ICLR 2022 Papers

Learning multiple tasks sequentially without forgetting previous knowledge, called Continual Learning(CL), remains a long-standing challenge for neural networks. Most existing methods rely on additional network capacity or data replay. In contrast, we introduce a novel approach which we refer to as Recursive Gradient Optimization(RGO). RGO is composed of an iteratively updated optimizer that modifies the gradient to minimize forgetting without data replay and a virtual Feature Encoding Layer(FEL) that represents different long-term structures with only task descriptors. Experiments demonstrate that RGO has significantly better performance on popular continual classification benchmarks when compared to the baselines and achieves new state-of-the-art performance on 20-split-CIFAR100(82.22%) and 20-split-miniImageNet(72.63%). With higher average accuracy than Single-Task Learning(STL), this method is flexible and reliable to provide continual learning capabilities for learning models that rely on gradient descent.

Most graph neural networks (GNNs) use the message passing paradigm, in which node features are propagated on the input graph. Recent works pointed to the distortion of information flowing from distant nodes as a factor limiting the efficiency of message passing for tasks relying on long-distance interactions. This phenomenon, referred to as 'over-squashing', has been heuristically attributed to graph bottlenecks where the number of $k$-hop neighbors grows rapidly with $k$. We provide a precise description of the over-squashing phenomenon in GNNs and analyze how it arises from bottlenecks in the graph. For this purpose, we introduce a new edge-based combinatorial curvature and prove that negatively curved edges are responsible for the over-squashing issue. We also propose and experimentally test a curvature-based graph rewiring method to alleviate the over-squashing.

Class-conditioning offers a direct means to control a Generative Adversarial Network (GAN) based on a discrete input variable. While necessary in many applications, the additional information provided by the class labels could even be expected to benefit the training of the GAN itself. On the contrary, we observe that class-conditioning causes mode collapse in limited data settings, where unconditional learning leads to satisfactory generative ability. Motivated by this observation, we propose a training strategy for class-conditional GANs (cGANs) that effectively prevents the observed mode-collapse by leveraging unconditional learning. Our training strategy starts with an unconditional GAN and gradually injects the class conditioning into the generator and the objective function. The proposed method for training cGANs with limited data results not only in stable training but also in generating high-quality images, thanks to the early-stage exploitation of the shared information across classes. We analyze the observed mode collapse problem in comprehensive experiments on four datasets. Our approach demonstrates outstanding results compared with state-of-the-art methods and established baselines. The code is available at https://github.com/mshahbazi72/transitional-cGAN

This paper studies an important problem in computational chemistry: predicting a molecule's spatial atom arrangements, or a molecular conformation. We propose a neural energy minimization formulation that casts the prediction problem into an unrolled optimization process, where a neural network is parametrized to learn the gradient fields of an implicit conformational energy landscape. Assuming different forms of the underlying potential energy function, we can not only reinterpret and unify many of the existing models but also derive new variants of SE(3)-equivariant neural networks in a principled manner. In our experiments, these new variants show superior performance in molecular conformation optimization comparing to existing SE(3)-equivariant neural networks. Moreover, our energy-inspired formulation is also suitable for molecular conformation generation, where we can generate more diverse and accurate conformers comparing to existing baselines.

We present a conditional variational auto-encoder (VAE) which, to avoid the substantial cost of training from scratch, uses an architecture and training objective capable of leveraging a foundation model in the form of a pretrained unconditional VAE. To train the conditional VAE, we only need to train an artifact to perform amortized inference over the unconditional VAE's latent variables given a conditioning input. We demonstrate our approach on tasks including image inpainting, for which it outperforms state-of-the-art GAN-based approaches at faithfully representing the inherent uncertainty. We conclude by describing a possible application of our inpainting model, in which it is used to perform Bayesian experimental design for the purpose of guiding a sensor.

In many reinforcement learning (RL) applications, the observation space is specified by human developers and restricted by physical realizations, and may thus be subject to dramatic changes over time (e.g. increased number of observable features). However, when the observation space changes, the previous policy will likely fail due to the mismatch of input features, and another policy must be trained from scratch, which is inefficient in terms of computation and sample complexity. Following theoretical insights, we propose a novel algorithm which extracts the latent-space dynamics in the source task, and transfers the dynamics model to the target task to use as a model-based regularizer. Our algorithm works for drastic changes of observation space (e.g. from vector-based observation to image-based observation), without any inter-task mapping or any prior knowledge of the target task. Empirical results show that our algorithm significantly improves the efficiency and stability of learning in the target task.

We study the problem of learning a good set of policies, so that when combined together, they can solve a wide variety of unseen reinforcement learning tasks with no or very little new data. Specifically, we consider the framework of generalized policy evaluation and improvement, in which the rewards for all tasks of interest are assumed to be expressible as a linear combination of a fixed set of features. We show theoretically that, under certain assumptions, having access to a specific set of diverse policies, which we call a set of independent policies, can allow for instantaneously achieving high-level performance on all possible downstream tasks which are typically more complex than the ones on which the agent was trained. Based on this theoretical analysis, we propose a simple algorithm that iteratively constructs this set of policies. In addition to empirically validating our theoretical results, we compare our approach with recently proposed diverse policy set construction methods and show that, while others fail, our approach is able to build a behavior basis that enables instantaneous transfer to all possible downstream tasks. We also show empirically that having access to a set of independent policies can better bootstrap the learning process on downstream tasks where the new reward function cannot be described as a linear combination of the features. Finally, we demonstrate how this policy set can be useful in a lifelong reinforcement learning setting.

We prove that $\mathbb{F}_p$ sketch, a well-celebrated streaming algorithm for frequency moments estimation, is differentially private as is when $p\in(0, 1]$. $\mathbb{F}_p$ sketch uses only polylogarithmic space, exponentially better than existing DP baselines and only worse than the optimal non-private baseline by a logarithmic factor. The evaluation shows that $\mathbb{F}_p$ sketch can achieve reasonable accuracy with strong privacy guarantees. The code for evaluation is included in the supplementary material.

Equivariant neural networks enforce symmetry within the structure of their convolutional layers, resulting in a substantial improvement in sample efficiency when learning an equivariant or invariant function. Such models are applicable to robotic manipulation learning which can often be formulated as a rotationally symmetric problem. This paper studies equivariant model architectures in the context of $Q$-learning and actor-critic reinforcement learning. We identify equivariant and invariant characteristics of the optimal $Q$-function and the optimal policy and propose equivariant DQN and SAC algorithms that leverage this structure. We present experiments that demonstrate that our equivariant versions of DQN and SAC can be significantly more sample efficient than competing algorithms on an important class of robotic manipulation problems.

We propose a novel probabilistic framework for modeling stochastic dynamics with the rigorous use of stochastic optimal control theory. The proposed model called the neural Markov controlled stochastic differential equation (CSDE) overcomes the fundamental and structural limitations of conventional dynamical models by introducing the following two components: (1) Markov dynamic programming to efficiently train the proposed CSDE and (2) multi-conditional forward-backward losses to provide rich information for accurate inference and to assure theoretical optimality. We demonstrate that our dynamical model efficiently generates a complex time series in the data space without extra networks while showing comparable performance against existing model-based methods on several datasets.

Spiking Neural Networks (SNNs) have gained great attraction due to their distinctive properties of low power consumption and fast inference on neuromorphic hardware. As the most effective method to get deep SNNs, ANN-SNN conversion has achieved comparable performance as ANNs on large-scale datasets. Despite this, it requires long time-steps to match the firing rates of SNNs to the activation of ANNs. As a result, the converted SNN suffers severe performance degradation problems with short time-steps, which hamper the practical application of SNNs. In this paper, we theoretically analyze ANN-SNN conversion error and derive the estimated activation function of SNNs. Then we propose the quantization clip-floor-shift activation function to replace the ReLU activation function in source ANNs, which can better approximate the activation function of SNNs. We prove that the expected conversion error between SNNs and ANNs is zero, enabling us to achieve high-accuracy and ultra-low-latency SNNs. We evaluate our method on CIFAR-10/100 and ImageNet datasets, and show that it outperforms the state-of-the-art ANN-SNN and directly trained SNNs in both accuracy and time-steps. To the best of our knowledge, this is the first time to explore high-performance ANN-SNN conversion with ultra-low latency (4 time-steps). Code is available at https://github.com/putshua/SNN_conversion_QCFS

Reinforcement Learning (RL) has achieved significant successes, which aims to obtain a single policy maximizing the expected cumulative rewards for a given task. However, in many real-world scenarios, e.g., navigating in complex environments and controlling robots, one may need to find a set of policies having both high rewards and diverse behaviors, which can bring better exploration and robust few-shot adaptation. Recently, some methods have been developed by using evolutionary techniques, including iterative reproduction and selection of policies. However, due to the inefficient selection mechanisms, these methods cannot fully guarantee both high quality and diversity. In this paper, we propose EDO-CS, a new Evolutionary Diversity Optimization algorithm with Clustering-based Selection. In each iteration, the policies are divided into several clusters based on their behaviors, and a high-quality policy is selected from each cluster for reproduction. EDO-CS also adaptively balances the importance between quality and diversity in the reproduction process. Experiments on various (i.e., deceptive and multi-modal) continuous control tasks, show the superior performance of EDO-CS over previous methods, i.e., EDO-CS can achieve a set of policies with both high quality and diversity efficiently while previous methods cannot.

Despite our best efforts, deep learning models remain highly vulnerable to even tiny adversarial perturbations applied to the inputs. The ability to extract information from solely the output of a machine learning model to craft adversarial perturbations to black-box models is a practical threat against real-world systems, such as Machine Learning as a Service (MLaaS), particularly $sparse~attacks$. The realization of sparse attacks in black-box settings demonstrates that machine learning models are more vulnerable than we believe. Because, these attacks aim to $minimize~the~number~of~perturbed~pixels$—measured by $l_0$ norm—required to mislead a model by $solely$ observing the decision ($the~predicted~label$) returned to a model query; the so-called $decision-based~setting$. But, such an attack leads to an NP-hard optimization problem. We develop an evolution-based algorithm—$SparseEvo$—for the problem and evaluate it against both convolutional deep neural networks and $vision~transformers$. Notably, vision transformers are yet to be investigated under a decision-based attack setting. SparseEvo requires significantly fewer queries than the state-of-the-art sparse attack $Pointwise$ for both untargeted and targeted attacks. The attack algorithm, although conceptually simple, is competitive with only a limited query budget against the state-of-the-art gradient-based $white-box$ attacks in standard computer vision tasks such as $ImageNet$. Importantly, the query efficient SparseEvo, along with decision-based attacks, in general, raises new questions regarding the safety of deployed systems and poses new directions to study and understand the robustness of machine learning models.

Semi-Supervised Learning (SSL) is fundamentally a missing label problem, in which the label Missing Not At Random (MNAR) problem is more realistic and challenging, compared to the widely-adopted yet naive Missing Completely At Random assumption where both labeled and unlabeled data share the same class distribution. Different from existing SSL solutions that overlook the role of ''class'' in causing the non-randomness, e.g., users are more likely to label popular classes, we explicitly incorporate ''class'' into SSL. Our method is three-fold: 1) We propose Class-Aware Propensity (CAP) that exploits the unlabeled data to train an improved classifier using the biased labeled data. 2) To encourage rare class training, whose model is low-recall but high-precision that discards too many pseudo-labeled data, we propose Class-Aware Imputation (CAI) that dynamically decreases (or increases) the pseudo-label assignment threshold for rare (or frequent) classes. 3) Overall, we integrate CAP and CAI into a Class-Aware Doubly Robust (CADR) estimator for training an unbiased SSL model. Under various MNAR settings and ablations, our method not only significantly outperforms existing baselines, but also surpasses other label bias removal SSL methods.

Deep generative models have been widely used in practical applications such as the detection of out-of-distribution (OOD) data. In this work, we aim to re-examine the potential of generative flow models in OOD detection. We first propose a simple combination of univariate one-sample statistical test (e.g., Kolmogorov-Smirnov) and random projections in the latent space of flow models to perform OOD detection. Then, we propose a two-sample version of our test to account for imperfect flow models. Quite distinctly, our method does not pose parametric assumptions on OOD data and is capable of exploiting any flow model. Experimentally, firstly we confirm the efficacy of our method against state-of-the-art baselines through extensive experiments on several image datasets; secondly we investigate the relationship between model accuracy (e.g., the generation quality) and the OOD detection performance, and found surprisingly that they are not always positively correlated; and thirdly we show that detection in the latent space of flow models generally outperforms detection in the sample space across various OOD datasets, hence highlighting the benefits of training a flow model.

Humans decompose novel complex tasks into simpler ones to exploit previously learned skills. Analogously, hierarchical reinforcement learning seeks to leverage lower-level policies for simple tasks to solve complex ones. However, because each lower-level policy induces a different distribution of states, transitioning from one lower-level policy to another may fail due to an unexpected starting state. We introduce transition policies that smoothly connect lower-level policies by producing a distribution of states and actions that matches what is expected by the next policy. Training transition policies is challenging because the natural reward signal---whether the next policy can execute its subtask successfully---is sparse. By training transition policies via adversarial inverse reinforcement learning to match the distribution of expected states and actions, we avoid relying on task-based reward. To further improve performance, we use deep Q-learning with a binary action space to determine when to switch from a transition policy to the next pre-trained policy, using the success or failure of the next subtask as the reward. Although the reward is still sparse, the problem is less severe due to the simple binary action space. We demonstrate our method on continuous bipedal locomotion and arm manipulation tasks that require diverse skills. We show that it smoothly connects the lower-level policies, achieving higher success rates than previous methods that search for successful trajectories based on a reward function, but do not match the state distribution.

We present ReCo, a contrastive learning framework designed at a regional level to assist learning in semantic segmentation. ReCo performs pixel-level contrastive learning on a sparse set of hard negative pixels, with minimal additional memory footprint. ReCo is easy to implement, being built on top of off-the-shelf segmentation networks, and consistently improves performance, achieving more accurate segmentation boundaries and faster convergence. The strongest effect is in semi-supervised learning with very few labels. With ReCo, we achieve high quality semantic segmentation model, requiring only 5 examples of each semantic class.

We present a method for training neural network models with discrete stochastic variables. The core of the method is \emph{stability regularization}, which is a regularization procedure based on the idea of noise stability developed in Gaussian isoperimetric theory in the analysis of Gaussian functions. Stability regularization is method to make the output of continuous functions of Gaussian random variables close to discrete, that is binary or categorical, without the need for significant manual tuning. The method allows control over the extent to which a Gaussian function's output is close to discrete, thus allowing for continued flow of gradient. The method can be used standalone or in combination with existing continuous relaxation methods. We validate the method in a broad range of experiments using discrete variables including neural relational inference, generative modeling, clustering and conditional computing.

Molecule representation learning (MRL) methods aim to embed molecules into a real vector space. However, existing SMILES-based (Simplified Molecular-Input Line-Entry System) or GNN-based (Graph Neural Networks) MRL methods either take SMILES strings as input that have difficulty in encoding molecule structure information, or over-emphasize the importance of GNN architectures but neglect their generalization ability. Here we propose using chemical reactions to assist learning molecule representation. The key idea of our approach is to preserve the equivalence of molecules with respect to chemical reactions in the embedding space, i.e., forcing the sum of reactant embeddings and the sum of product embeddings to be equal for each chemical equation. This constraint is proven effective to 1) keep the embedding space well-organized and 2) improve the generalization ability of molecule embeddings. Moreover, our model can use any GNN as the molecule encoder and is thus agnostic to GNN architectures. Experimental results demonstrate that our method achieves state-of-the-art performance in a variety of downstream tasks, e.g., reaction product prediction, molecule property prediction, reaction classification, and graph-edit-distance prediction. The code is available at https://github.com/hwwang55/MolR.