Posts in category: Google AI

Posted by Junjie Ke, Senior Software Engineer, and Feng Yang, Senior Staff Software Engineer, Google Research

Understanding the aesthetic and technical quality of images is important for providing a better user visual experience. Image quality assessment (IQA) uses models to build a bridge between an image and a user’s subjective perception of its quality. In the deep learning era, many IQA approaches, such as NIMA, have achieved success by leveraging the power of convolutional neural networks (CNNs). However, CNN-based IQA models are often constrained by the fixed-size input requirement in batch training, i.e., the input images need to be resized or cropped to a fixed size shape. This preprocessing is problematic for IQA because images can have very different aspect ratios and resolutions. Resizing and cropping can impact image composition or introduce distortions, thus changing the quality of the image.

In CNN-based models, images need to be resized or cropped to a fixed shape for batch training. However, such preprocessing can alter the image aspect ratio and composition, thus impacting image quality. Original image used under CC BY 2.0 license.

In “MUSIQ: Multi-scale Image Quality Transformer”, published at ICCV 2021, we propose a patch-based multi-scale image quality transformer (MUSIQ) to bypass the CNN constraints on fixed input size and predict the image quality effectively on native-resolution images. The MUSIQ model supports the processing of full-size image inputs with varying aspect ratios and resolutions and allows multi-scale feature extraction to capture image quality at different granularities. To support positional encoding in the multi-scale representation, we propose a novel hash-based 2D spatial embedding combined with an embedding that captures the image scaling. We apply MUSIQ on four large-scale IQA datasets, demonstrating consistent state-of-the-art results across three technical quality datasets (PaQ-2-PiQ, KonIQ-10k, and SPAQ) and comparable performance to that of state-of-the-art models on the aesthetic quality dataset AVA.

The patch-based MUSIQ model can process the full-size image and extract multi-scale features, which better aligns with a person’s typical visual response.

In the following figure, we show a sample of images, their MUSIQ score, and their mean opinion score (MOS) from multiple human raters in the brackets. The range of the score is from 0 to 100, with 100 being the highest perceived quality. As we can see from the figure, MUSIQ predicts high scores for images with high aesthetic quality and high technical quality, and it predicts low scores for images that are not aesthetically pleasing (low aesthetic quality) or that contain visible distortions (low technical quality).

High quality
	76.10 [74.36]	69.29 [70.92]
Low aesthetics quality
	55.37 [53.18]	32.50 [35.47]
Low technical quality
	14.93 [14.38]	15.24 [11.86]

The Multi-scale Image Quality Transformer

MUSIQ tackles the challenge of learning IQA on full-size images. Unlike CNN-models that are often constrained to fixed resolution, MUSIQ can handle inputs with arbitrary aspect ratios and resolutions.

To accomplish this, we first make a multi-scale representation of the input image, containing the native resolution image and its resized variants. To preserve the image composition, we maintain its aspect ratio during resizing. After obtaining the pyramid of images, we then partition the images at different scales into fixed-size patches that are fed into the model.

Illustration of the multi-scale image representation in MUSIQ.

Since patches are from images of varying resolutions, we need to effectively encode the multi-aspect-ratio multi-scale input into a sequence of tokens, capturing both the pixel, spatial, and scale information. To achieve this, we design three encoding components in MUSIQ, including: 1) a patch encoding module to encode patches extracted from the multi-scale representation; 2) a novel hash-based spatial embedding module to encode the 2D spatial position for each patch; and 3) a learnable scale embedding to encode different scales. In this way, we can effectively encode the multi-scale input as a sequence of tokens, serving as the input to the Transformer encoder.

To predict the final image quality score, we use the standard approach of prepending an additional learnable “classification token” (CLS). The CLS token state at the output of the Transformer encoder serves as the final image representation. We then add a fully connected layer on top to predict the IQS. The figure below provides an overview of the MUSIQ model.

Overview of MUSIQ. The multi-scale multi-resolution input will be encoded by three components: the scale embedding (SCE), the hash-based 2D spatial embedding (HSE), and the multi-scale patch embedding (MPE).

Since MUSIQ only changes the input encoding, it is compatible with any Transformer variants. To demonstrate the effectiveness of the proposed method, in our experiments we use the classic Transformer with a relatively lightweight setting so that the model size is comparable to ResNet-50.

Benchmark and Evaluation

To evaluate MUSIQ, we run experiments on multiple large-scale IQA datasets. On each dataset, we report the Spearman’s rank correlation coefficient (SRCC) and Pearson linear correlation coefficient (PLCC) between our model prediction and the human evaluators’ mean opinion score. SRCC and PLCC are correlation metrics ranging from -1 to 1. Higher PLCC and SRCC means better alignment between model prediction and human evaluation. The graph below shows that MUSIQ outperforms other methods on PaQ-2-PiQ, KonIQ-10k, and SPAQ.

Performance comparison of MUSIQ and previous state-of-the-art (SOTA) methods on four large-scale IQA datasets. On each dataset we compare the Spearman’s rank correlation coefficient (SRCC) and Pearson linear correlation coefficient (PLCC) of model prediction and ground truth.

Notably, the PaQ-2-PiQ test set is entirely composed of large pictures having at least one dimension exceeding 640 pixels. This is very challenging for traditional deep learning approaches, which require resizing. MUSIQ can outperform previous methods by a large margin on the full-size test set, which verifies its robustness and effectiveness.

It is also worth mentioning that previous CNN-based methods often required sampling as many as 20 crops for each image during testing. This kind of multi-crop ensemble is a way to mitigate the fixed shape constraint in the CNN models. But since each crop is only a sub-view of the whole image, the ensemble is still an approximate approach. Moreover, CNN-based methods both add additional inference cost for every crop and, because they sample different crops, they can introduce randomness in the result. In contrast, because MUSIQ takes the full-size image as input, it can directly learn the best aggregation of information across the full image and it only needs to run the inference once.

To further verify that the MUSIQ model captures different information at different scales, we visualize the attention weights on each image at different scales.

Attention visualization from the output tokens to the multi-scale representation, including the original resolution image and two proportionally resized images. Brighter areas indicate higher attention, which means that those areas are more important for the model output. Images for illustration are taken from the AVA dataset.

We observe that MUSIQ tends to focus on more detailed areas in the full, high-resolution images and on more global areas on the resized ones. For example, for the flower photo above, the model’s attention on the original image is focusing on the pedal details, and the attention shifts to the buds at lower resolutions. This shows that the model learns to capture image quality at different granularities.

Conclusion

We propose a multi-scale image quality transformer (MUSIQ), which can handle full-size image input with varying resolutions and aspect ratios. By transforming the input image to a multi-scale representation with both global and local views, the model can capture the image quality at different granularities. Although MUSIQ is designed for IQA, it can be applied to other scenarios where task labels are sensitive to image resolution and aspect ratio. The MUSIQ model and checkpoints are available at our GitHub repository.

Acknowledgements

This work is made possible through a collaboration spanning several teams across Google. We’d like to acknowledge contributions from Qifei Wang, Yilin Wang and Peyman Milanfar.

Read More

Posted by Manoj Kumar, Research Engineer, and Ekin Dogus Cubuk, Research Scientist, Google Research

The task of determining the similarity between images is an open problem in computer vision and is crucial for evaluating the realism of machine-generated images. Though there are a number of straightforward methods of estimating image similarity (e.g., low-level metrics that measure pixel differences, such as FSIM and SSIM), in many cases, the measured similarity differences do not match the differences perceived by a person. However, more recent work has demonstrated that intermediate representations of neural network classifiers, such as AlexNet, VGG and SqueezeNet trained on ImageNet, exhibit perceptual similarity as an emergent property. That is, Euclidean distances between encoded representations of images by ImageNet-trained models correlate much better with a person’s judgment of differences between images than estimating perceptual similarity directly from image pixels.

Two sets of sample images from the BAPPS dataset. Trained networks agree more with human judgements as compared to low-level metrics (PSNR, SSIM, FSIM). Image source: Zhang et al. (2018).

In “Do better ImageNet classifiers assess perceptual similarity better?” published in Transactions on Machine Learning Research, we contribute an extensive experimental study on the relationship between the accuracy of ImageNet classifiers and their emergent ability to capture perceptual similarity. To evaluate this emergent ability, we follow previous work in measuring the perceptual scores (PS), which is roughly the correlation between human preferences to that of a model for image similarity on the BAPPS dataset. While prior work studied the first generation of ImageNet classifiers, such as AlexNet, SqueezeNet and VGG, we significantly increase the scope of the analysis incorporating modern classifiers, such as ResNets and Vision Transformers (ViTs), across a wide range of hyper-parameters.

Relationship Between Accuracy and Perceptual Similarity

It is well established that features learned via training on ImageNet transfer well to a number of downstream tasks, making ImageNet pre-training a standard recipe. Further, better accuracy on ImageNet usually implies better performance on a diverse set of downstream tasks, such as robustness to common corruptions, out-of-distribution generalization and transfer learning on smaller classification datasets. Contrary to prevailing evidence that suggests models with high validation accuracies on ImageNet are likely to transfer better to other tasks, surprisingly, we find that representations from underfit ImageNet models with modest validation accuracies achieve the best perceptual scores.

Plot of perceptual scores (PS) on the 64 × 64 BAPPS dataset (y-axis) against the ImageNet 64 × 64 validation accuracies (x-axis). Each blue dot represents an ImageNet classifier. Better ImageNet classifiers achieve better PS up to a certain point (dark blue), beyond which improving the accuracy lowers the PS. The best PS are attained by classifiers with moderate accuracy (20.0–40.0).

<!–

Plot of perceptual scores (PS) on the 64 × 64 BAPPS Dataset (y-axis) against the ImageNet 64 × 64 validation accuracies (x-axis). Each blue dot represents an ImageNet classifier. Better ImageNet classifiers achieve better PS up to a certain point (dark blue), beyond which improving the accuracy lowers the PS. The best PS are attained by classifiers with moderate accuracy (20.0–40.0).

–>

We study the variation of perceptual scores as a function of neural network hyperparameters: width, depth, number of training steps, weight decay, label smoothing and dropout. For each hyperparameter, there exists an optimal accuracy up to which improving accuracy improves PS. This optimum is fairly low and is attained quite early in the hyperparameter sweep. Beyond this point, improved classifier accuracy corresponds to worse PS.

As illustration, we present the variation of PS with respect to two hyperparameters: training steps in ResNets and width in ViTs. The PS of ResNet-50 and ResNet-200 peak very early at the first few epochs of training. After the peak, PS of better classifiers decrease more drastically. ResNets are trained with a learning rate schedule that causes a stepwise increase in accuracy as a function of training steps. Interestingly, after the peak, they also exhibit a step-wise decrease in PS that matches this step-wise accuracy increase.

Early-stopped ResNets attain the best PS across different depths of 6, 50 and 200.

ViTs consist of a stack of transformer blocks applied to the input image. The width of a ViT model is the number of output neurons of a single transformer block. Increasing its width is an effective way to improve its accuracy. Here, we vary the width of two ViT variants, B/8 and L/4 (i.e., Base and Large ViT models with patch sizes 4 and 8 respectively), and evaluate both the accuracy and PS. Similar to our observations with early-stopped ResNets, narrower ViTs with lower accuracies perform better than the default widths. Surprisingly, the optimal width of ViT-B/8 and ViT-L/4 are 6 and 12% of their default widths. For a more comprehensive list of experiments involving other hyperparameters such as width, depth, number of training steps, weight decay, label smoothing and dropout across both ResNets and ViTs, check out our paper.

Narrow ViTs attain the best PS.

Scaling Down Models Improves Perceptual Scores

Our results prescribe a simple strategy to improve an architecture’s PS: scale down the model to reduce its accuracy until it attains the optimal perceptual score. The table below summarizes the improvements in PS obtained by scaling down each model across every hyperparameter. Except for ViT-L/4, early stopping yields the highest improvement in PS, regardless of architecture. In addition, early stopping is the most efficient strategy as there is no need for an expensive grid search.

Global Perceptual Functions

In prior work, the perceptual similarity function was computed using Euclidean distances across the spatial dimensions of the image. This assumes a direct correspondence between pixels, which may not hold for warped, translated or rotated images. Instead, we adopt two perceptual functions that rely on global representations of images, namely the style-loss function from the Neural Style Transfer work that captures stylistic similarity between two images, and a normalized mean pool distance function. The style-loss function compares the inter-channel cross-correlation matrix between two images while the mean pool function compares the spatially averaged global representations.

Global perceptual functions consistently improve PS across both networks trained with default hyperparameters (top) and ResNet-200 as a function of train epochs (bottom).

We probe a number of hypotheses to explain the relationship between accuracy and PS and come away with a few additional insights. For example, the accuracy of models without commonly used skip-connections also inversely correlate with PS, and layers close to the input on average have lower PS as compared to layers close to the output. For further exploration involving distortion sensitivity, ImageNet class granularity, and spatial frequency sensitivity, check out our paper.

Conclusion

In this paper, we explore the question of whether improving classification accuracy yields better perceptual metrics. We study the relationship between accuracy and PS on ResNets and ViTs across many different hyperparameters and observe that PS exhibits an inverse-U relationship with accuracy, where accuracy correlates with PS up to a certain point, and then exhibits an inverse-correlation. Finally, in our paper, we discuss in detail a number of explanations for the observed relationship between accuracy and PS, involving skip connections, global similarity functions, distortion sensitivity, layerwise perceptual scores, spatial frequency sensitivity and ImageNet class granularity. While the exact explanation for the observed tradeoff between ImageNet accuracy and perceptual similarity is a mystery, we are excited that our paper opens the door for further research in this area.

Acknowledgements

This is joint work with Neil Houlsby and Nal Kalchbrenner. We would additionally like to thank Basil Mustafa, Kevin Swersky, Simon Kornblith, Johannes Balle, Mike Mozer, Mohammad Norouzi and Jascha Sohl-Dickstein for useful discussions.

Read More

Posted by Avi Singh, Research Scientist, and Laura Graesser, Research Engineer, Robotics at Google

Robot learning has been applied to a wide range of challenging real world tasks, including dexterous manipulation, legged locomotion, and grasping. It is less common to see robot learning applied to dynamic, high-acceleration tasks requiring tight-loop human-robot interactions, such as table tennis. There are two complementary properties of the table tennis task that make it interesting for robotic learning research. First, the task requires both speed and precision, which puts significant demands on a learning algorithm. At the same time, the problem is highly-structured (with a fixed, predictable environment) and naturally multi-agent (the robot can play with humans or another robot), making it a desirable testbed to investigate questions about human-robot interaction and reinforcement learning. These properties have led to several research groups developing table tennis research platforms [1, 2, 3, 4].

The Robotics team at Google has built such a platform to study problems that arise from robotic learning in a multi-player, dynamic and interactive setting. In the rest of this post we introduce two projects, Iterative-Sim2Real (to be presented at CoRL 2022) and GoalsEye (IROS 2022), which illustrate the problems we have been investigating so far. Iterative-Sim2Real enables a robot to hold rallies of over 300 hits with a human player, while GoalsEye enables learning goal-conditioned policies that match the precision of amateur humans.

Iterative-Sim2Real: Leveraging a Simulator to Play Cooperatively with Humans

In this project, the goal for the robot is cooperative in nature: to carry out a rally with a human for as long as possible. Since it would be tedious and time-consuming to train directly against a human player in the real world, we adopt a simulation-based (i.e., sim-to-real) approach. However, because it is difficult to simulate human behavior accurately, applying sim-to-real learning to tasks that require tight, close-loop interaction with a human participant is difficult.

In Iterative-Sim2Real, (i.e., i-S2R), we present a method for learning human behavior models for human-robot interaction tasks, and instantiate it on our robotic table tennis platform. We have built a system that can achieve rallies of up to 340 hits with an amateur human player (shown below).

Learning Human Behavior Models: a Chicken and Egg Problem

The central problem in learning accurate human behavior models for robotics is the following: if we do not have a good-enough robot policy to begin with, then we cannot collect high-quality data on how a person might interact with the robot. But without a human behavior model, we cannot obtain robot policies in the first place. An alternative would be to train a robot policy directly in the real world, but this is often slow, cost-prohibitive, and poses safety-related challenges, which are further exacerbated when people are involved. i-S2R, visualized below, is a solution to this chicken and egg problem. It uses a simple model of human behavior as an approximate starting point and alternates between training in simulation and deploying in the real world. In each iteration, both the human behavior model and the policy are refined.

i-S2R Methodology.

Results

To evaluate i-S2R, we repeated the training process five times with five different human opponents and compared it with a baseline approach of ordinary sim-to-real plus fine-tuning (S2R+FT). When aggregated across all players, the i-S2R rally length is higher than S2R+FT by about 9% (below on the left). The histogram of rally lengths for i-S2R and S2R+FT (below on the right) shows that a large fraction of the rallies for S2R+FT are shorter (i.e., less than 5), while i-S2R achieves longer rallies more frequently.

Summary of i-S2R results. Boxplot details: The white circle is the mean, the horizontal line is the median, box bounds are the 25th and 75th percentiles.

We also break down the results based on player type: beginner (40% players), intermediate (40% of players) and advanced (20% players). We see that i-S2R significantly outperforms S2R+FT for both beginner and intermediate players (80% of players).

i-S2R Results by player type.

More details on i-S2R can be found on our preprint, website, and also in the following summary video.

GoalsEye: Learning to Return Balls Precisely on a Physical Robot

While we focused on sim-to-real learning in i-S2R, it is sometimes desirable to learn using only real-world data — closing the sim-to-real gap in this case is unnecessary. Imitation learning (IL) provides a simple and stable approach to learning in the real world, but it requires access to demonstrations and cannot exceed the performance of the teacher. Collecting expert human demonstrations of precise goal-targeting in high speed settings is challenging and sometimes impossible (due to limited precision in human movements). While reinforcement learning (RL) is well-suited to such high-speed, high-precision tasks, it faces a difficult exploration problem (especially at the start), and can be very sample inefficient. In GoalsEye, we demonstrate an approach that combines recent behavior cloning techniques [5, 6] to learn a precise goal-targeting policy, starting from a small, weakly-structured, non-targeting dataset.

Here we consider a different table tennis task with an emphasis on precision. We want the robot to return the ball to an arbitrary goal location on the table, e.g. “hit the back left corner” or ”land the ball just over the net on the right side” (see left video below). Further, we wanted to find a method that can be applied directly on our real world table tennis environment with no simulation involved. We found that the synthesis of two existing imitation learning techniques, Learning from Play (LFP) and Goal-Conditioned Supervised Learning (GCSL), scales to this setting. It is safe and sample efficient enough to train a policy on a physical robot which is as accurate as amateur humans at the task of returning balls to specific goals on the table.

The essential ingredients of success are:

1. A minimal, but non-goal-directed “bootstrap” dataset of the robot hitting the ball to overcome an initial difficult exploration problem.
2. Hindsight relabeled goal conditioned behavioral cloning (GCBC) to train a goal-directed policy to reach any goal in the dataset.
3. Iterative self-supervised goal reaching. The agent improves continuously by setting random goals and attempting to reach them using the current policy. All attempts are relabeled and added into a continuously expanding training set. This self-practice, in which the robot expands the training data by setting and attempting to reach goals, is repeated iteratively.

GoalsEye methodology.

Demonstrations and Self-Improvement Through Practice Are Key

The synthesis of techniques is crucial. The policy’s objective is to return a variety of incoming balls to any location on the opponent’s side of the table. A policy trained on the initial 2,480 demonstrations only accurately reaches within 30 cm of the goal 9% of the time. However, after a policy has self-practiced for ~13,500 attempts, goal-reaching accuracy rises to 43% (below on the right). This improvement is clearly visible as shown in the videos below. Yet if a policy only self-practices, training fails completely in this setting. Interestingly, the number of demonstrations improves the efficiency of subsequent self-practice, albeit with diminishing returns. This indicates that demonstration data and self-practice could be substituted depending on the relative time and cost to gather demonstration data compared with self-practice.

Self-practice substantially improves accuracy. Left: simulated training. Right: real robot training. The demonstration datasets contain ~2,500 episodes, both in simulation and the real world.

More details on GoalsEye can be found in the preprint and on our website.

Conclusion and Future Work

We have presented two complementary projects using our robotic table tennis research platform. i-S2R learns RL policies that are able to interact with humans, while GoalsEye demonstrates that learning from real-world unstructured data combined with self-supervised practice is effective for learning goal-conditioned policies in a precise, dynamic setting.

One interesting research direction to pursue on the table tennis platform would be to build a robot “coach” that could adapt its play style according to the skill level of the human participant to keep things challenging and exciting.

Acknowledgements

We thank our co-authors, Saminda Abeyruwan, Alex Bewley, Krzysztof Choromanski, David B. D’Ambrosio, Tianli Ding, Deepali Jain, Corey Lynch, Pannag R. Sanketi, Pierre Sermanet and Anish Shankar. We are also grateful for the support of many members of the Robotics Team who are listed in the acknowledgement sections of the papers.

Read More

Posted by Yi Tay and Mostafa Dehghani, Research Scientists, Google Research, Brain Team

Building models that understand and generate natural language well is one the grand goals of machine learning (ML) research and has a direct impact on building smart systems for everyday applications. Improving the quality of language models is a key target for researchers to make progress toward such a goal.

Most common paradigms to build and train language models use either autoregressive decoder-only architectures (e.g., PaLM or GPT-3), where the model is trained to predict the next word for a given prefix phrase, or span corruption-based encoder-decoder architectures (e.g., T5, ST-MoE), where the training objective is to recover the subset of words masked out of the input. On the one hand, T5-like models perform well on supervised fine-tuning tasks, but struggle with few-shot in-context learning. On the other hand, autoregressive language models are great for open-ended generation (e.g., dialog generation with LaMDA) and prompt-based learning (e.g., in-context learning with PaLM), but may perform suboptimally on fine-tuning tasks. Thus, there remains an opportunity to create an effective unified framework for pre-training models.

In “Unifying Language Learning Paradigms”, we present a novel language pre-training paradigm called Unified Language Learner (UL2) that improves the performance of language models universally across datasets and setups. UL2 frames different objective functions for training language models as denoising tasks, where the model has to recover missing sub-sequences of a given input. During pre-training it uses a novel mixture-of-denoisers that samples from a varied set of such objectives, each with different configurations. We demonstrate that models trained using the UL2 framework perform well in a variety of language domains, including prompt-based few-shot learning and models fine-tuned for down-stream tasks. Additionally, we show that UL2 excels in generation, language understanding, retrieval, long-text understanding and question answering tasks. Finally, we are excited to publicly release the checkpoints for our best performing UL2 20 billion parameter model.

Background: Language Modeling Objectives and Architectures

Common objective functions for training language models can mostly be framed as learning data transformations that map inputs to targets. The model is conditioned on different forms of input to predict target tokens. To this end, different objectives utilize different properties of the inputs.

The standard Causal Language modeling objective (CausalLM) is trained to predict full sequence lengths and so, only recognizes tokens in the target output. The prefix language modeling objective (PrefixLM) modifies this process by randomly sampling a contiguous span of k tokens from the given tokenized text to form the input of the model, referred to as the “prefix”. The span corruption objective masks contiguous spans from the inputs and trains the model to predict these masked spans.

In the table below, we list the common objectives on which state-of-the-art language models are trained along with different characteristics of the input, i.e., how it is presented to the model. Moreover, we characterize the example efficiency of each objective in terms of the ability of the model for exploiting supervision signals from a single input, e.g., how much of the input tokens contribute to the calculation of the loss.

UL2 leverages the strengths of each of these objective functions through a framework that generalizes over each of them, which enables the ability to reason and unify common pre-training objectives. Based on this framework, the main task for training a language model is to learn the transformation of a sequence of input tokens to a sequence of target tokens. Then all the objective functions introduced above can be simply reduced to different ways of generating input and target tokens. For instance, the PrefixLM objective can be viewed as a transformation that moves a segment of k contiguous tokens from the inputs to the targets. Meanwhile, the span corruption objective is a data transformation that corrupts spans (a subsequence of tokens in the input), replacing them with mask tokens that are shifted to the targets.

It is worth noting that one can decouple the model architecture and the objective function with which it’s trained. Thus, it is possible to train different architectures, such as the common single stack decoder-only and two-stack encoder-decoder models, with any of these objectives.

Mixture of Denoisers

The UL2 framework can be used to train a model on a mixture of pre-training objectives and supply it with capabilities and inductive bias benefits from different pre-training tasks. Training on the mixture helps the model leverage the strengths of different tasks and mitigates the weaknesses of others. For instance, the mixture-of-denoisers objective can strongly improve the prompt-based learning capability of the model as opposed to a span corruption-only T5 model.

UL2 is trained using a mixture of three denoising tasks: (1) R-denoising (or regular span corruption), which emulates the standard T5 span corruption objective; (2) X-denoising (or extreme span corruption); and (3) S-denoising (or sequential PrefixLM). During pre-training, we sample from the available denoising tasks based on user-specified ratios (i.e., different combinations of the R, X, and S-denoisers) and prepare the input and target appropriately. Then, a paradigm token is appended to the input (one of [R], [X], or [S]) indicating the denoising task at hand.

An overview of the denoising objectives used in UL2’s mixture-of-denoisers.

Improving Trade-Offs Across Learning Paradigms

Many existing commonly used language learning paradigms typically excel at one type of task or application, such as fine-tuning performance or prompt-based in-context learning. In the plot below, we show baseline objective functions on different tasks compared to UL2: CausalLM (referred to as GPT-like), PrefixLM, Span Corrupt (also referred to as T5 in the plot), and a baseline objective function proposed by UniLM. We use these objectives for training decoder only architectures (green) and encoder-decoder architectures (blue) and evaluate different combinations of objective functions and architectures on two main sets of tasks:

1. Fine-tuning, by measuring performance on SuperGLUE (y-axis of the plot below)
2. In-context learning, by measuring performance of the model on a suite of 1-shot GEM tasks (e.g., XSUM, SGD or Schema guided dialog and TOTTO) (x-axis of the plot below).

For most of the existing language learning paradigms, there is a trade-off between the quality of the model on these two sets of tasks. We show that UL2 bridges this trade-off across in-context learning and fine-tuning.

In both decoder-only and encoder-decoder setups, UL2 strikes a significantly improved balance in performance between fine-tuned discriminative tasks and prompt-based 1-shot open-ended text generation compared to previous methods. (All models are comparable in terms of computational costs, i.e., FLOPs (EncDec models are 300M and Dec models are 150M parameters).

UL2 for Few-Shot Prompting and Chain-of-Thought Reasoning

We scale up UL2 and train a 20 billion parameter encoder-decoder model on the public C4 corpus and demonstrate some impressive capabilities of the UL2 20B model.

UL2 is a powerful in-context learner that excels at both few-shot and chain-of-thought (CoT) prompting. In the table below, we compare UL2 with other state-of-the-art models (e.g, T5 XXL and PaLM) for few-shot prompting on the XSUM summarization dataset. Our results show that UL2 20B outperforms PaLM and T5, both of which are in the same ballpark of compute cost.

Most CoT prompting results have been obtained using much larger language models, such as GPT-3 175B, PaLM 540B, or LaMDA 137B. We show that reasoning via CoT prompting can be achieved with UL2 20B, which is both publicly available and several times smaller than prior models that leverage chain-of-thought prompting. This enables an open avenue for researchers to conduct research on CoT prompting and reasoning at an accessible scale. In the table below, we show that for UL2, CoT prompting outperforms standard prompting on math word problems with a range of difficulties (GSM8K, SVAMP, ASDiv, AQuA, and MAWPS). We also show that self-consistency further improves performance.

Chain-of-thought (CoT) prompting and self-consistency (SC) results on five arithmetic reasoning benchmarks.

Conclusion and Future Directions

UL2 demonstrates superior performance on a plethora of fine-tuning and few-shot tasks. We publicly release checkpoints of our best performing UL2 model with 20 billion parameters, which we hope will inspire faster progress in developing better language models in the machine learning community as a whole.

Acknowledgements

It was an honor and privilege to work on this with Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Denny Zhou, Neil Houlsby and Donald Metzler. We further acknowledge Alexey Gritsenko, Andrew M. Dai, Jacob Devlin, Jai Gupta, William Fedus, Orhan Firat, Sebastian Gerhmann, Nan Du, Dave Uthus, Siamak Shakeri, Slav Petrov and Quoc Le for support and discussions. We thank the Jax and T5X team for building such wonderful infrastructure that made this research possible.

Read More

Posted by Ashish Thapliyal, Software Engineer, and Jordi Pont-Tuset, Research Scientist, Google Research

Image captioning is the machine learning task of automatically generating a fluent natural language description for a given image. This task is important for improving accessibility for visually impaired users and is a core task in multimodal research encompassing both vision and language modeling.

However, datasets for image captioning are primarily available in English. Beyond that, there are only a few datasets covering a limited number of languages that represent just a small fraction of the world’s population. Further, these datasets feature images that severely under-represent the richness and diversity of cultures from across the globe. These aspects have hindered research on image captioning for a wide variety of languages, and directly hamper the deployment of accessibility solutions for a large potential audience around the world.

Today we present and make publicly available the Crossmodal 3600 (XM3600) image captioning evaluation dataset as a robust benchmark for multilingual image captioning that enables researchers to reliably compare research contributions in this emerging field. XM3600 provides 261,375 human-generated reference captions in 36 languages for a geographically diverse set of 3600 images. We show that the captions are of high quality and the style is consistent across languages.

Overview of the Crossmodal 3600 Dataset

Creating large training and evaluation datasets in multiple languages is a resource-intensive endeavor. Recent work has shown that it is feasible to build multilingual image captioning models trained on machine-translated data with English captions as the starting point. However, some of the most reliable automatic metrics for image captioning are much less effective when applied to evaluation sets with translated image captions, resulting in poorer agreement with human evaluations compared to the English case. As such, trustworthy model evaluation at present can only be based on extensive human evaluation. Unfortunately, such evaluations usually cannot be replicated across different research efforts, and therefore do not offer a fast and reliable mechanism to automatically evaluate multiple model parameters and configurations (e.g., model hill climbing) or to compare multiple lines of research.

XM3600 provides 261,375 human-generated reference captions in 36 languages for a geographically diverse set of 3600 images from the Open Images dataset. We measure the quality of generated captions by comparing them to the manually provided captions using the CIDEr metric, which ranges from 0 (unrelated to the reference captions) to 10 (perfectly matching the reference captions). When comparing pairs of models, we observed strong correlations between the differences in the CIDEr scores of the model outputs, and side-by-side human evaluations comparing the model outputs. , making XM3600 is a reliable tool for high-quality automatic comparisons between image captioning models on a wide variety of languages beyond English.

Language Selection

We chose 30 languages beyond English, roughly based on their percentage of web content. In addition, we chose an additional five languages that include under-resourced languages that have many native speakers or major native languages from continents that would not be covered otherwise. Finally, we also included English as a baseline, thus resulting in a total of 36 languages, as listed in the table below.

Image Selection

The images were selected from among those in the Open Images dataset that have location metadata. Since there are many regions where more than one language is spoken, and some areas are not well covered by these images, we designed an algorithm to maximize the correspondence between selected images and the regions where the targeted languages are spoken. The algorithm starts with the selection of images with geo-data corresponding to the languages for which we have the smallest pool (e.g., Persian) and processes them in increasing order of their candidate image pool size. If there aren’t enough images in an area where a language is spoken, then we gradually expand the geographic selection radius to: (i) a country where the language is spoken; (ii) a continent where the language is spoken; and, as last resort, (iii) from anywhere in the world. This strategy succeeded in providing our target number of 100 images from an appropriate region for most of the 36 languages, except for Persian (where 14 continent-level images are used) and Hindi (where all 100 images are at the global level, because the in-region images were assigned to Bengali and Telugu).

English Photo by Chris Sampson		Swahili Photo by Henrik Palm		Telugu Photo by rojypala
Cusco Quechua Photo by McKay Savage		Filipino Photo by Simon Schoeters		Chinese Photo by Stefan Krasowski

Caption Generation

In total, all 3600 images (100 images per language) are annotated in all 36 languages, each with an average of two annotations per language, yielding a total of 261,375 captions.

Annotators work in batches of 15 images. The first screen shows all 15 images with their captions in English as generated by a captioning model trained to output a consistent style of the form “<main salient objects> doing <activities> in the <environment>”, often with object attributes, such as a “smiling” person, “red” car, etc. The annotators are asked to rate the caption quality given guidelines for a 4-point scale from “excellent” to “bad”, plus an option for “not_enough_information”. This step forces the annotators to carefully assess caption quality and it primes them to internalize the style of the captions. The following screens show the images again but individually and without the English captions, and the annotators are asked to produce descriptive captions in the target language for each image.

The image batch size of 15 was chosen so that the annotators would internalize the style without remembering the exact captions. Thus, we expect the raters to generate captions based on the image content only and lacking translation artifacts. For example in the example shown below, the Spanish caption mentions “number 42” and the Thai caption mentions “convertibles”, none of which are mentioned in the English captions. The annotators were also provided with a protocol to use when creating the captions, thus achieving style consistency across languages.

Photo by Brian Solis		English		• A vintage sports car in a showroom with many other vintage sports cars
				• The branded classic cars in a row at display
		Spanish		• Automóvil clásico deportivo en exhibición de automóviles de galería — (Classic sports car in gallery car show)
				• Coche pequeño de carreras color plateado con el número 42 en una exhibición de coches — (Small silver racing car with the number 42 at a car show)
		Thai		• รถเปิดประทุนหลายสีจอดเรียงกันในที่จัดแสดง — (Multicolored convertibles line up in the exhibit)
				• รถแข่งวินเทจจอดเรียงกันหลายคันในงานจัดแสดง — (Several vintage racing cars line up at the show.)

Caption Quality and Statistics

We ran two to five pilot studies per language to troubleshoot the caption generation process and to ensure high quality captions. We then manually evaluated a random subset of captions. First we randomly selected a sample of 600 images. Then, to measure the quality of captions in a particular language, for each image, we selected for evaluation one of the manually generated captions. We found that:

• For 25 out of 36 languages, the percentage of captions rated as “Good” or “Excellent” is above 90%, and the rest are all above 70%.
• For 26 out of 36 languages, the percentage of captions rated as “Bad” is below 2%, and the rest are all below 5%.

For languages that use spaces to separate words, the number of words per caption can be as low as 5 or 6 for some agglutinative languages like Cusco Quechua and Czech, and as high as 18 for an analytic language like Vietnamese. The number of characters per caption also varies drastically — from mid-20s for Korean to mid-90s for Indonesian — depending on the alphabet and the script of the language.

Empirical Evaluation and Results

We empirically measured the ability of the XM3600 annotations to rank image captioning model variations by training four variations of a multilingual image captioning model and comparing the CIDEr differences of the models’ outputs over the XM3600 dataset for 30+ languages, to side-by-side human evaluations. We observed strong correlations between the CIDEr differences and the human evaluations. These results support the use of the XM3600 references as a means to achieve high-quality automatic comparisons between image captioning models on a wide variety of languages beyond English.

Recent Uses

Recently PaLI used XM3600 to evaluate model performance beyond English for image captioning, image-to-text retrieval and text-to-image retrieval. The key takeaways they found when evaluating on XM3600 were that multilingual captioning greatly benefits from scaling the PaLI models, especially for low-resource languages.

Acknowledgements

We would like to acknowledge the coauthors of this work: Xi Chen and Radu Soricut.

Read More

10:25

Read More

Read More

Posted by Zalán Borsos, Research Software Engineer, and Neil Zeghidour, Research Scientist, Google Research

Generating realistic audio requires modeling information represented at different scales. For example, just as music builds complex musical phrases from individual notes, speech combines temporally local structures, such as phonemes or syllables, into words and sentences. Creating well-structured and coherent audio sequences at all these scales is a challenge that has been addressed by coupling audio with transcriptions that can guide the generative process, be it text transcripts for speech synthesis or MIDI representations for piano. However, this approach breaks when trying to model untranscribed aspects of audio, such as speaker characteristics necessary to help people with speech impairments recover their voice, or stylistic components of a piano performance.

In “AudioLM: a Language Modeling Approach to Audio Generation”, we propose a new framework for audio generation that learns to generate realistic speech and piano music by listening to audio only. Audio generated by AudioLM demonstrates long-term consistency (e.g., syntax in speech, melody in music) and high fidelity, outperforming previous systems and pushing the frontiers of audio generation with applications in speech synthesis or computer-assisted music. Following our AI Principles, we’ve also developed a model to identify synthetic audio generated by AudioLM.

From Text to Audio Language Models

In recent years, language models trained on very large text corpora have demonstrated their exceptional generative abilities, from open-ended dialogue to machine translation or even common-sense reasoning. They have further shown their capacity to model other signals than texts, such as natural images. The key intuition behind AudioLM is to leverage such advances in language modeling to generate audio without being trained on annotated data.

However, some challenges need to be addressed when moving from text language models to audio language models. First, one must cope with the fact that the data rate for audio is significantly higher, thus leading to much longer sequences — while a written sentence can be represented by a few dozen characters, its audio waveform typically contains hundreds of thousands of values. Second, there is a one-to-many relationship between text and audio. This means that the same sentence can be rendered by different speakers with different speaking styles, emotional content and recording conditions.

To overcome both challenges, AudioLM leverages two kinds of audio tokens. First, semantic tokens are extracted from w2v-BERT, a self-supervised audio model. These tokens capture both local dependencies (e.g., phonetics in speech, local melody in piano music) and global long-term structure (e.g., language syntax and semantic content in speech, harmony and rhythm in piano music), while heavily downsampling the audio signal to allow for modeling long sequences.

However, audio reconstructed from these tokens demonstrates poor fidelity. To overcome this limitation, in addition to semantic tokens, we rely on acoustic tokens produced by a SoundStream neural codec, which capture the details of the audio waveform (such as speaker characteristics or recording conditions) and allow for high-quality synthesis. Training a system to generate both semantic and acoustic tokens leads simultaneously to high audio quality and long-term consistency.

Training an Audio-Only Language Model

AudioLM is a pure audio model that is trained without any text or symbolic representation of music. AudioLM models an audio sequence hierarchically, from semantic tokens up to fine acoustic tokens, by chaining several Transformer models, one for each stage. Each stage is trained for the next token prediction based on past tokens, as one would train a text language model. The first stage performs this task on semantic tokens to model the high-level structure of the audio sequence.

In the second stage, we concatenate the entire semantic token sequence, along with the past coarse acoustic tokens, and feed both as conditioning to the coarse acoustic model, which then predicts the future tokens. This step models acoustic properties such as speaker characteristics in speech or timbre in music.

In the third stage, we process the coarse acoustic tokens with the fine acoustic model, which adds even more detail to the final audio. Finally, we feed acoustic tokens to the SoundStream decoder to reconstruct a waveform.

After training, one can condition AudioLM on a few seconds of audio, which enables it to generate consistent continuation. In order to showcase the general applicability of the AudioLM framework, we consider two tasks from different audio domains:

• Speech continuation, where the model is expected to retain the speaker characteristics, prosody and recording conditions of the prompt while producing new content that is syntactically correct and semantically consistent.
• Piano continuation, where the model is expected to generate piano music that is coherent with the prompt in terms of melody, harmony and rhythm.

In the video below, you can listen to examples where the model is asked to continue either speech or music and generate new content that was not seen during training. As you listen, note that everything you hear after the gray vertical line was generated by AudioLM and that the model has never seen any text or musical transcription, but rather just learned from raw audio. We release more samples on this webpage.

To validate our results, we asked human raters to listen to short audio clips and decide whether it is an original recording of human speech or a synthetic continuation generated by AudioLM. Based on the ratings collected, we observed a 51.2% success rate, which is not statistically significantly different from the 50% success rate achieved when assigning labels at random. This means that speech generated by AudioLM is hard to distinguish from real speech for the average listener.

Our work on AudioLM is for research purposes and we have no plans to release it more broadly at this time. In alignment with our AI Principles, we sought to understand and mitigate the possibility that people could misinterpret the short speech samples synthesized by AudioLM as real speech. For this purpose, we trained a classifier that can detect synthetic speech generated by AudioLM with very high accuracy (98.6%). This shows that despite being (almost) indistinguishable to some listeners, continuations generated by AudioLM are very easy to detect with a simple audio classifier. This is a crucial first step to help protect against the potential misuse of AudioLM, with future efforts potentially exploring technologies such as audio “watermarking”.

Conclusion

We introduce AudioLM, a language modeling approach to audio generation that provides both long-term coherence and high audio quality. Experiments on speech generation show not only that AudioLM can generate syntactically and semantically coherent speech without any text, but also that continuations produced by the model are almost indistinguishable from real speech by humans. Moreover, AudioLM goes well beyond speech and can model arbitrary audio signals such as piano music. This encourages the future extensions to other types of audio (e.g., multilingual speech, polyphonic music, and audio events) as well as integrating AudioLM into an encoder-decoder framework for conditioned tasks such as text-to-speech or speech-to-speech translation.

Acknowledgments

The work described here was authored by Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Olivier Teboul, David Grangier, Marco Tagliasacchi and Neil Zeghidour. We are grateful for all discussions and feedback on this work that we received from our colleagues at Google.

Read More