ICML 2022 Papers

We present a simple and effective self-supervised learning approach for speech recognition. The approach learns a model to predict the masked speech signals, in the form of discrete labels generated with a random-projection quantizer. In particular the quantizer projects speech inputs with a randomly initialized matrix, and does a nearest-neighbor lookup in a randomly-initialized codebook. Neither the matrix nor the codebook are updated during self-supervised learning. Since the random-projection quantizer is not trained and is separated from the speech recognition model, the design makes the approach flexible and is compatible with universal speech recognition architecture. On LibriSpeech our approach achieves similar word-error-rates as previous work using self-supervised learning with non-streaming models, and provides lower word-error-rates than previous work with streaming models. On multilingual tasks the approach also provides significant improvement over wav2vec 2.0 and w2v-BERT.

Transformers have shown remarkable progress on computer vision tasks in the past year. Compared to their CNN counterparts, transformers usually need the help of distillation to achieve comparable results on middle or small sized datasets. Meanwhile, recent researches discover that when transformers are trained with supervised and self-supervised manner respectively, the captured patterns are quite different both qualitatively and quantitatively. These findings motivate us to introduce an self-supervised teaching assistant (SSTA) besides the commonly used supervised teacher to improve the performance of transformers. Specifically, we propose a head-level knowledge distillation method that selects the most important head of the supervised teacher and self-supervised teaching assistant, and let the student mimic the attention distribution of these two heads, so as to make the student focus on the relationship between tokens deemed by the teacher and the teacher assistant. Extensive experiments verify the effectiveness of SSTA and demonstrate that the proposed SSTA is a good compensation to the supervised teacher. Meanwhile, some analytical experiments towards multiple perspectives (e.g. prediction, shape bias, robustness, and transferability to downstream tasks) with supervised teachers, self-supervised teaching assistants and students are inductive and may inspire future researches.

In this paper we aim to investigate the mechanisms that guide text generation with pre-trained Transformer-based Language Models (TLMs). Grounded on the Product of Experts formulation by Hinton (1999), we describe a generative mechanism that exploits expert units which naturally exist in TLMs. Such units are responsible for detecting concepts in the input and conditioning text generation on such concepts. We describe how to identify expert units and how to activate them during inference in order to induce any desired concept in the generated output. We find that the activation of a surprisingly small amount of units is sufficient to steer text generation (as little as 3 units in a model with 345M parameters). While the objective of this work is to learn more about how TLMs work, we show that our method is effective for conditioning without fine-tuning or using extra parameters, even on fine-grained homograph concepts. Additionally, we show that our method can be used to correct gender bias present in the output of TLMs and achieves gender parity for all evaluated contexts. We compare our method with FUDGE and PPLM-BoW, and show that our approach is able to achieve gender parity at a lower perplexity and better Self-BLEU score. The proposed method is accessible to a wide audience thanks to its simplicity and minimal compute needs. The findings in this paper are a step forward in understanding the generative mechanisms of TLMs.

Self-supervised learning (SSL) of graph neural networks is emerging as a promising way of leveraging unlabeled data. Currently, most methods are based on contrastive learning adapted from the image domain, which requires view generation and a sufficient number of negative samples. In contrast, existing predictive models do not require negative sampling, but lack theoretical guidance on the design of pretext training tasks. In this work, we propose the LaGraph, a theoretically grounded predictive SSL framework based on latent graph prediction. Learning objectives of LaGraph are derived as self-supervised upper bounds to objectives for predicting unobserved latent graphs. In addition to its improved performance, LaGraph provides explanations for recent successes of predictive models that include invariance-based objectives. We provide theoretical analysis comparing LaGraph to related methods in different domains. Our experimental results demonstrate the superiority of LaGraph in performance and the robustness to decreasing of training sample size on both graph-level and node-level tasks.

Self-supervised language models are very effective at predicting high-level cortical responses during language comprehension. However, the best current models of lower-level auditory processing in the human brain rely on either hand-constructed acoustic filters or representations from supervised audio neural networks. In this work, we capitalize on the progress of self-supervised speech representation learning (SSL) to create new state-of-the-art models of the human auditory system. Compared against acoustic baselines, phonemic features, and supervised models, representations from the middle layers of self-supervised models (APC, wav2vec, wav2vec 2.0, and HuBERT) consistently yield the best prediction performance for fMRI recordings within the auditory cortex (AC). Brain areas involved in low-level auditory processing exhibit a preference for earlier SSL model layers, whereas higher-level semantic areas prefer later layers. We show that these trends are due to the models’ ability to encode information at multiple linguistic levels (acoustic, phonetic, and lexical) along their representation depth. Overall, these results show that self-supervised models effectively capture the hierarchy of information relevant to different stages of speech processing in human cortex.

Coordination graph is a promising approach to model agent collaboration in multi-agent reinforcement learning. It conducts a graph-based value factorization and induces explicit coordination among agents to complete complicated tasks. However, one critical challenge in this paradigm is the complexity of greedy action selection with respect to the factorized values. It refers to the decentralized constraint optimization problem (DCOP), which and whose constant-ratio approximation are NP-hard problems. To bypass this systematic hardness, this paper proposes a novel method, named Self-Organized Polynomial-time Coordination Graphs (SOP-CG), which uses structured graph classes to guarantee the accuracy and the computational efficiency of collaborated action selection. SOP-CG employs dynamic graph topology to ensure sufficient value function expressiveness. The graph selection is unified into an end-to-end learning paradigm. In experiments, we show that our approach learns succinct and well-adapted graph topologies, induces effective coordination, and improves performance across a variety of cooperative multi-agent tasks.

Selective regression allows abstention from prediction if the confidence to make an accurate prediction is not sufficient. In general, by allowing a reject option, one expects the performance of a regression model to increase at the cost of reducing coverage (i.e., by predicting on fewer samples). However, as we show, in some cases, the performance of a minority subgroup can decrease while we reduce the coverage, and thus selective regression can magnify disparities between different sensitive subgroups. Motivated by these disparities, we propose new fairness criteria for selective regression requiring the performance of every subgroup to improve with a decrease in coverage. We prove that if a feature representation satisfies the

Private inference (PI) enables inferences directly on cryptographically secure data. While promising to address many privacy issues, it has seen limited use due to extreme runtimes. Unlike plaintext inference, where latency is dominated by FLOPs, in PI non-linear functions (namely ReLU) are the bottleneck. Thus, practical PI demands novel ReLU-aware optimizations. To reduce PI latency we propose a gradient-based algorithm that selectively linearizes ReLUs while maintaining prediction accuracy. We evaluate our algorithm on several standard PI benchmarks. The results demonstrate up to $4.25%$ more accuracy (iso-ReLU count at 50K) or $2.2\times$ less latency (iso-accuracy at 70%) than the current state of the art and advance the Pareto frontier across the latency-accuracy space. To complement empirical results, we present a “no free lunch" theorem that sheds light on how and when network linearization is possible while maintaining prediction accuracy.

We implement training of neural networks in secure multi-party computation (MPC) using quantization commonly used in said setting. We are the first to present an MNIST classifier purely trained in MPC that comes within 0.2 percent of the accuracy of the same convolutional neural network trained via plaintext computation. More concretely, we have trained a network with two convolutional and two dense layers to 99.2% accuracy in 3.5 hours (under one hour for 99% accuracy). We have also implemented AlexNet for CIFAR-10, which converges in a few hours. We develop novel protocols for exponentiation and inverse square root. Finally, we present experiments in a range of MPC security models for up to ten parties, both with honest and dishonest majority as well as semi-honest and malicious security.

Many areas of deep learning benefit from using increasingly larger neural networks trained on public data, as is the case for pre-trained models for NLP and computer vision. Training such models requires a lot of computational resources (e.g., HPC clusters) that are not available to small research groups and independent researchers. One way to address it is for several smaller groups to pool their computational resources together and train a model that benefits all participants. Unfortunately, in this case, any participant can jeopardize the entire training run by sending incorrect updates, deliberately or by mistake. Training in presence of such peers requires specialized distributed training algorithms with Byzantine tolerance. These algorithms often sacrifice efficiency by introducing redundant communication or passing all updates through a trusted server, making it infeasible to apply them to large-scale deep learning, where models can have billions of parameters. In this work, we propose a novel protocol for secure (Byzantine-tolerant) decentralized training that emphasizes communication efficiency.

With the success of Transformers in the computer vision field, the automated design of vision Transformers has attracted significant attention. Recently, MetaFormer found that simple average pooling can achieve impressive performance, which naturally raises the question of how to design a search space to search diverse and high-performance Transformer-like architectures. By revisiting typical search spaces, we design micro-meso-macro space to search for Transformer-like architectures, namely BurgerFormer. Micro, meso, and macro correspond to the granularity levels of operation, block and stage, respectively. At the microscopic level, we enrich the atomic operations to include various normalizations, activation functions, and basic operations (e.g., multi-head self attention, average pooling). At the mesoscopic level, a hamburger structure is searched out as the basic BurgerFormer block. At the macroscopic level, we search for the depth, width, and expansion ratio of the network based on the multi-stage architecture. Meanwhile, we propose a hybrid sampling method for effectively training the supernet. Experimental results demonstrate that the searched BurgerFormer architectures achieve comparable even superior performance compared with current state-of-the-art Transformers on the ImageNet and COCO datasets. The codes can be available at https://github.com/xingxing-123/BurgerFormer.

Generating graph-structured data requires learning the underlying distribution of graphs. Yet, this is a challenging problem, and the previous graph generative methods either fail to capture the permutation-invariance property of graphs or cannot sufficiently model the complex dependency between nodes and edges, which is crucial for generating real-world graphs such as molecules. To overcome such limitations, we propose a novel score-based generative model for graphs with a continuous-time framework. Specifically, we propose a new graph diffusion process that models the joint distribution of the nodes and edges through a system of stochastic differential equations (SDEs). Then, we derive novel score matching objectives tailored for the proposed diffusion process to estimate the gradient of the joint log-density with respect to each component, and introduce a new solver for the system of SDEs to efficiently sample from the reverse diffusion process. We validate our graph generation method on diverse datasets, on which it either achieves significantly superior or competitive performance to the baselines. Further analysis shows that our method is able to generate molecules that lie close to the training distribution yet do not violate the chemical valency rule, demonstrating the effectiveness of the system of SDEs in modeling the node-edge relationships.

We prove fast mixing and characterize the stationary distribution of the Langevin Algorithm for inverting random weighted DNN generators. This result extends the work of Hand and Voroninski from efficient inversion to efficient posterior sampling. In practice, to allow for increased expressivity, we propose to do posterior sampling in the latent space of a pre-trained generative model. To achieve that, we train a score-based model in the latent space of a StyleGAN-2 and we use it to solve inverse problems. Our framework, Score-Guided Intermediate Layer Optimization (SGILO), extends prior work by replacing the sparsity regularization with a generative prior in the intermediate layer. Experimentally, we obtain significant improvements over the previous state-of-the-art, especially in the low measurement regime.

Enforcing orthogonality in convolutional neural networks is a remedy for gradient vanishing/exploding problems and sensitivity to perturbation. Many previous approaches for orthogonal convolutions enforce orthogonality on its flattened kernel, which, however, do not lead to the orthogonality of the operation. Some recent approaches consider orthogonality for standard convolutional layers and propose specific classes of their realizations. In this work, we propose a theoretical framework that establishes the equivalence between diverse orthogonal convolutional layers in the spatial domain and the paraunitary systems in the spectral domain. Since 1D paraunitary systems admit a complete factorization, we can parameterize any separable orthogonal convolution as a composition of spatial filters. As a result, our framework endows high expressive power to various convolutional layers while maintaining their exact orthogonality. Furthermore, our layers are memory and computationally efficient for deep networks compared to previous designs. Our versatile framework, for the first time, enables the study of architectural designs for deep orthogonal networks, such as choices of skip connection, initialization, stride, and dilation. Consequently, we scale up orthogonal networks to deep architectures, including ResNet and ShuffleNet, substantially outperforming their shallower counterparts. Finally, we show how to construct residual flows, a flow-based generative model that requires strict Lipschitzness, using our orthogonal networks. Our code will be publicly available at https://github.com/umd-huang-lab/ortho-conv

Deep discrete structured models have seen considerable progress recently, but traditional inference using dynamic programming (DP) typically works with a small number of states (less than hundreds), which severely limits model capacity. At the same time, across machine learning, there is a recent trend of using randomized truncation techniques to accelerate computations involving large sums. Here, we propose a family of randomized dynamic programming (RDP) algorithms for scaling structured models to tens of thousands of latent states. Our method is widely applicable to classical DP-based inference (partition, marginal, reparameterization, entropy) and different graph structures (chains, trees, and more general hypergraphs). It is also compatible with automatic differentiation: it can be integrated with neural networks seamlessly and learned with gradient-based optimizers. Our core technique approximates the sum-product by restricting and reweighting DP on a small subset of nodes, which reduces computation by orders of magnitude. We further achieve low bias and variance via Rao-Blackwellization and importance sampling. Experiments over different graphs demonstrate the accuracy and efficiency of our approach. Furthermore, when using RDP for training a structured variational autoencoder with a scaled inference network, we achieve better test likelihood than baselines and successfully prevent posterior collapse.

Detecting out-of-distribution examples is important for safety-critical machine learning applications such as detecting novel biological phenomena and self-driving cars. However, existing research mainly focuses on simple small-scale settings. To set the stage for more realistic out-of-distribution detection, we depart from small-scale settings and explore large-scale multiclass and multi-label settings with high-resolution images and thousands of classes. To make future work in real-world settings possible, we create new benchmarks for three large-scale settings. To test ImageNet multiclass anomaly detectors, we introduce the Species dataset containing over 700,000 images and over a thousand anomalous species. We leverage ImageNet-21K to evaluate PASCAL VOC and COCO multilabel anomaly detectors. Third, we introduce a new benchmark for anomaly segmentation by introducing a segmentation benchmark with road anomalies. We conduct extensive experiments in these more realistic settings for out-of-distribution detection and find that a surprisingly simple detector based on the maximum logit outperforms prior methods in all the large-scale multi-class, multi-label, and segmentation tasks, establishing a simple new baseline for future work.

Computing a Gaussian process (GP) posterior has a computational cost cubical in the number of historical points. A reformulation of the same GP posterior highlights that this complexity mainly depends on how many

Spike-and-slab priors are commonly used for Bayesian variable selection, due to their interpretability and favorable statistical properties. However, existing samplers for spike-and-slab posteriors incur prohibitive computational costs when the number of variables is large. In this article, we propose Scalable Spike-and-Slab (S^3), a scalable Gibbs sampling implementation for high-dimensional Bayesian regression with the continuous spike-and-slab prior of George & McCulloch (1993). For a dataset with n observations and p covariates, S^3 has order max{n^2 p_t, np} computational cost at iteration t where p_t never exceeds the number of covariates switching spike-and-slab states between iterations t and t-1 of the Markov chain. This improves upon the order n^2 p per-iteration cost of state-of-the-art implementations as, typically, p_t is substantially smaller than p. We apply S^3 on synthetic and real-world datasets, demonstrating orders of magnitude speed-ups over existing exact samplers and significant gains in inferential quality over approximate samplers with comparable cost.

Bayesian Optimization (BO) has shown great promise for the global optimization of functions that are expensive to evaluate, but despite many successes, standard approaches can struggle in high dimensions. To improve the performance of BO, prior work suggested incorporating gradient information into a Gaussian process surrogate of the objective, giving rise to kernel matrices of size $nd$ {\texttimes} $nd$ for $n$ observations in $d$ dimensions. Naı̈vely multiplying with (resp. inverting) these matrices requires $O(n^2d^2)$ (resp. $O(n^3d^3)$) operations, which becomes infeasible for moderate dimensions and sample sizes. Here, we observe that a wide range of kernels gives rise to structured matrices, enabling an exact $O(n^2d)$ matrix-vector multiply for gradient observations and $O(n^2d^2)$ for Hessian observations. Beyond canonical kernel classes, we derive a programmatic approach to leveraging this type of structure for transformations and combinations of the discussed kernel classes, which constitutes a structure-aware automatic differentiation algorithm. Our methods apply to virtually all canonical kernels and automatically extend to complex kernels, like the neural network, radial basis function network, and spectral mixture kernels without any additional derivations, enabling flexible, problem-dependent modeling while scaling first-order BO to high $d$.