. PnG NAT is a new text-to-speech synthesis (TTS) model that merges two state-of-the-art technologies, PnG BERT and Non-Attentive Tacotron (NAT), into a single model. It demonstrates significantly better quality and fluency than previous technologies, and represents a promising approach that can be extended to a wider array of users.

Recreating a Voice
Non-Attentive Tacotron (NAT) is the successor to Tacotron 2, a sequence-to-sequence neural TTS model proposed in 2017. Tacotron 2 used an attention module to connect the input text sequence and the output speech spectrogram frame sequence, so that the model knows which part of the text to pay attention to when generating each time step of the synthesized speech spectrogram. Tacotron 2 was the first TTS model that was able to synthesize speech that sounds as natural as a person speaking. However, with extensive experimentation we discovered that there is a small probability that the model can suffer from robustness issues — such as babbling, repeating, or skipping part of the text — due to the inherent flexibility of the attention mechanism.

NAT improves upon Tacotron 2 by replacing the attention module with a duration-based upsampler, which predicts a duration for each input phoneme and upsamples the encoded phoneme representation so that the output length corresponds to the length of the predicted speech spectrogram. Such a change both resolves the robustness issue, and improves the naturalness of the synthesized speech. This approach also enables precise control of the speech duration for each phoneme of the input text while still maintaining highly natural synthesis quality. Because recordings of people with ALS often exhibit disfluent speech, this ability to exert per-phoneme control is key for achieving the fluency of the recreated voice.

Non-Attentive Tacotron (NAT) model.

While NAT addresses the robustness issue and enables precise duration control in neural TTS, we build upon it to further improve the natural language understanding of the TTS input. For this, we apply PnG BERT, which uses an approach similar to BERT, but is specifically designed for TTS. It is pre-trained with self-supervision on both the phoneme representation and the grapheme representation of the same content from a large text corpus, and then is used as the encoder of the TTS model. This results in a significant improvement of the prosody and pronunciation of the synthesized speech, especially in difficult cases.

Take, for example, the following audio, which was synthesized from a regular NAT model that takes only phonemes as input:

In comparison, the audio synthesized from PnG NAT on the same input text includes an additional pause that makes the meaning more clear.

The input text to both models is, “To cancel the payment, press one; or to continue, two.” Notice the different pause lengths before the ending “two” in the two versions. The word “two” in the version output by the regular NAT model could be confused for “too”. Because “too” and “two” have identical pronunciation (and thus the same phoneme representation), the regular NAT model does not understand which of the two is appropriate, and assumes it to be the word that more frequently follows a comma, “too”. In contrast, the PnG NAT model can more easily tell the difference, because it takes graphemes in addition to phonemes as input, and thus makes more appropriate pause.

The PnG NAT model integrates the pre-trained PnG BERT model as the encoder to the NAT model. The hidden representations output from the encoder are used by NAT to predict the duration of each phoneme, and are then upsampled to match the length of the audio spectrogram, as outlined above. In the final step, a non-attentive decoder converts the upsampled hidden representations into audio speech spectrograms, which are finally converted into audio waveforms by a neural vocoder.

PnG BERT and the pre-training objectives. Yellow boxes represent phonemes, and pink boxes represent graphemes.

PnG NAT: PnG BERT replaces the original encoder in the NAT model. The random masking for the Masked Language Model (MLM) pre-training is removed.

To recreate Steve Gleason’s voice, we first trained a PnG NAT model with recordings from 31 professional speakers, and then fine-tuned it with 30 minutes of Gleason’s recordings. Because these latter recordings were made after he was diagnosed with ALS, they exhibit signs of slurring. The fine tuned model was able to synthesize speech that sounds very similar to these recordings. However, because the symptoms of ALS were already present in Gleason’s speech, they exhibited some similar disfluencies.

To mitigate this, we leveraged the phoneme duration control of NAT as well as the model trained with professional speakers. We first predicted the durations of each phoneme for both a professional speaker and for Gleason, and then used the geometric mean of the two durations for each phoneme to guide the NAT output. As a result, the model is able to speak in Gleason’s voice, but more fluently than in the original recordings.

Here is the full version of the synthesized Lou Gehrig speech in Gleason’s voice:

<!– As a comparison, following is one of Gleason’s recordings that was used to train the model:

–>

Besides recreating voices for people with ALS, PnG NAT is also powering voices for a variety of customers through Google Cloud Custom Voice.

Project Euphonia
Of the millions of people around the world who have neurologic conditions that may impact their speech, such as ALS, cerebral palsy or Down syndrome, many may find it difficult to be understood, which can make face-to-face communication challenging. Using voice-activated technologies can be frustrating too, as they don’t always work reliably. Project Euphonia is a Google Research initiative focused on helping people with impaired speech be better understood. The team is researching ways to improve speech recognition for individuals with speech impairments (see recent blog post and segment in TODAY show), as well as customized text-to-speech technology (see Age of AI documentary featuring former NFL player Tim Shaw).

Acknowledgements
Many people across Google Research, Google Cloud and Consumer Apps, and Google Accessibility teams contributed to this project and the event, including Michael Brenner, Bob MacDonald, Heiga Zen, Yu Zhang, Jonathan Shen, Isaac Elias‎, Yonghui Wu, Anne Keck, Danielle Notaro, Kevin Hogan, Zack Kaplan, KR Liu, Kyndra Price, Zoe Ortiz.

Geometric deep learning is a “program” that aspires to situate deep learning architectures and techniques in a framework of mathematical priors. The priors, such as various types of invariance, first arise in some physical domain. A neural network that well matches the domain will preserve as many invariances as possible. In this post, we present a very conceptual, high-level overview, and highlight a few applications.

Posted by Neil Zeghidour, Research Scientist and Marco Tagliasacchi, Staff Research Scientist, Google Research

Audio codecs are used to efficiently compress audio to reduce either storage requirements or network bandwidth. Ideally, audio codecs should be transparent to the end user, so that the decoded audio is perceptually indistinguishable from the original and the encoding/decoding process does not introduce perceivable latency.

Over the past few years, different audio codecs have been successfully developed to meet these requirements, including Opus and Enhanced Voice Services (EVS). Opus is a versatile speech and audio codec, supporting bitrates from 6 kbps (kilobits per second) to 510 kbps, which has been widely deployed across applications ranging from video conferencing platforms, like Google Meet, to streaming services, like YouTube. EVS is the latest codec developed by the 3GPP standardization body targeting mobile telephony. Like Opus, it is a versatile codec operating at multiple bitrates, 5.9 kbps to 128 kbps. The quality of the reconstructed audio using either of these codecs is excellent at medium-to-low bitrates (12–20 kbps), but it degrades sharply when operating at very low bitrates (⪅3 kbps). While these codecs leverage expert knowledge of human perception as well as carefully engineered signal processing pipelines to maximize the efficiency of the compression algorithms, there has been recent interest in replacing these handcrafted pipelines by machine learning approaches that learn to encode audio in a data-driven manner.

Earlier this year, we released Lyra, a neural audio codec for low-bitrate speech. In “SoundStream: an End-to-End Neural Audio Codec”, we introduce a novel neural audio codec that extends those efforts by providing higher-quality audio and expanding to encode different sound types, including clean speech, noisy and reverberant speech, music, and environmental sounds. SoundStream is the first neural network codec to work on speech and music, while being able to run in real-time on a smartphone CPU. It is able to deliver state-of-the-art quality over a broad range of bitrates with a single trained model, which represents a significant advance in learnable codecs.

Learning an Audio Codec from Data
The main technical ingredient of SoundStream is a neural network, consisting of an encoder, decoder and quantizer, all of which are trained end-to-end. The encoder converts the input audio stream into a coded signal, which is compressed using the quantizer and then converted back to audio using the decoder. SoundStream leverages state-of-the-art solutions in the field of neural audio synthesis to deliver audio at high perceptual quality, by training a discriminator that computes a combination of adversarial and reconstruction loss functions that induce the reconstructed audio to sound like the uncompressed original input. Once trained, the encoder and decoder can be run on separate clients to efficiently transmit high-quality audio over a network.

SoundStream training and inference. During training, the encoder, quantizer and decoder parameters are optimized using a combination of reconstruction and adversarial losses, computed by a discriminator, which is trained to distinguish between the original input audio and the reconstructed audio. During inference, the encoder and quantizer on a transmitter client send the compressed bitstream to a receiver client that can then decode the audio signal.

Learning a Scalable Codec with Residual Vector Quantization
The encoder of SoundStream produces vectors that can take an indefinite number of values. In order to transmit them to the receiver using a limited number of bits, it is necessary to replace them by close vectors from a finite set (called a codebook), a process known as vector quantization. This approach works well at bitrates around 1 kbps or lower, but quickly reaches its limits when using higher bitrates. For example, even at a bitrate as low as 3 kbps, and assuming the encoder produces 100 vectors per second, one would need to store a codebook with more than 1 billion vectors, which is infeasible in practice.

In SoundStream, we address this issue by proposing a new residual vector quantizer (RVQ), consisting of several layers (up to 80 in our experiments). The first layer quantizes the code vectors with moderate resolution, and each of the following layers processes the residual error from the previous one. By splitting the quantization process in several layers, the codebook size can be reduced drastically. As an example, with 100 vectors per second at 3 kbps, and using 5 quantizer layers, the codebook size goes from 1 billion to 320. Moreover, we can easily increase or decrease the bitrate by adding or removing quantizer layers, respectively.

Because network conditions can vary while transmitting audio, ideally a codec should be “scalable” so that it can change its bitrate from low to high depending on the state of the network. While most traditional codecs are scalable, previous learnable codecs need to be trained and deployed specifically for each bitrate.

To circumvent this limitation, we leverage the fact that the number of quantization layers in SoundStream controls the bitrate, and propose a new method called “quantizer dropout”. During training, we randomly drop some quantization layers to simulate a varying bitrate. This pushes the decoder to perform well at any bitrate of the incoming audio stream, and thus helps SoundStream to become “scalable” so that a single trained model can operate at any bitrate, performing as well as models trained specifically for these bitrates.

Comparison of SoundStream models (higher is better) that are trained at 18 kbps with quantizer dropout (bitrate scalable), without quantizer dropout (not bitrate scalable) and evaluated with a variable number of quantizers, or trained and evaluated at a fixed bitrate (bitrate specific). The bitrate-scalable model (a single model for all bitrates) does not lose any quality when compared to bitrate-specific models (a different model for each bitrate), thanks to quantizer dropout.

A State-of-the-Art Audio Codec
SoundStream at 3 kbps outperforms Opus at 12 kbps and approaches the quality of EVS at 9.6 kbps, while using 3.2x–4x fewer bits. This means that encoding audio with SoundStream can provide a similar quality while using a significantly lower amount of bandwidth. Moreover, at the same bitrate, SoundStream outperforms the current version of Lyra, which is based on an autoregressive network. Unlike Lyra, which is already deployed and optimized for production usage, SoundStream is still at an experimental stage. In the future, Lyra will incorporate the components of SoundStream to provide both higher audio quality and reduced complexity.

SoundStream at 3kbps vs. state-of-the-art codecs. MUSHRA score is an indication of subjective quality (the higher the better).

The demonstration of SoundStream’s performance compared to Opus, EVS, and the original Lyra codec is presented in these audio examples, a selection of which are provided below.

Speech

Reference
Lyra (3kbps)
Opus (6kbps)
EVS (5.9kbps)
SoundStream (3kbps)

Music

Reference
Lyra (3kbps)
Opus (6kbps)
EVS (5.9kbps)
SoundStream (3kbps)

Joint Audio Compression and Enhancement
In traditional audio processing pipelines, compression and enhancement (the removal of background noise) are typically performed by different modules. For example, it is possible to apply an audio enhancement algorithm at the transmitter side, before audio is compressed, or at the receiver side, after audio is decoded. In such a setup, each processing step contributes to the end-to-end latency. Conversely, we design SoundStream in such a way that compression and enhancement can be carried out jointly by the same model, without increasing the overall latency. In the following examples, we show that it is possible to combine compression with background noise suppression, by activating and deactivating denoising dynamically (no denoising for 5 seconds, denoising for 5 seconds, no denoising for 5 seconds, etc.).

Original noisy audio
Denoised output*

* Demonstrated by turning denoising on and off every 5 seconds.

Conclusion
Efficient compression is necessary whenever one needs to transmit audio, whether when streaming a video, or during a conference call. SoundStream is an important step towards improving machine learning-driven audio codecs. It outperforms state-of-the-art codecs, such as Opus and EVS, can enhance audio on demand, and requires deployment of only a single scalable model, rather than many.

SoundStream will be released as a part of the next, improved version of Lyra. By integrating SoundStream with Lyra, developers can leverage the existing Lyra APIs and tools for their work, providing both flexibility and better sound quality. We will also release it as a separate TensorFlow model for experimentation.

AcknowledgmentsThe work described here was authored by Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund and Marco Tagliasacchi. We are grateful for all discussions and feedback on this work that we received from our colleagues at Google.

Posted by Jimmy Chen, Quantum Research Scientist and Matt McEwen, Student Researcher, Google Quantum AI

The Google Quantum AI team has been building quantum processors made of superconducting quantum bits (qubits) that have achieved the first beyond-classical computation, as well as the largest quantum chemical simulations to date. However, current generation quantum processors still have high operational error rates — in the range of 10^-3 per operation, compared to the 10^-12 believed to be necessary for a variety of useful algorithms. Bridging this tremendous gap in error rates will require more than just making better qubits — quantum computers of the future will have to use quantum error correction (QEC).

The core idea of QEC is to make a logical qubit by distributing its quantum state across many physical data qubits. When a physical error occurs, one can detect it by repeatedly checking certain properties of the qubits, allowing it to be corrected, preventing any error from occurring on the logical qubit state. While logical errors may still occur if a series of physical qubits experience an error together, this error rate should exponentially decrease with the addition of more physical qubits (more physical qubits need to be involved to cause a logical error). This exponential scaling behavior relies on physical qubit errors being sufficiently rare and independent. In particular, it’s important to suppress correlated errors, where one physical error simultaneously affects many qubits at once or persists over many cycles of error correction. Such correlated errors produce more complex patterns of error detections that are more difficult to correct and more easily cause logical errors.

Our team has recently implemented the ideas of QEC in our Sycamore architecture using quantum repetition codes. These codes consist of one-dimensional chains of qubits that alternate between data qubits, which encode the logical qubit, and measure qubits, which we use to detect errors in the logical state. While these repetition codes can only correct for one kind of quantum error at a time¹, they contain all of the same ingredients as more sophisticated error correction codes and require fewer physical qubits per logical qubit, allowing us to better explore how logical errors decrease as logical qubit size grows.

In “Removing leakage-induced correlated errors in superconducting quantum error correction”, published in Nature Communications, we use these repetition codes to demonstrate a new technique for reducing the amount of correlated errors in our physical qubits. Then, in “Exponential suppression of bit or phase flip errors with repetitive error correction”, published in Nature, we show that the logical errors of these repetition codes are exponentially suppressed as we add more and more physical qubits, consistent with expectations from QEC theory.

Layout of the repetition code (21 qubits, 1D chain) and distance-2 surface code (7 qubits) on the Sycamore device.

Leaky Qubits
The goal of the repetition code is to detect errors on the data qubits without measuring their states directly. It does so by entangling each pair of data qubits with their shared measure qubit in a way that tells us whether those data qubit states are the same or different (i.e., their parity) without telling us the states themselves. We repeat this process over and over in rounds that last only one microsecond. When the measured parities change between rounds, we’ve detected an error.

However, one key challenge stems from how we make qubits out of superconducting circuits. While a qubit needs only two energy states, which are usually labeled |0⟩ and |1⟩, our devices feature a ladder of energy states, |0⟩, |1⟩, |2⟩, |3⟩, and so on. We use the two lowest energy states to encode our qubit with information to be used for computation (we call these the computational states). We use the higher energy states (|2⟩, |3⟩ and higher) to help achieve high-fidelity entangling operations, but these entangling operations can sometimes allow the qubit to “leak” into these higher states, earning them the name leakage states.

Population in the leakage states builds up as operations are applied, which increases the error of subsequent operations and even causes other nearby qubits to leak as well — resulting in a particularly challenging source of correlated error. In our early 2015 experiments on error correction, we observed that as more rounds of error correction were applied, performance declined as leakage began to build.

Mitigating the impact of leakage required us to develop a new kind of qubit operation that could “empty out” leakage states, called multi-level reset. We manipulate the qubit to rapidly pump energy out into the structures used for readout, where it will quickly move off the chip, leaving the qubit cooled to the |0⟩ state, even if it started in |2⟩ or |3⟩. Applying this operation to the data qubits would destroy the logical state we’re trying to protect, but we can apply it to the measure qubits without disturbing the data qubits. Resetting the measure qubits at the end of every round dynamically stabilizes the device so leakage doesn’t continue to grow and spread, allowing our devices to behave more like ideal qubits.

Applying the multi-level reset gate to the measure qubits almost totally removes leakage, while also reducing the growth of leakage on the data qubits.

Exponential Suppression
Having mitigated leakage as a significant source of correlated error, we next set out to test whether the repetition codes give us the predicted exponential reduction in error when increasing the number of qubits. Every time we run our repetition code, it produces a collection of error detections. Because the detections are linked to pairs of qubits rather than individual qubits, we have to look at all of the detections to try to piece together where the errors have occurred, a procedure known as decoding. Once we’ve decoded the errors, we then know which corrections we need to apply to the data qubits. However, decoding can fail if there are too many error detections for the number of data qubits used, resulting in a logical error.

To test our repetition codes, we run codes with sizes ranging from 5 to 21 qubits while also varying the number of error correction rounds. We also run two different types of repetition codes — either a phase-flip code or bit-flip code — that are sensitive to different kinds of quantum errors. By finding the logical error probability as a function of the number of rounds, we can fit a logical error rate for each code size and code type. In our data, we see that the logical error rate does in fact get suppressed exponentially as the code size is increased.

Probability of getting a logical error after decoding versus number of rounds run, shown for various sizes of phase-flip repetition code.

We can quantify the error suppression with the error scaling parameter Lambda (Λ), where a Lambda value of 2 means that we halve the logical error rate every time we add four data qubits to the repetition code. In our experiments, we find Lambda values of 3.18 for the phase-flip code and 2.99 for the bit-flip code. We can compare these experimental values to a numerical simulation of the expected Lambda based on a simple error model with no correlated errors, which predicts values of 3.34 and 3.78 for the bit- and phase-flip codes respectively.

Logical error rate per round versus number of qubits for the phase-flip (X) and bit-flip (Z) repetition codes. The line shows an exponential decay fit, and Λ is the scale factor for the exponential decay.

This is the first time Lambda has been measured in any platform while performing multiple rounds of error detection. We’re especially excited about how close the experimental and simulated Lambda values are, because it means that our system can be described with a fairly simple error model without many unexpected errors occurring. Nevertheless, the agreement is not perfect, indicating that there’s more research to be done in understanding the non-idealities of our QEC architecture, including additional sources of correlated errors.

What’s Next
This work demonstrates two important prerequisites for QEC: first, the Sycamore device can run many rounds of error correction without building up errors over time thanks to our new reset protocol, and second, we were able to validate QEC theory and error models by showing exponential suppression of error in a repetition code. These experiments were the largest stress test of a QEC system yet, using 1000 entangling gates and 500 qubit measurements in our largest test. We’re looking forward to taking what we learned from these experiments and applying it to our target QEC architecture, the 2D surface code, which will require even more qubits with even better performance.

---

¹A true quantum error correcting code would require a two dimensional array of qubits in order to correct for all of the errors that could occur. ^↩

Posted by Felix Stahlberg and Shankar Kumar, Research Scientists, Google Research

Grammatical error correction (GEC) attempts to model grammar and other types of writing errors in order to provide grammar and spelling suggestions, improving the quality of written output in documents, emails, blog posts and even informal chats. Over the past 15 years, there has been a substantial improvement in GEC quality, which can in large part be credited to recasting the problem as a “translation” task. When introduced in Google Docs, for example, this approach resulted in a significant increase in the number of accepted grammar correction suggestions.

One of the biggest challenges for GEC models, however, is data sparsity. Unlike other natural language processing (NLP) tasks, such as speech recognition and machine translation, there is very limited training data available for GEC, even for high-resource languages like English. A common remedy for this is to generate synthetic data using a range of techniques, from heuristic-based random word- or character-level corruptions to model-based approaches. However, such methods tend to be simplistic and do not reflect the true distribution of error types from actual users.

In “Synthetic Data Generation for Grammatical Error Correction with Tagged Corruption Models”, presented at the EACL 16th Workshop on Innovative Use of NLP for Building Educational Applications, we introduce tagged corruption models. Inspired by the popular back-translation data synthesis technique for machine translation, this approach enables the precise control of synthetic data generation, ensuring diverse outputs that are more consistent with the distribution of errors seen in practice. We used tagged corruption models to generate a new 200M sentence dataset, which we have released in order to provide researchers with realistic pre-training data for GEC. By integrating this new dataset into our training pipeline, we were able to significantly improve on GEC baselines.

Tagged Corruption Models
The idea behind applying a conventional corruption model to GEC is to begin with a grammatically correct sentence and then to “corrupt” it by adding errors. A corruption model can be easily trained by switching the source and target sentences in existing GEC datasets, a method that previous studies have shown that can be very effective for generating improved GEC datasets.

A conventional corruption model generates an ungrammatical sentence (red) given a clean input sentence (green).

The tagged corruption model that we propose builds on this idea by taking a clean sentence as input along with an error type tag that describes the kind of error one wishes to reproduce. It then generates an ungrammatical version of the input sentence that contains the given error type. Choosing different error types for different sentences increases the diversity of corruptions compared to a conventional corruption model.

Tagged corruption models generate corruptions (red) for the clean input sentence (green) depending on the error type tag. A determiner error may lead to dropping the “a”, whereas a noun-inflection error may produce the incorrect plural “sheeps”.

To use this model for data generation we first randomly selected 200M clean sentences from the C4 corpus, and assigned an error type tag to each sentence such that their relative frequencies matched the error type tag distribution of the small development set BEA-dev. Since BEA-dev is a carefully curated set that covers a wide range of different English proficiency levels, we expect its tag distribution to be representative for writing errors found in the wild. We then used a tagged corruption model to synthesize the source sentence.

Synthetic data generation with tagged corruption models. The clean C4 sentences (green) are paired with the corrupted sentences (red) in the synthetic GEC training corpus. The corrupted sentences are generated using a tagged corruption model by following the error type frequencies in the development set (bar chart).

Results
In our experiments, tagged corruption models outperformed untagged corruption models on two standard development sets (CoNLL-13 and BEA-dev) by more than three F0.5-points (a standard metric in GEC research that combines precision and recall with more weight on precision), advancing the state-of-the-art on the two widely used academic test sets, CoNLL-14 and BEA-test.

In addition, the use of tagged corruption models not only yields gains on standard GEC test sets, it is also able to adapt GEC systems to the proficiency levels of users. This could be useful, for example, because the error tag distribution for native English writers often differs significantly from the distributions for non-native English speakers. For example, native speakers tend to make more punctuation and spelling mistakes, whereas determiner errors (e.g., missing or superfluous articles, like “a”, “an” or “the”) are more common in text from non-native writers.

Conclusion
Neural sequence models are notoriously data-hungry, but the availability of annotated training data for grammatical error correction is rare. Our new C4_200M corpus is a synthetic dataset containing diverse grammatical errors, which yields state-of-the-art performance when used to pre-train GEC systems. By releasing the dataset we hope to provide GEC researchers with a valuable resource to train strong baseline systems.

Using the torch just-in-time (JIT) compiler, it is possible to query a model trained in R from a different language, provided that language can make use of the low-level libtorch library. This post shows how. In addition, we try to untangle a bit of the terminological jumble surrounding the topic.

Posted by Mark Díaz and Emily Denton, Research Scientists, Google Research, Ethical AI Team

Data underlies much of machine learning (ML) research and development, helping to structure what a machine learning algorithm learns and how models are evaluated and benchmarked. However, data collection and labeling can be complicated by unconscious biases, data access limitations and privacy concerns, among other challenges. As a result, machine learning datasets can reflect unfair social biases along dimensions of race, gender, age, and more.

Methods of examining datasets that can surface information about how different social groups are represented within are a key component of ensuring development of ML models and datasets is aligned with our AI Principles. Such methods can inform the responsible use of ML datasets and point toward potential mitigations of unfair outcomes. For example, prior research has demonstrated that some object recognition datasets are biased toward images sourced from North America and Western Europe, prompting Google’s Crowdsource effort to balance out image representations in other parts of the world.

Today, we demonstrate some of the functionality of a dataset exploration tool, Know Your Data (KYD), recently introduced at Google I/O, using the COCO Captions dataset as a case study. Using this tool, we find a range of gender and age biases in COCO Captions — biases that can be traced to both dataset collection and annotation practices. KYD is a dataset analysis tool that complements the growing suite of responsible AI tools being developed across Google and the broader research community. Currently, KYD only supports analysis of a small set of image datasets, but we’re working hard to make the tool accessible beyond this set.

Introducing Know Your Data
Know Your Data helps ML research, product and compliance teams understand datasets, with the goal of improving data quality, and thus helping to mitigate fairness and bias issues. KYD offers a range of features that allow users to explore and examine machine learning datasets — users can filter, group, and study correlations based on annotations already present in a given dataset. KYD also presents automatically computed labels from Google’s Cloud Vision API, providing users with a simple way to explore their data based on signals that weren’t originally present in the dataset.

A KYD Case Study
As a case study, we explore some of these features using the COCO Captions dataset, an image dataset that contains five human-generated captions for each of over 300k images. Given the rich annotations provided by free-form text, we focus our analysis on signals already present within the dataset.

Exploring Gender Bias
Previous research has demonstrated undesirable gender biases within computer vision datasets, including pornographic imagery of women and image label correlations that align with harmful gender stereotypes. We use KYD to explore gender biases within COCO Captions by examining gendered correlations within the image captions. We find a gender bias in the depiction of different activities across the images in the dataset, as well as biases relating to how people of different genders are described by annotators.

The first part of our analysis aimed to surface gender biases with respect to different activities depicted in the dataset. We examined images captioned with words describing different activities and analyzed their relation to gendered caption words, such as “man” or “woman”. The KYD Relations tab makes it easy to examine the relation between two different signals in a dataset by visualizing the extent to which two signals co-occur more (or less) than would be expected by chance. Each cell indicates either a positive (blue color) or negative (orange color) correlation between two specific signal values along with the strength of that correlation.

KYD also allows users to filter rows of a relations table based on substring matching. Using this functionality, we initially probed for caption words containing “-ing”, as a simple way to filter by verbs. We immediately saw strong gendered correlations:

Using KYD to analyze the relationship between any word and gendered words. Each cell shows if the two respective words co-occur in the same caption more (up arrow) or less often (down arrow) than pure chance.

Digging further into these correlations, we found that several activities stereotypically associated with women, such as “shopping” and “cooking”, co-occur with images captioned with “women” or “woman” at a higher rate than with images captioned with “men” or “man”. In contrast captions describing many physically intensive activities, such as “skateboarding”, “surfing”, and “snowboarding”, co-occur with images captioned with “man” or “men” at higher rates.

While individual image captions may not use stereotypical or derogatory language, such as with the example below, if certain gender groups are over (or under) represented within a particular activity across the whole dataset, models developed from the dataset risk learning stereotypical associations. KYD makes it easy to surface, quantify, and make plans to mitigate this risk.

An image with one of the captions: “Two women cooking in a beige and white kitchen.” Image licensed under CC-BY 2.0.

In addition to examining biases with respect to the social groups depicted with different activities, we also explored biases in how annotators described the appearance of people they perceived as male or female. Inspired by media scholars who have examined the “male gaze” embedded in other forms of visual media, we examined the frequency with which individuals perceived as women in COCO are described using adjectives that position them as an object of desire. KYD allowed us to easily examine co-occurrences between words associated with binary gender (e.g. “female/girl/woman” vs. “male/man/boy”) and words associated with evaluating physical attractiveness. Importantly, these are captions written by human annotators, who are making subjective assessments about the gender of people in the image and choosing a descriptor for attractiveness. We see that the words “attractive”, “beautiful”, “pretty”, and “sexy” are overrepresented in describing people perceived as women as compared to those perceived as men, confirming what prior work has said about how gender is viewed in visual media.

A screenshot from KYD showing the relationship between words that describe attractiveness and gendered words. For example, “attractive” and “male/man/boy” co-occur 12 times, but we expect ~60 times by chance (the ratio is 0.2x). On the other hand, “attractive” and “female/woman/girl” co-occur 2.62 times more than chance.

KYD also allows us to manually inspect images for each relation by clicking on the relation in question. For example, we can see images whose captions include female terms (e.g. “woman”) and the word “beautiful”.

Exploring Age Bias
Adults older than 65 have been shown to be underrepresented in datasets relative to their presence in the general population — a first step toward improving age representation is to allow developers to assess it in their datasets. By looking at caption words describing different activities and analyzing their relation to caption words describing age, KYD helped us to assess the range of example captions depicting older adults. Having example captions of adults in a range of environments and activities is important for a variety of tasks, such as image captioning or pedestrian detection.

The first trend that KYD made clear is how rarely annotators described people as older adults in captions detailing different activities. The relations tab also shows a trend wherein “elderly”, “old”, and “older” tend not to occur with verbs that describe a variety of physical activities that might be important for a system to be able to detect. Important to note is that, relative to “young”, “old” is more often used to describe things other than people, such as belongings or clothing, so these relations are also capturing some uses that don’t describe people.

The relationship between words associated with age and movement from a screenshot of KYD.

The underrepresentation of captions containing the references to older adults that we examined here could be rooted in a relative lack of images depicting older adults as well as in a tendency for annotators to omit older age-related terms when describing people in images. While manual inspection of the intersection of “old” and “running” shows a negative relation, we notice that it shows no older people and a number of locomotives. KYD makes it easy to quantitatively and qualitatively inspect relations to identify dataset strengths and areas for improvement.

Conclusion
Understanding the contents of ML datasets is a critical first step to developing suitable strategies to mitigate the downstream impact of unfair dataset bias. The above analysis points towards several potential mitigations. For example, correlations between certain activities and social groups, which can lead trained models to reproduce social stereotypes, can be potentially mitigated by “dataset balancing” — increasing the representation of under-represented group/activity combinations. However, mitigations focused exclusively on dataset balancing are not sufficient, as our analysis of how different genders are described by annotators demonstrated. We found annotators’ subjective judgements of people portrayed in images were reflected within the final dataset, suggesting a deeper look at methods of image annotations are needed. One solution for data practitioners who are developing image captioning datasets is to consider integrating guidelines that have been developed for writing image descriptions that are sensitive to race, gender, and other identity categories.

The above case studies highlight only some of the KYD features. For example, Cloud Vision API signals are also integrated into KYD and can be used to infer signals that annotators haven’t labeled directly. We encourage the broader ML community to perform their own KYD case studies and share their findings.

KYD complements other dataset analysis tools being developed across the ML community, including Google’s growing Responsible AI toolkit. We look forward to ML practitioners using KYD to better understand their datasets and mitigate potential bias and fairness concerns. If you have feedback on KYD, please write to knowyourdata-feedback@google.com.

Acknowledgements
The analysis and write-up in this post were conducted with equal contribution by Emily Denton, Mark Díaz, and Alex Hanna. We thank Marie Pellat, Ludovic Peran, Daniel Smilkov, Nikhil Thorat and Tsung-Yi for their contributions to and reviews of this post.