Quantum processors are made of superconducting quantum bits (qubits) that — being quantum objects — are highly susceptible to even tiny amounts of environmental noise. This noise can cause errors in quantum computation that need to be addressed to continue advancing quantum computers. Our Sycamore processors are installed in specially designed cryostats, where they are sealed away from stray light and electromagnetic fields and are cooled down to very low temperatures to reduce thermal noise.
However, the world is full of high-energy radiation. In fact, there’s a tiny background of high-energy gamma rays and muons that pass through everything around us all the time. While these particles interact so weakly that they don’t cause any harm in our day-to-day lives, qubits are sensitive enough that even weak particle interactions can cause significant interference.
In “Resolving Catastrophic Error Bursts from Cosmic Rays in Large Arrays of Superconducting Qubits”, published in Nature Physics, we identify the effects of these high-energy particles when they impact the quantum processor. To detect and study individual impact events, we use new techniques in rapid, repetitive measurement to operate our processor like a particle detector. This allows us to characterize the resulting burst of errors as they spread through the chip, helping to better understand this important source of correlated errors.
The Dynamics of a High-Energy Impact The Sycamore quantum processor is constructed with a very thin layer of superconducting aluminum on a silicon substrate, onto which a pattern is etched to define the qubits. At the center of each qubit is the Josephson junction, a superconducting component that defines the distinct energy levels of the qubit, which are used for computation. In a superconducting metal, electrons bind together into a macroscopic, quantum state, which allows electrons to flow as a current with zero resistance (a supercurrent). In superconducting qubits, information is encoded in different patterns of oscillating supercurrent going back and forth through the Josephson junction.
If enough energy is added to the system, the superconducting state can be broken up to produce quasiparticles. These quasiparticles are a problem, as they can absorb energy from the oscillating supercurrent and jump across the Josephson junction, which changes the qubit state and produces errors. To prevent any energy from being absorbed by the chip and producing quasiparticles, we use extensive shielding for electric and magnetic fields, and powerful cryogenic refrigerators to keep the chip near absolute zero temperature, thus minimizing the thermal energy.
A source of energy that we can’t effectively shield against is high-energy radiation, which includes charged particles and photons that can pass straight through most materials. One source of these particles are tiny amounts of radioactive elements that can be found everywhere, e.g., in building materials, the metal that makes up our cryostats, and even in the air. Another source is cosmic rays, which are extremely energetic particles produced by supernovae and black holes. When cosmic rays impact the upper atmosphere, they create a shower of high-energy particles that can travel all the way down to the surface and through our chip. Between radioactive impurities and cosmic ray showers, we expect a high energy particle to pass through a quantum chip every few seconds.
When one of these particles impinges on the chip, it passes straight through and deposits a small amount of its energy along its path through the substrate. Even a small amount of energy from these particles is a very large amount of energy for the qubits. Regardless of where the impact occurs, the energy quickly spreads throughout the entire chip through quantum vibrations called phonons. When these phonons hit the aluminum layer that makes up the qubits, they have more than enough energy to break the superconducting state and produce quasiparticles. So many quasiparticles are produced that the probability of the qubits interacting with one becomes very high. We see this as a sudden and significant increase in errors over the whole chip as those quasiparticles absorb energy from the qubits. Eventually, as phonons escape and the chip cools, these quasiparticles recombine back into the superconducting state, and the qubit error rates slowly return to normal.
Detecting Particles with a Computer The Sycamore processor is designed to perform quantum error correction (QEC) to improve the error rates and enable it to execute a variety of quantum algorithms. QEC provides an effective way of identifying and mitigating errors, provided they are sufficiently rare and independent. However, in the case of a high-energy particle going through the chip, all of the qubits will experience high error rates until the event cools off, producing a correlated error burst that QEC won’t be able to correct. In order to successfully perform QEC, we first have to understand what these impact events look like on the processor, which requires operating it like a particle detector.
To do so, we take advantage of recent advances in qubit state preparation and measurement to quickly prepare each qubit in their excited state, similar to flipping a classical bit from 0 to 1. We then wait for a short idle time and measure whether they are still excited. If the qubits are behaving normally, almost all of them will be. Further, the qubits that experience a decay out of their excited state won’t be correlated, meaning the qubits that have errors will be randomly distributed over the chip.
However, during the experiment we occasionally observe large error bursts, where all the qubits on the chip suddenly become more error prone all at once. This correlated error burst is a clear signature of a high-energy impact event. We also see that, while all qubits on the chip are affected by the event, the qubits with the highest error rates are all concentrated in a “hotspot” around the impact site, where slightly more energy is deposited into the qubit layer by the spreading phonons.
Next Steps Because these error bursts are severe and quickly cover the whole chip, they are a type of correlated error that QEC is unable to correct. Therefore, it’s very important to find a solution to mitigate these events in future processors that are expected to rely on QEC.
Shielding against these particles is very difficult and typically requires careful engineering and design of the cryostat and many meters of shielding, which becomes more impractical as processors grow in size. Another approach is to modify the chip, allowing it to tolerate impacts without causing widespread correlated errors. This is an approach taken in other complex superconducting devices like detectors for astronomical telescopes, where it’s not possible to use shielding. Examples of such mitigation strategies include adding additional metal layers to the chip to absorb phonons and prevent them from getting to the qubit, adding barriers in the chip to prevent phonons spreading over long distances, and adding traps for quasiparticles in the qubits themselves. By employing these techniques, future processors will be much more robust to these high-energy impact events.
As the error rates of quantum processors continue to decrease, and as we make progress in building a prototype of an error-corrected logical qubit, we're increasingly pushed to study more exotic sources of error. While QEC is a powerful tool for correcting many kinds of errors, understanding and correcting more difficult sources of correlated errors will become increasingly important. We’re looking forward to future processor designs that can handle high energy impacts and enable the first experimental demonstrations of working quantum error correction.
Acknowledgements This work wouldn’t have been possible without the contributions of the entire Google Quantum AI Team, especially those who worked to design, fabricate, install and calibrate the Sycamore processors used for this experiment. Special thanks to Rami Barends and Lev Ioffe, who led this project.
Image matting is the process of extracting a precise alpha matte that separates foreground and background objects in an image. This technique has been traditionally used in the filmmaking and photography industry for image and video editing purposes, e.g., background replacement, synthetic bokeh and other visual effects. Image matting assumes that an image is a composite of foreground and background images, and hence, the intensity of each pixel is a linear combination of the foreground and the background.
In the case of traditional image segmentation, the image is segmented in a binary manner, in which a pixel either belongs to the foreground or background. This type of segmentation, however, is unable to deal with natural scenes that contain fine details, e.g., hair and fur, which require estimating a transparency value for each pixel of the foreground object.
Alpha mattes, unlike segmentation masks, are usually extremely precise, preserving strand-level hair details and accurate foreground boundaries. While recent deep learning techniques have shown their potential in image matting, many challenges remain, such as generation of accurate ground truth alpha mattes, improving generalization on in-the-wild images and performing inference on mobile devices treating high-resolution images.
With the Pixel 6, we have significantly improved the appearance of selfies taken in Portrait Mode by introducing a new approach to estimate a high-resolution and accurate alpha matte from a selfie image. When synthesizing the depth-of-field effect, the usage of the alpha matte allows us to extract a more accurate silhouette of the photographed subject and have a better foreground-background separation. This allows users with a wide variety of hairstyles to take great-looking Portrait Mode shots using the selfie camera. In this post, we describe the technology we used to achieve this improvement and discuss how we tackled the challenges mentioned above.
Portrait Matting In designing Portrait Matting, we trained a fully convolutional neural network consisting of a sequence of encoder-decoder blocks to progressively estimate a high-quality alpha matte. We concatenate the input RGB image together with a coarse alpha matte (generated using a low-resolution person segmenter) that is passed as an input to the network. The new Portrait Matting model uses a MobileNetV3 backbone and a shallow (i.e., having a low number of layers) decoder to first predict a refined low-resolution alpha matte that operates on a low-resolution image. Then we use a shallow encoder-decoder and a series of residual blocks to process a high-resolution image and the refined alpha matte from the previous step. The shallow encoder-decoder relies more on lower-level features than the previous MobileNetV3 backbone, focusing on high-resolution structural features to predict final transparency values for each pixel. In this way, the model is able to refine an initial foreground alpha matte and accurately extract very fine details like hair strands. The proposed neural network architecture efficiently runs on Pixel 6 using Tensorflow Lite.
Most recent deep learning work for image matting relies on manually annotated per-pixel alpha mattes used to separate the foreground from the background that are generated with image editing tools or green screens. This process is tedious and does not scale for the generation of large datasets. Also, it often produces inaccurate alpha mattes and foreground images that are contaminated (e.g., by reflected light from the background, or “green spill”). Moreover, this does nothing to ensure that the lighting on the subject appears consistent with the lighting in the new background environment.
To address these challenges, Portrait Matting is trained using a high-quality dataset generated using a custom volumetric capture system, Light Stage. Compared with previous datasets, this is more realistic, as relighting allows the illumination of the foreground subject to match the background. Additionally, we supervise the training of the model using pseudo–ground truth alpha mattes from in-the-wild images to improve model generalization, explained below. This ground truth data generation process is one of the key components of this work.
Ground Truth Data Generation To generate accurate ground truth data, Light Stage produces near-photorealistic models of people using a geodesic sphere outfitted with 331 custom color LED lights, an array of high-resolution cameras, and a set of custom high-resolution depth sensors. Together with Light Stage data, we compute accurate alpha mattes using time-multiplexed lights and a previously recorded “clean plate”. This technique is also known as ratio matting.
Then, we extrapolate the recorded alpha mattes to all the camera viewpoints in Light Stage using a deep learning–based matting network that leverages captured clean plates as an input. This approach allows us to extend the alpha mattes computation to unconstrained backgrounds without the need for specialized time-multiplexed lighting or a clean background. This deep learning architecture was solely trained using ground truth mattes generated using the ratio matting approach.
Leveraging the reflectance field for each subject and the alpha matte generated with our ground truth matte generation system, we can relight each portrait using a given HDR lighting environment. We composite these relit subjects into backgrounds corresponding to the target illumination following the alpha blending equation. The background images are then generated from the HDR panoramas by positioning a virtual camera at the center and ray-tracing into the panorama from the camera’s center of projection. We ensure that the projected view into the panorama matches its orientation as used for relighting. We use virtual cameras with different focal lengths to simulate the different fields-of-view of consumer cameras. This pipeline produces realistic composites by handling matting, relighting, and compositing in one system, which we then use to train the Portrait Matting model.
Training Supervision Using In-the-Wild Portraits To bridge the gap between portraits generated using Light Stage and in-the-wild portraits, we created a pipeline to automatically annotate in-the-wild photos generating pseudo–ground truth alpha mattes. For this purpose, we leveraged the Deep Matting model proposed in Total Relighting to create an ensemble of models that computes multiple high-resolution alpha mattes from in-the-wild images. We ran this pipeline on an extensive dataset of portrait photos captured in-house using Pixel phones. Additionally, during this process we performed test-time augmentation by doing inference on input images at different scales and rotations, and finally aggregating per-pixel alpha values across all estimated alpha mattes.
Generated alpha mattes are visually evaluated with respect to the input RGB image. The alpha mattes that are perceptually correct, i.e., following the subject's silhouette and fine details (e.g., hair), are added to the training set. During training, both datasets are sampled using different weights. Using the proposed supervision strategy exposes the model to a larger variety of scenes and human poses, improving its predictions on photos in the wild (model generalization).
Portrait Mode Selfies The Portrait Mode effect is particularly sensitive to errors around the subject boundary (see image below). For example, errors caused by the usage of a coarse alpha matte keep sharp focus on background regions near the subject boundaries or hair area. The usage of a high-quality alpha matte allows us to extract a more accurate silhouette of the photographed subject and improve foreground-background separation.
Try It Out Yourself We have made front-facing camera Portrait Mode on the Pixel 6 better by improving alpha matte quality, resulting in fewer errors in the final rendered image and by improving the look of the blurred background around the hair region and subject boundary. Additionally, our ML model uses diverse training datasets that cover a wide variety of skin tones and hair styles. You can try this improved version of Portrait Mode by taking a selfie shot with the new Pixel 6 phones.
Acknowledgments This work wouldn’t have been possible without Sergio Orts Escolano, Jana Ehmann, Sean Fanello, Christoph Rhemann, Junlan Yang, Andy Hsu, Hossam Isack, Rohit Pandey, David Aguilar, Yi Jinn, Christian Hane, Jay Busch, Cynthia Herrera, Matt Whalen, Philip Davidson, Jonathan Taylor, Peter Lincoln, Geoff Harvey, Nisha Masharani, Alexander Schiffhauer, Chloe LeGendre, Paul Debevec, Sofien Bouaziz, Adarsh Kowdle, Thabo Beeler, Chia-Kai Liang and Shahram Izadi. Special thanks to our photographers James Adamson, Christopher Farro and Cort Muller who took numerous test photographs for us.
Birds are all around us, and just by listening, we can learn many things about our environment. Ecologists use birds to understand food systems and forest health — for example, if there are more woodpeckers in a forest, that means there’s a lot of dead wood. Because birds communicate and mark territory with songs and calls, it’s most efficient to identify them by ear. In fact, experts may identify up to 10x as many birds by ear as by sight.
In recent years, autonomous recording units (ARUs) have made it easy to capture thousands of hours of audio in forests that could be used to better understand ecosystems and identify critical habitat. However, manually reviewing the audio data is very time consuming, and experts in birdsong are rare. But an approach based on machine learning (ML) has the potential to greatly reduce the amount of expert review needed for understanding a habitat.
However, ML-based audio classification of bird species can be challenging for several reasons. For one, birds often sing over one another, especially during the “dawn chorus” when many birds are most active. Also, there aren’t clear recordings of individual birds to learn from — almost all of the available training data is recorded in noisy outdoor conditions, where other sounds from the wind, insects, and other environmental sources are often present. As a result, existing birdsong classification models struggle to identify quiet, distant and overlapping vocalizations. Additionally, some of the most common species often appear unlabeled in the background of training recordings for less common species, leading models to discount the common species. These difficult cases are very important for ecologists who want to identify endangered or invasive species using automated systems.
To address the general challenge of training ML models to automatically separate audio recordings without access to examples of isolated sounds, we recently proposed a new unsupervised method called mixture invariant training (MixIT) in our paper, “Unsupervised Sound Separation Using Mixture Invariant Training”. Moreover, in our new paper, “Improving Bird Classification with Unsupervised Sound Separation,” we use MixIT training to separate birdsong and improve species classification. We found that including the separated audio in the classification improves precision and classification quality on three independent soundscape datasets. We are also happy to announce the open-source release of the birdsong separation models on GitHub.
Birdsong Audio Separation MixIT learns to separate single-channel recordings into multiple individual tracks, and can be trained entirely with noisy, real-world recordings. To train the separation model, we create a “mixture of mixtures” (MoM) by mixing together two real-world recordings. The separation model then learns to take the MoM apart into many channels to minimize a loss function that uses the two original real-world recordings as ground-truth references. The loss function uses these references to group the separated channels such that they can be mixed back together to recreate the two original real-world recordings. Since there’s no way to know how the different sounds in the MoM were grouped together in the original recordings, the separation model has no choice but to separate the individual sounds themselves, and thus learns to place each singing bird in a different output audio channel, also separate from wind and other background noise.
We trained a new MixIT separation model using birdsong recordings from Xeno-Canto and the Macaulay Library. We found that for separating birdsong, this new model outperformed a MixIT separation model trained on a large amount of general audio from the AudioSet dataset. We measure the quality of the separation by mixing two recordings together, applying separation, and then remixing the separated audio channels such that they reconstruct the original two recordings. We measure the signal-to-noise ratio (SNR) of the remixed audio relative to the original recordings. We found that the model trained specifically for birds achieved 6.1 decibels (dB) better SNR than the model trained on AudioSet (10.5 dB vs 4.4 dB). Subjectively, we also found many examples where the system worked incredibly well, separating very difficult to distinguish calls in real-world data.
The following videos demonstrate separation of birdsong from two different regions (Caples and the High Sierras). The videos show the mel-spectrogram of the mixed audio (a 2D image that shows the frequency content of the audio over time) and highlight the audio separated into different tracks.
Classifying Bird Species To classify birds in real-world audio captured with ARUs, we first split the audio into five-second segments and then create a mel-spectrogram of each segment. We then train an EfficientNet classifier to identify bird species from the mel-spectrogram images, training on audio from Xeno-Canto and the Macaulay Library. We trained two separate classifiers, one for species in the Sierra Nevada mountains and one for upstate New York. Note that these classifiers are not trained on separated audio; that’s an area for future improvement.
We also introduced some new techniques to improve classifier training. Taxonomic training asks the classifier to provide labels for each level of the species taxonomy (genus, family, and order), which allows the model to learn groupings of species before learning the sometimes-subtle differences between similar species. Taxonomic training also allows the model to benefit from expert information about the taxonomic relationships between different species. We also found that random low-pass filtering was helpful for simulating distant sounds during training: As an audio source gets further away, the high-frequency parts fade away before the low-frequency parts. This was particularly effective for identifying species from the High Sierras region, where birdsongs cover very long distances, unimpeded by trees.
Classifying Separated Audio We found that separating audio with the new MixIT model before classification improved the classifier performance on three independent real-world datasets. The separation was particularly successful for identification of quiet and background birds, and in many cases helped with overlapping vocalizations as well.
The separation model does have some potential limitations. Occasionally we observe over-separation, where a single song is broken into multiple channels, which can cause misclassifications. We also notice that when multiple birds are vocalizing, the most prominent song often gets a lower score after separation. This may be due to loss of environmental context or other artifacts introduced by separation that do not appear during classifier training. For now, we get the best results by running the classifier on the separated channels and the original audio, and taking the maximum score for each species. We expect that further work will allow us to reduce over-separation and find better ways to combine separation and classification. You can see and hear more examples of the full system at our GitHub repo.
Future Directions We are currently working with partners at the California Academy of Sciences to understand how habitat and species mix changes after prescribed fires and wildfires, applying these models to ARU audio collected over many years.
We also foresee many potential applications for the unsupervised separation models in ecology, beyond just birds. For example, the separated audio can be used to create better acoustic indices, which could measure ecosystem health by tracking the total activity of birds, insects, and amphibians without identifying particular species. Similar methods could also be adapted for use underwater to track coral reef health.
Acknowledgements We would like to thank Mary Clapp, Jack Dumbacher, and Durrell Kapan from the California Academy of Sciences for providing extensive annotated soundscapes from the Sierra Nevadas. Stefan Kahl and Holger Klinck from the Cornell Lab of Ornithology provided soundscapes from Sapsucker Woods. Training data for both the separation and classification models came from Xeno-Canto and the Macaulay Library. Finally, we would like to thank Julie Cattiau, Lauren Harrell, Matt Harvey, and our co-author, John Hershey, from the Google Bioacoustics and Sound Separation teams.
Language models are becoming more capable than ever before and are helpful in a variety of tasks — translating one language into another, summarizing a long document into a brief highlight, or answering information-seeking questions. Among these, open-domain dialog, where a model needs to be able to converse about any topic, is probably one of the most difficult, with a wide range of potential applications and open challenges. In addition to producing responses that humans judge as sensible, interesting, and specific to the context, dialog models should adhere to Responsible AI practices, and avoid making factual statements that are not supported by external information sources.
Today we’re excited to share recent advances in our “LaMDA: Language Models for Dialog Applications” project. In this post, we’ll give an overview on how we’re making progress towards safe, grounded, and high-quality dialog applications. LaMDA is built by fine-tuning a family of Transformer-based neural language models specialized for dialog, with up to 137B model parameters, and teaching the models to leverage external knowledge sources.
Objectives & Metrics Defining objectives and metrics is critical to guide training dialog models. LaMDA has three key objectives — Quality, Safety, and Groundedness — each of which we measure using carefully designed metrics:
Quality: We decompose Quality into three dimensions, Sensibleness, Specificity, and Interestingness (SSI), which are evaluated by human raters. Sensibleness refers to whether the model produces responses that make sense in the dialog context (e.g., no common sense mistakes, no absurd responses, and no contradictions with earlier responses). Specificity is measured by judging whether the system's response is specific to the preceding dialog context, and not a generic response that could apply to most contexts (e.g., “ok” or “I don’t know”). Finally, Interestingness measures whether the model produces responses that are also insightful, unexpected or witty, and are therefore more likely to create better dialog.
Safety: We’re also making progress towards addressing important questions related to the development and deployment of Responsible AI. Our Safety metric is composed of an illustrative set of safety objectives that captures the behavior that the model should exhibit in a dialog. These objectives attempt to constrain the model’s output to avoid any unintended results that create risks of harm for the user, and to avoid reinforcing unfair bias. For example, these objectives train the model to avoid producing outputs that contain violent or gory content, promote slurs or hateful stereotypes towards groups of people, or contain profanity. Our research towards developing a practical Safety metric represents very early work, and there is still a great deal of progress for us to make in this area.
Groundedness: The current generation of language models often generate statements that seem plausible, but actually contradict facts established in known external sources. This motivates our study of groundedness in LaMDA. Groundedness is defined as the percentage of responses with claims about the external world that can be supported by authoritative external sources, as a share of all responses containing claims about the external world. A related metric, Informativeness, is defined as the percentage of responses with information about the external world that can be supported by known sources, as a share of all responses. Therefore, casual responses that do not carry any real world information (e.g., “That’s a great idea”), affect Informativeness but not Groundedness. While grounding LaMDA generated responses in known sources does not in itself guarantee factual accuracy, it allows users or external systems to judge the validity of a response based on the reliability of its source.
LaMDA Pre-Training With the objectives and metrics defined, we describe LaMDA’s two-stage training: pre-training and fine-tuning. In the pre-training stage, we first created a dataset of 1.56T words — nearly 40 times more words than what were used to train previous dialog models — from public dialog data and other public web documents. After tokenizing the dataset into 2.81T SentencePiece tokens, we pre-train the model using GSPMD to predict every next token in a sentence, given the previous tokens. The pre-trained LaMDA model has also been widely used for natural language processing research across Google, including program synthesis, zero-shot learning, style transfer, as well as in the BIG-bench workshop.
LaMDA Fine-Tuning In the fine-tuning stage, we train LaMDA to perform a mix of generative tasks to generate natural-language responses to given contexts, and classification tasks on whether a response is safe and high-quality, resulting in a single multi-task model that can do both. The LaMDA generator is trained to predict the next token on a dialog dataset restricted to back-and-forth dialog between two authors, while the LaMDA classifiers are trained to predict the Safety and Quality (SSI) ratings for the response in context using annotated data. During a dialog, the LaMDA generator first generates several candidate responses given the current multi-turn dialog context, and the LaMDA classifiers predict the SSI and Safety scores for every response candidate. Candidate responses with low Safety scores are first filtered out. Remaining candidates are re-ranked by their SSI scores, and the top result is selected as the response. We further filter the training data used for the generation task with LaMDA classifiers to increase the density of high-quality response candidates.
Factual Grounding While people are capable of checking their facts by using tools and referencing established knowledge bases, many language models draw their knowledge on their internal model parameters only. To improve the groundedness of LaMDA’s original response, we collect a dataset of dialogs between people and LaMDA, which are annotated with information retrieval queries and the retrieved results where applicable. We then fine-tune LaMDA’s generator and classifier on this dataset to learn to call an external information retrieval system during its interaction with the user to improve the groundedness of its responses. While this is very early work, we’re seeing promising results.
Evaluation In order to quantify progress against our key metrics, we collect responses from the pre-trained model, fine-tuned model, and human raters (i.e., human-generated responses) to multi-turn two-author dialogs, and then ask a different set of human raters a series of questions to evaluate these responses against the Quality, Safety, and Groundedness metrics.
We observe that LaMDA significantly outperforms the pre-trained model in every dimension and across all model sizes. Quality metrics (Sensibleness, Specificity, and Interestingness, in the first column below) generally improve with the number of model parameters, with or without fine-tuning. Safety does not seem to benefit from model scaling alone, but it does improve with fine-tuning. Groundedness improves as model size increases, perhaps because larger models have a greater capacity to memorize uncommon knowledge, but fine-tuning allows the model to access external knowledge sources and effectively shift some of the load of remembering knowledge to an external knowledge source. With fine-tuning, the quality gap to human levels can be narrowed, though the model’s performance remains below human levels in safety and groundedness.
Future Research & Challenges LaMDA’s level of Sensibleness, Specificity and Interestingness unlocks new avenues for understanding the benefits and risks of open-ended dialog agents. It also presents encouraging evidence that key challenges with neural language models, such as using a safety metric and improving groundedness, can improve with larger models and fine-tuning with more well-labeled data. However, this is very early work, and there are significant limitations. Exploring new ways to improve our Safety metric and LaMDA's groundedness, aligned with our AI Principles, will continue to be our main areas of focus going forward.
Acknowledgements We'd to like to thank everyone for contributing to the project and paper, including: Blaise Aguera-Arcas, Javier Alberca, Thushan Amarasiriwardena, Lora Aroyo, Martin Baeuml, Leslie Baker, Rachel Bernstein, Taylor Bos, Maarten Bosma, Jonas Bragagnolo, Alena Butryna, Bill Byrne, Chung-Ching Chang, Zhifeng Chen, Dehao Chen, Heng-Tze Cheng, Ed Chi, Aaron Cohen, Eli Collins, Marian Croak, Claire Cui, Andrew Dai, Dipanjan Das, Daniel De Freitas, Jeff Dean, Rajat Dewan, Mark Diaz, Tulsee Doshi, Yu Du, Toju Duke, Doug Eck, Joe Fenton, Noah Fiedel, Christian Frueh, Harish Ganapathy, Saravanan Ganesh, Amin Ghafouri, Zoubin Ghahramani, Kourosh Gharachorloo, Jamie Hall, Erin Hoffman-John, Sissie Hsiao, Yanping Huang, Ben Hutchinson, Daphne Ippolito, Alicia Jin, Thomas Jurdi, Ashwin Kakarla, Nand Kishore, Maxim Krikun, Karthik Krishnamoorthi, Igor Krivokon, Apoorv Kulshreshtha, Ray Kurzweil, Viktoriya Kuzmina, Vivek Kwatra, Matthew Lamm, Quoc Le, Max Lee, Katherine Lee, Hongrae Lee, Josh Lee, Dmitry Lepikhin, YaGuang Li, Yifeng Lu, David Luan, Daphne Luong, Laichee Man, Jianchang (JC) Mao, Yossi Matias, Kathleen Meier-Hellstern, Marcelo Menegali, Muqthar Mohammad,, Muqthar Mohammad, Alejandra Molina, Erica Moreira, Meredith Ringel Morris, Maysam Moussalem, Jiaqi Mu, Tyler Mullen, Tyler Mullen, Eric Ni, Kristen Olson, Alexander Passos, Fernando Pereira, Slav Petrov, Marc Pickett, Roberto Pieraccini, Christian Plagemann, Sahitya Potluri, Vinodkumar Prabhakaran, Andy Pratt, James Qin, Ravi Rajakumar, Adam Roberts, Will Rusch, Renelito Delos Santos, Noam Shazeer, RJ Skerry-Ryan, Grigori Somin, Johnny Soraker, Pranesh Srinivasan, Amarnag Subramanya, Mustafa Suleyman, Romal Thoppilan, Song Wang, Sheng Wang, Chris Wassman, Yuanzhong Xu, Yuanzhong Xu, Ni Yan, Ben Zevenbergen, Vincent Zhao, Huaixiu Steven Zheng, Denny Zhou, Hao Zhou, Yanqi Zhou, and more.
Neural networks can perform certain tasks remarkably well, but understanding how they reach their decisions — e.g., identifying which signals in an image cause a model to determine it to be of one class and not another — is often a mystery. Explaining a neural model’s decision process may have high social impact in certain areas, such as analysis of medical images and autonomous driving, where human oversight is critical. These insights can also be helpful in guiding health care providers, revealing model biases, providing support for downstream decision makers, and even aiding scientific discovery.
Previous approaches for visual explanations of classifiers, such as attention maps (e.g., Grad-CAM), highlight which regions in an image affect the classification, but they do not explain what attributes within those regions determine the classification outcome: For example, is it their color? Their shape? Another family of methods provides an explanation by smoothly transforming the image between one class and another (e.g., GANalyze). However, these methods tend to change all attributes at once, thus making it difficult to isolate the individual affecting attributes.
In “Explaining in Style: Training a GAN to explain a classifier in StyleSpace”, presented at ICCV 2021, we propose a new approach for a visual explanation of classifiers. Our approach, StylEx, automatically discovers and visualizes disentangled attributes that affect a classifier. It allows exploring the effect of individual attributes by manipulating those attributes separately (changing one attribute does not affect others). StylEx is applicable to a wide range of domains, including animals, leaves, faces, and retinal images. Our results show that StylEx finds attributes that align well with semantic ones, generate meaningful image-specific explanations, and are interpretable by people as measured in user studies.
For instance, to understand a cat vs. dog classifier on a given image, StylEx can automatically detect disentangled attributes and visualize how manipulating each attribute can affect the classifier probability. The user can then view these attributes and make semantic interpretations for what they represent. For example, in the figure above, one can draw conclusions such as “dogs are more likely to have their mouth open than cats” (attribute #4 in the GIF above), “cats’ pupils are more slit-like” (attribute #5), “cats’ ears do not tend to be folded” (attribute #1), and so on.
The video below provides a short explanation of the method:
How StylEx Works: Training StyleGAN to Explain a Classifier Given a classifier and an input image, we want to find and visualize the individual attributes that affect its classification. For that, we utilize the StyleGAN2 architecture, which is known to generate high quality images. Our method consists of two phases:
Phase 1: Training StylEx
A recent work showed that StyleGAN2 contains a disentangled latent space called “StyleSpace”, which contains individual semantically meaningful attributes of the images in the training dataset. However, because StyleGAN training is not dependent on the classifier, it may not represent those attributes that are important for the decision of the specific classifier we want to explain. Therefore, we train a StyleGAN-like generator to satisfy the classifier, thus encouraging its StyleSpace to accommodate classifier-specific attributes. This is achieved by training the StyleGAN generator with two additional components. The first is an encoder, trained together with the GAN with a reconstruction-loss, which forces the generated output image to be visually similar to the input. This allows us to apply the generator on any given input image. However, visual similarity of the image is not enough, as it may not necessarily capture subtle visual details important for a particular classifier (such as medical pathologies). To ensure this, we add a classification-loss to the StyleGAN training, which forces the classifier probability of the generated image to be the same as the classifier probability of the input image. This guarantees that subtle visual details important for the classifier (such as medical pathologies) will be included in the generated image.
A recent work showed that StyleGAN2 contains a disentangled latent space called “StyleSpace”, which contains individual semantically meaningful attributes of the images in the training dataset. However, because StyleGAN training is not dependent on the classifier, it may not represent those attributes that are important for the decision of the specific classifier we want to explain. Therefore, we train a StyleGAN-like generator to satisfy the classifier, thus encouraging its StyleSpace to accommodate classifier-specific attributes.
This is achieved by training the StyleGAN generator with two additional components. The first is an encoder, trained together with the GAN with a reconstruction-loss, which forces the generated output image to be visually similar to the input. This allows us to apply the generator on any given input image. However, visual similarity of the image is not enough, as it may not necessarily capture subtle visual details important for a particular classifier (such as medical pathologies). To ensure this, we add a classification-loss to the StyleGAN training, which forces the classifier probability of the generated image to be the same as the classifier probability of the input image. This guarantees that subtle visual details important for the classifier (such as medical pathologies) will be included in the generated image.
Phase 2: Extracting Disentangled Attributes
Once trained, we search the StyleSpace of the trained Generator for attributes that significantly affect the classifier. To do so, we manipulate each StyleSpace coordinate and measure its effect on the classification probability. We seek the top attributes that maximize the change in classification probability for the given image. This provides the top-K image-specific attributes. By repeating this process for a large number of images per class, we can further discover the top-K class-specific attributes, which teaches us what the classifier has learned about the specific class. We call our end-to-end system “StylEx”. A visual illustration of image-specific attribute extraction: once trained, we search for the StyleSpace coordinates that have the highest effect on the classification probability of a given image.
Once trained, we search the StyleSpace of the trained Generator for attributes that significantly affect the classifier. To do so, we manipulate each StyleSpace coordinate and measure its effect on the classification probability. We seek the top attributes that maximize the change in classification probability for the given image. This provides the top-K image-specific attributes. By repeating this process for a large number of images per class, we can further discover the top-K class-specific attributes, which teaches us what the classifier has learned about the specific class. We call our end-to-end system “StylEx”.
StylEx is Applicable to a Wide Range of Domains and Classifiers Our method works on a wide variety of domains and classifiers (binary and multi-class). Below are some examples of class-specific explanations. In all the domains tested, the top attributes detected by our method correspond to coherent semantic notions when interpreted by humans, as verified by human evaluation.
For perceived gender and age classifiers, below are the top four detected attributes per classifier. Our method exemplifies each attribute on multiple images that are automatically selected to best demonstrate that attribute. For each attribute we flicker between the source and attribute-manipulated image. The degree to which manipulating the attribute affects the classifier probability is shown at the top-left corner of each image.
Note that our method explains a classifier, not reality. That is, the method is designed to reveal image attributes that a given classifier has learned to utilize from data; those attributes may not necessarily characterize actual physical differences between class labels (e.g., a younger or older age) in reality. In particular, these detected attributes may reveal biases in the classifier training or dataset, which is another key benefit of our method. It can further be used to improve fairness of neural networks, for example, by augmenting the training dataset with examples that compensate for the biases our method reveals.
Adding the classifier loss into StyleGAN training turns out to be crucial in domains where the classification depends on fine details. For example, a GAN trained on retinal images without a classifier loss will not necessarily generate fine pathological details corresponding to a particular disease. Adding the classification loss causes the GAN to generate these subtle pathologies as an explanation of the classifier. This is exemplified below for a retinal image classifier (DME disease) and a sick/healthy leaf classifier. StylEx is able to discover attributes that are aligned with disease indicators, for instance “hard exudates”, which is a well known marker for retinal DME, and rot for leaf diseases.
Finally, this method is also applicable to multi-class problems, as demonstrated on a 200-way bird species classifier.
Broader Impact and Next Steps Overall, we have introduced a new technique that enables the generation of meaningful explanations for a given classifier on a given image or class. We believe that our technique is a promising step towards detection and mitigation of previously unknown biases in classifiers and/or datasets, in line with Google’s AI Principles. Additionally, our focus on multiple-attribute based explanation is key to providing new insights about previously opaque classification processes and aiding in the process of scientific discovery. Finally, our GitHub repository includes a Colab and model weights for the GANs used in our paper.
Acknowledgements The research described in this post was done by Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald (as an intern), Gal Elidan, Avinatan Hassidim, William T. Freeman, Phillip Isola, Amir Globerson, Michal Irani and Inbar Mosseri. We would like to thank Jenny Huang and Marilyn Zhang for leading the writing process for this blogpost, and Reena Jana, Paul Nicholas, and Johnny Soraker for ethics reviews of our research paper and this post.
Scaling large language models has resulted in significant quality improvements in natural language understanding (T5), generation (GPT-3) and multilingual neural machine translation (M4). One common approach to building a larger model is to increase the depth (number of layers) and width (layer dimensionality), simply enlarging existing dimensions of the network. Such dense models take an input sequence (divided into smaller components, called tokens) and pass every token through the full network, activating every layer and parameter. While these large, dense models have achieved state-of-the-art results on multiple natural language processing (NLP) tasks, their training cost increases linearly with model size.
An alternative, and increasingly popular, approach is to build sparsely activated models based on a mixture of experts (MoE) (e.g., GShard-M4 or GLaM), where each token passed to the network follows a separate subnetwork by skipping some of the model parameters. The choice of how to distribute the input tokens to each subnetwork (the “experts”) is determined by small router networks that are trained together with the rest of the network. This allows researchers to increase model size (and hence, performance) without a proportional increase in training cost.
While this is an effective strategy at training time, sending tokens of a long sequence to multiple experts, again makes inference computationally expensive because the experts have to be distributed among a large number of accelerators. For example, serving the 1.2T parameter GLaM model requires 256 TPU-v3 chips. Much like dense models, the number of processors needed to serve an MoE model still scales linearly with respect to the model size, increasing compute requirements while also resulting in significant communication overhead and added engineering complexity.
In “Beyond Distillation: Task-level Mixture-of-Experts for Efficient Inference”, we introduce a method called Task-level Mixture-of-Experts (TaskMoE), that takes advantage of the quality gains of model scaling while still being efficient to serve. Our solution is to train a large multi-task model from which we then extract smaller, stand-alone per-task subnetworks suitable for inference with no loss in model quality and with significantly reduced inference latency. We demonstrate the effectiveness of this method for multilingual neural machine translation (NMT) compared to other mixture of experts models and to models compressed using knowledge distillation.
Training Large Sparsely Activated Models with Task Information We train a sparsely activated model, where router networks learn to send tokens of each task-specific input to different subnetworks of the model associated with the task of interest. For example, in the case of multilingual NMT, every token of a given language is routed to the same subnetwork. This differs from other recent approaches, such as the sparsely gated mixture of expert models (e.g., TokenMoE), where router networks learn to send different tokens in an input to different subnetworks independent of task.
Inference: Bypassing Distillation by Extracting Subnetworks A consequence of this difference in training between TaskMoE and models like TokenMoE is in how we approach inference. Because TokenMoE follows the practice of distributing tokens of the same task to many experts at both training and inference time, it is still computationally expensive at inference.
For TaskMoE, we dedicate a smaller subnetwork to a single task identity during training and inference. At inference time, we extract subnetworks by discarding unused experts for each task. TaskMoE and its variants enable us to train a single large multi-task network and then use a separate subnetwork at inference time for each task without using any additional compression methods post-training. We illustrate the process of training a TaskMoE network and then extracting per-task subnetworks for inference below.
To demonstrate this approach, we train models based on the Transformer architecture. Similar to GShard-M4 and GLaM, we replace the feedforward network of every other transformer layer with a Mixture-of-Experts (MoE) layer that consists of multiple identical feedforward networks, the “experts”. For each task, the routing network, trained along with the rest of the model, keeps track of the task identity for all input tokens and chooses a certain number of experts per layer (two in this case) to form the task-specific subnetwork. The baseline dense Transformer model has 143M parameters and 6 layers on both the encoder and decoder. The TaskMoE and TokenMoE that we train are also both 6 layers deep but with 32 experts for every MoE layer and have a total of 533M parameters. We train our models using publicly available WMT datasets, with over 431M sentences across 30 language pairs from different language families and scripts. We point the reader to the full paper for further details.
A popular approach to building a smaller network that still performs well is through knowledge distillation, in which a large teacher model trains a smaller student model with the goal of matching the teacher’s performance. However, this method comes at the cost of additional computation needed to train the student from the teacher. So, we also compare TaskMoE to a baseline TokenMoE model that we compress using knowledge distillation. The compressed TokenMoE model has a size comparable to the per-task subnetwork extracted from TaskMoE.
We find that in addition to being a simpler method that does not need any additional training, TaskMoE improves upon a distilled TokenMoE model by 2.1 BLEU on average across all languages in our multilingual translation model. We note that distillation retains 43% of the performance gains achieved from scaling a dense multilingual model to a TokenMoE, whereas extracting the smaller subnetwork from the TaskMoE model results in no loss of quality.
Next Steps The quality improvements often seen with scaling machine learning models has incentivized the research community to work toward advancing scaling technology to enable efficient training of large models. The emerging need to train models capable of generalizing to multiple tasks and modalities only increases the need for scaling models even further. However, the practicality of serving these large models remains a major challenge. Efficiently deploying large models is an important direction of research, and we believe TaskMoE is a promising step towards more inference friendly algorithms that retain the quality gains of scaling.
Acknowledgements We would like to first thank our coauthors - Yanping Huang, Ankur Bapna, Maxim Krikun, Dmitry Lepikhin and Minh-Thang Luong. We would also like to thank Wolfgang Macherey, Yuanzhong Xu, Zhifeng Chen and Macduff Richard Hughes for their helpful feedback. Special thanks to the Translate and Brain teams for their useful input and discussions, and the entire GShard development team for their foundational contributions to this project. We would also like to thank Tom Small for creating the animations for the blog post.
Advances in deep learning over the last few decades have been driven by a few key elements. With a small number of simple but flexible mechanisms (i.e., inductive biases such as convolutions or sequence attention), increasingly large datasets, and more specialized hardware, neural networks can now achieve impressive results on a wide range of tasks, such as image classification, machine translation, and protein folding prediction.
However, the use of large models and datasets comes at the expense of significant computational requirements. Yet, recent works suggest that large model sizes might be necessary for strong generalization and robustness, so training large models while limiting resource requirements is becoming increasingly important. One promising approach involves the use of conditional computation: rather than activating the whole network for every single input, different parts of the model are activated for different inputs. This paradigm has been featured in the Pathways vision and recent works on large language models, while it has not been well explored in the context of computer vision.
In “Scaling Vision with Sparse Mixture of Experts”, we present V-MoE, a new vision architecture based on a sparse mixture of experts, which we then use to train the largest vision model to date. We transfer V-MoE to ImageNet and demonstrate matching state-of-the-art accuracy while using about 50% fewer resources than models of comparable performance. We have also open-sourced the code to train sparse models and provided several pre-trained models.
Vision Mixture of Experts (V-MoEs) Vision Transformers (ViT) have emerged as one of the best architectures for vision tasks. ViT first partitions an image into equally-sized square patches. These are called tokens, a term inherited from language models. Still, compared to the largest language models, ViT models are several orders of magnitude smaller in terms of number of parameters and compute.
To massively scale vision models, we replace some dense feedforward layers (FFN) in the ViT architecture with a sparse mixture of independent FFNs (which we call experts). A learnable router layer selects which experts are chosen (and how they are weighted) for every individual token. That is, different tokens from the same image may be routed to different experts. Each token is only routed to at most K (typically 1 or 2) experts, among a total of E experts (in our experiments, E is typically 32). This allows scaling the model’s size while keeping its computation per token roughly constant. The figure below shows the structure of the encoder blocks in more detail.
Experimental Results We first pre-train the model once on JFT-300M, a large dataset of images. The left plot below shows our pre-training results for models of all sizes: from the small S/32 to the huge H/14.
We then transfer the model to new downstream tasks (such as ImageNet), by using a new head (the last layer in a model). We explore two transfer setups: either fine-tuning the entire model on all available examples of the new task, or freezing the pre-trained network and tuning only the new head using a few examples (known as few-shot transfer). The right plot in the figure below summarizes our transfer results to ImageNet, training on only 5 images per class (called 5-shot transfer).
In both cases, the sparse model strongly outperforms its dense counterpart at a given amount of training compute (shown by the V-MoE line being above the ViT line), or achieves similar performance much faster (shown by the V-MoE line being to the left of the ViT line).
To explore the limits of vision models, we trained a 15-billion parameter model with 24 MoE layers (out of 48 blocks) on an extended version of JFT-300M. This massive model — the largest to date in vision as far as we know — achieved 90.35% test accuracy on ImageNet after fine-tuning, near the current state-of-the-art.
Priority Routing In practice, due to hardware constraints, it is not efficient to use buffers with a dynamic size, so models typically use a pre-defined buffer capacity for each expert. Assigned tokens beyond this capacity are dropped and not processed once the expert becomes "full". As a consequence, higher capacities yield higher accuracy, but they are also more computationally expensive.
We leverage this implementation constraint to make V-MoEs faster at inference time. By decreasing the total combined buffer capacity below the number of tokens to be processed, the network is forced to skip processing some tokens in the expert layers. Instead of choosing the tokens to skip in some arbitrary fashion (as previous works did), the model learns to sort tokens according to an importance score. This maintains high quality predictions while saving a lot of compute. We refer to this approach as Batch Priority Routing (BPR), illustrated below.
Dropping the right tokens turns out to be essential to deliver high-quality and more efficient inference predictions. When the expert capacity decreases, performance quickly decreases with the vanilla routing mechanism. Conversely, BPR is much more robust to low capacities.
Overall, we observed that V-MoEs are highly flexible at inference time: for instance, one can decrease the number of selected experts per token to save time and compute, without any further training on the model weights.
Exploring V-MoEs Because much is yet to be discovered about the internal workings of sparse networks, we also explored the routing patterns of the V-MoE.
One hypothesis is that routers would learn to discriminate and assign tokens to experts based on some semantic grounds (the “car” expert, the “animal” experts, and so on). To test this, below we show plots for two different MoE layers (a very early-on one, and another closer to the head). The x-axis corresponds to each of the 32 experts, and the y-axis shows the ID of the image classes (from 1 to 1000). Each entry in the plot shows how often an expert was selected for tokens corresponding to the specific image class, with darker colors indicating higher frequency. While in the early layers there is little correlation, later in the network, each expert receives and processes tokens from only a handful of classes. Therefore, we can conclude that some semantic clustering of the patches emerges in the deeper layers of the network.
Final Thoughts We train very large vision models using conditional computation, delivering significant improvements in representation and transfer learning for relatively little training cost. Alongside V-MoE, we introduced BPR, which requires the model to process only the most useful tokens in the expert layers.
We believe this is just the beginning of conditional computation at scale for computer vision; extensions include multi-modal and multi-task models, scaling up the expert count, and improving transfer of the representations produced by sparse models. Heterogeneous expert architectures and conditional variable-length routes are also promising directions. Sparse models can especially help in data rich domains such as large-scale video modeling. We hope our open-source code and models help attract and engage researchers new to this field.
Acknowledgments We thank our co-authors: Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. We thank Alex Kolesnikov, Lucas Beyer, and Xiaohua Zhai for providing continuous help and details about scaling ViT models. We are also grateful to Josip Djolonga, Ilya Tolstikhin, Liam Fedus, and Barret Zoph for feedback on the paper; James Bradbury, Roy Frostig, Blake Hechtman, Dmitry Lepikhin, Anselm Levskaya, and Parker Schuh for invaluable support helping us run our JAX models efficiently on TPUs; and many others from the Brain team for their support. Finally, we would also like to thank and acknowledge Tom Small for the awesome animated figure used in this post.
