ICLR 2022 Papers

Meta-Imitation Learning is a promising technique for the robot to learn a new task from observing one or a few human demonstrations. However, it usually requires a significant number of demonstrations both from humans and robots during the meta-training phase, which is a laborious and hard work for data collection, especially in recording the actions and specifying the correspondence between human and robot. In this work, we present an approach of meta-imitation learning by watching video demonstrations from humans. In comparison to prior works, our approach is able to translate human videos into practical robot demonstrations and train the meta-policy with adaptive loss based on the quality of the translated data. Our approach relies only on human videos and does not require robot demonstration, which facilitates data collection and is more in line with human imitation behavior. Experiments reveal that our method achieves the comparable performance to the baseline on fast learning a set of vision-based tasks through watching a single video demonstration.

Currently, it is hard to reap the benefits of deep learning for Bayesian methods, which allow the explicit specification of prior knowledge and accurately capture model uncertainty. We present Prior-Data Fitted Networks (PFNs). PFNs leverage large-scale machine learning techniques to approximate a large set of posteriors. The only requirement for PFNs to work is the ability to sample from a prior distribution over supervised learning tasks (or functions). Our method restates the objective of posterior approximation as a supervised classification problem with a set-valued input: it repeatedly draws a task (or function) from the prior, draws a set of data points and their labels from it, masks one of the labels and learns to make probabilistic predictions for it based on the set-valued input of the rest of the data points. Presented with a set of samples from a new supervised learning task as input, PFNs make probabilistic predictions for arbitrary other data points in a single forward propagation, having learned to approximate Bayesian inference. We demonstrate that PFNs can near-perfectly mimic Gaussian processes and also enable efficient Bayesian inference for intractable problems, with over 200-fold speedups in multiple setups compared to current methods. We obtain strong results in very diverse areas such as Gaussian process regression, Bayesian neural networks, classification for small tabular data sets, and few-shot image classification, demonstrating the generality of PFNs. Code and trained PFNs are released at https://github.com/automl/TransformersCanDoBayesianInference.

In Multiple Instance Learning (MIL), models are trained using bags of instances, where only a single label is provided for each bag. A bag label is often only determined by a handful of key instances within a bag, making it difficult to interpret what information a classifier is using to make decisions. In this work, we establish the key requirements for interpreting MIL models. We then go on to develop several model-agnostic approaches that meet these requirements. Our methods are compared against existing inherently interpretable MIL models on several datasets, and achieve an increase in interpretability accuracy of up to 30%. We also examine the ability of the methods to identify interactions between instances and scale to larger datasets, improving their applicability to real-world problems.

Differentiable physics modeling combines physics models with gradient-based learning to provide model explicability and data efficiency. It has been used to learn dynamics, solve inverse problems and facilitate design, and is at its inception of impact. Current successes have concentrated on general physics models such as rigid bodies, deformable sheets, etc, assuming relatively simple structures and forces. Their granularity is intrinsically coarse and therefore incapable of modelling complex physical phenomena. Fine-grained models are still to be developed to incorporate sophisticated material structures and force interactions with gradient-based learning. Following this motivation, we propose a new differentiable fabrics model for composite materials such as cloths, where we dive into the granularity of yarns and model individual yarn physics and yarn-to-yarn interactions. To this end, we propose several differentiable forces, whose counterparts in empirical physics are indifferentiable, to facilitate gradient-based learning. These forces, albeit applied to cloths, are ubiquitous in various physical systems. Through comprehensive evaluation and comparison, we demonstrate our model's $\textit{explicability}$ in learning meaningful physical parameters, $\textit{versatility}$ in incorporating complex physical structures and heterogeneous materials, $\textit{data-efficiency}$ in learning, and $\textit{high-fidelity}$ in capturing subtle dynamics.

Offline reinforcement learning, which seeks to utilize offline/historical data to optimize sequential decision-making strategies, has gained surging prominence in recent studies. Due to the advantage that appropriate function approximators can help mitigate the sample complexity burden in modern reinforcement learning problems, existing endeavors usually enforce powerful function representation models (e.g. neural networks) to learn the optimal policies. However, a precise understanding of the statistical limits with function representations, remains elusive, even when such a representation is linear. Towards this goal, we study the statistical limits of offline reinforcement learning with linear model representations. To derive the tight offline learning bound, we design the variance-aware pessimistic value iteration (VAPVI), which adopts the conditional variance information of the value function for time-inhomogeneous episodic linear Markov decision processes (MDPs). VAPVI leverages estimated variances of the value functions to reweight the Bellman residuals in the least-square pessimistic value iteration and provides improved offline learning bounds over the best-known existing results (whereas the Bellman residuals are equally weighted by design). More importantly, our learning bounds are expressed in terms of system quantities, which provide natural instance-dependent characterizations that previous results are short of. We hope our results draw a clearer picture of what offline learning should look like when linear representations are provided.

Solving goal-conditioned tasks with sparse rewards using self-supervised learning is promising because of its simplicity and stability over current reinforcement learning (RL) algorithms. A recent work, called Goal-Conditioned Supervised Learning (GCSL), provides a new learning framework by iteratively relabeling and imitating self-generated experiences. In this paper, we revisit the theoretical property of GCSL --- optimizing a lower bound of the goal reaching objective, and extend GCSL as a novel offline goal-conditioned RL algorithm. The proposed method is named Weighted GCSL (WGCSL), in which we introduce an advanced compound weight consisting of three parts (1) discounted weight for goal relabeling, (2) goal-conditioned exponential advantage weight, and (3) best-advantage weight. Theoretically, WGCSL is proved to optimize an equivalent lower bound of the goal-conditioned RL objective and generates monotonically improved policies via an iterated scheme. The monotonic property holds for any behavior policies, and therefore WGCSL can be applied to both online and offline settings. To evaluate algorithms in the offline goal-conditioned RL setting, we provide a benchmark including a range of point and simulated robot domains. Experiments in the introduced benchmark demonstrate that WGCSL can consistently outperform GCSL and existing state-of-the-art offline methods in the fully offline goal-conditioned setting.

Learning on graphs has attracted significant attention in the learning community due to numerous real-world applications. In particular, graph neural networks (GNNs), which take \emph{numerical} node features and graph structure as inputs, have been shown to achieve state-of-the-art performance on various graph-related learning tasks. Recent works exploring the correlation between numerical node features and graph structure via self-supervised learning have paved the way for further performance improvements of GNNs. However, methods used for extracting numerical node features from \emph{raw data} are still \emph{graph-agnostic} within standard GNN pipelines. This practice is sub-optimal as it prevents one from fully utilizing potential correlations between graph topology and node attributes. To mitigate this issue, we propose a new self-supervised learning framework, Graph Information Aided Node feature exTraction (GIANT). GIANT makes use of the eXtreme Multi-label Classification (XMC) formalism, which is crucial for fine-tuning the language model based on graph information, and scales to large datasets. We also provide a theoretical analysis that justifies the use of XMC over link prediction and motivates integrating XR-Transformers, a powerful method for solving XMC problems, into the GIANT framework. We demonstrate the superior performance of GIANT over the standard GNN pipeline on Open Graph Benchmark datasets: For example, we improve the accuracy of the top-ranked method GAMLP from $68.25\%$ to $69.67\%$, SGC from $63.29\%$ to $66.10\%$ and MLP from $47.24\%$ to $61.10\%$ on the ogbn-papers100M dataset by leveraging GIANT.

This work explores how to design a single neural network capable of adapting to multiple heterogeneous vision tasks, such as image segmentation, 3D detection, and video recognition. This goal is challenging because both network architecture search (NAS) spaces and methods in different tasks are inconsistent. We solve this challenge from both sides. We first introduce a unified design space for multiple tasks and build a multitask NAS benchmark (NAS-Bench-MR) on many widely used datasets, including ImageNet, Cityscapes, KITTI, and HMDB51. We further propose Network Coding Propagation (NCP), which back-propagates gradients of neural predictors to directly update architecture codes along the desired gradient directions to solve various tasks. In this way, optimal architecture configurations can be found by NCP in our large search space in seconds. Unlike prior arts of NAS that typically focus on a single task, NCP has several unique benefits. (1) NCP transforms architecture optimization from data-driven to architecture-driven, enabling joint search an architecture among multitasks with different data distributions. (2) NCP learns from network codes but not original data, enabling it to update the architecture efficiently across datasets. (3) In addition to our NAS-Bench-MR, NCP performs well on other NAS benchmarks, such as NAS-Bench-201. (4) Thorough studies of NCP on inter-, cross-, and intra-tasks highlight the importance of cross-task neural architecture design, i.e., multitask neural architectures and architecture transferring between different tasks. Code is available at https://github.com/dingmyu/NCP.

Despite progress across a broad range of applications, Transformers have limited success in systematic generalization. The situation is especially frustrating in the case of algorithmic tasks, where they often fail to find intuitive solutions that route relevant information to the right node/operation at the right time in the grid represented by Transformer columns. To facilitate the learning of useful control flow, we propose two modifications to the Transformer architecture, copy gate and geometric attention. Our novel Neural Data Router (NDR) achieves 100% length generalization accuracy on the classic compositional table lookup task, as well as near-perfect accuracy on the simple arithmetic task and a new variant of ListOps testing for generalization across computational depths. NDR’s attention and gating patterns tend to be interpretable as an intuitive form of neural routing

Goal-conditioned reinforcement learning (RL) has shown great success recently at solving a wide range of tasks(e.g., navigation, robotic manipulation). However, learning to reach distant goals remains a central challenge to the field, and the task is particularly hard without any offline data, expert demonstrations, and reward shaping. In this paper, we propose to solve the distant goal-reaching task by using search at training time to generate a curriculum of intermediate states. Specifically, we introduce the algorithm Classifier-Planning (C-Planning) by framing the learning of the goal-conditioned policies as variational inference. C-Planning naturally follows expectation maximization (EM): the E-step corresponds to planning an optimal sequence of waypoints using graph search, while the M-step aims to learn a goal-conditioned policy to reach those waypoints. One essential difficulty of designing such an algorithm is accurately modeling the distribution over way-points to sample from. In C-Planning, we propose to sample the waypoints using contrastive methods to learn a value function. Unlike prior methods that combine goal-conditioned RL with graph search, ours performs search only during training and not testing, significantly decreasing the compute costs of deploying the learned policy. Empirically, we demonstrate that our method not only improves the sample efficiency of prior methods but also successfully solves temporally extended navigation and manipulation tasks, where prior goal-conditioned RL methods (including those based on graph search) fail to solve.

A simple and natural algorithm for reinforcement learning (RL) is Monte Carlo Exploring Starts (MCES), where the Q-function is estimated by averaging the Monte Carlo returns, and the policy is improved by choosing actions that maximize the current estimate of the Q-function. Exploration is performed by "exploring starts", that is, each episode begins with a randomly chosen state and action, and then follows the current policy to the terminal state. In the classic book on RL by Sutton & Barto (2018), it is stated that establishing convergence for the MCES algorithm is one of the most important remaining open theoretical problems in RL. However, the convergence question for MCES turns out to be quite nuanced. Bertsekas & Tsitsiklis (1996) provide a counter-example showing that the MCES algorithm does not necessarily converge. Tsitsiklis (2002) further shows that if the original MCES algorithm is modified so that the Q-function estimates are updated at the same rate for all state-action pairs, and the discount factor is strictly less than one, then the MCES algorithm converges. In this paper we make headway with the original and more efficient MCES algorithm given in Sutton et al. (1998), establishing almost sure convergence for Optimal Policy Feed-Forward MDPs, which are MDPs whose states are not revisited within any episode when using an optimal policy. Such MDPs include a large class of environments such as all deterministic environments and all episodic environments with a timestep or any monotonically changing values as part of the state. Different from the previous proofs using stochastic approximations, we introduce a novel inductive approach, which is very simple and only makes use of the strong law of large numbers.

Discovery and learning of an underlying spatiotemporal hierarchy in sequential data is an important topic for machine learning. Despite this, little work has been done to explore hierarchical generative models that can flexibly adapt their layerwise representations in response to datasets with different temporal dynamics. Here, we present Variational Predictive Routing (VPR) – a neural probabilistic inference system that organizes latent representations of video features in a temporal hierarchy, based on their rates of change, thus modeling continuous data as a hierarchical renewal process. By employing an event detection mechanism that relies solely on the system’s latent representations (without the need of a separate model), VPR is able to dynamically adjust its internal state following changes in the observed features, promoting an optimal organisation of representations across the levels of the model’s latent hierarchy. Using several video datasets, we show that VPR is able to detect event boundaries, disentangle spatiotemporal features across its hierarchy, adapt to the dynamics of the data, and produce accurate time-agnostic rollouts of the future. Our approach integrates insights from neuroscience and introduces a framework with high potential for applications in model-based reinforcement learning, where flexible and informative state-space rollouts are of particular interest.

State-of-the-art models in natural language processing rely on separate rigid subword tokenization algorithms, which limit their generalization ability and adaptation to new settings. In this paper, we propose a new model inductive bias that learns a subword tokenization end-to-end as part of the model. To this end, we introduce a soft gradient-based subword tokenization module (GBST) that automatically learns latent subword representations from characters in a data-driven fashion. Concretely, GBST enumerates candidate subword blocks and learns to score them in a position-wise fashion using a block scoring network. We additionally introduce Charformer, a deep Transformer model that integrates GBST and operates on the character level. Via extensive experiments on English GLUE, multilingual, and noisy text datasets, we show that Charformer outperforms a series of competitive character-level baselines while generally performing on par and sometimes outperforming subword-based models. Additionally, Charformer is fast, improving the speed of vanilla character-level Transformers by up to while maintaining quality. We believe this work paves the way for highly performant token-free models that are trained completely end-to-end.

The pretrain-finetune paradigm has shown outstanding performance on many applications of deep learning, where a model is pre-trained on an upstream large dataset (e.g. ImageNet), and is then fine-tuned to different downstream tasks. Though for most cases, the pre-training stage is conducted based on supervised methods, recent works on self-supervised pre-training have shown powerful transferability and even outperform supervised pre-training on multiple downstream tasks. It thus remains an open question how to better generalize supervised pre- training model to downstream tasks. In this paper, we argue that the worse transferability of existing supervised pre-training methods arise from the negligence of valuable intra-class semantic difference. This is because these methods tend to push images from the same class close to each other despite of the large diversity in their visual contents, a problem to which referred as “overfit of upstream tasks”. To alleviate this problem, we propose a new supervised pre-training method based on Leave-One-Out K-Nearest-Neighbor, or LOOK for short. It relieves the problem of overfitting upstream tasks by only requiring each image to share its class label with most of its k nearest neighbors, thus allowing each class to exhibit a multi-mode distribution and consequentially preserving part of intra-class difference for better transferring to downstream tasks. We developed efficient implementation of the proposed method that scales well to large datasets. Experimental studies on multiple downstream tasks show that LOOK outperforms other state-of-the-art methods for supervised and self-supervised pre-training.

In contrast to standard statistical wisdom, modern learning algorithms typically find their best performance in the overparameterized regime in which the model has many more parameters than needed to fit the training data. A growing number of recent works have shown that random feature models can offer a detailed theoretical explanation for this unexpected behavior, but typically these analyses have utilized isotropic distributional assumptions on the underlying data generation process, thereby failing to provide a realistic characterization of real-world models that are designed to identify and harness the structure in natural data. In this work, we examine the high-dimensional asymptotics of random feature regression in the presence of structured data, allowing for arbitrary input correlations and arbitrary alignment between the data and the weights of the target function. We define a partial order on the space of weight-data alignments and prove that generalization performance improves in response to stronger alignment. We also clarify several previous observations in the literature by distinguishing the behavior of the sample-wise and parameter-wise learning curves, finding that sample-wise multiple descent can occur at scales dictated by the eigenstructure of the data covariance, but that parameter-wise multiple descent is limited to double descent, although strong anisotropy can induce additional signatures such as wide plateaus and steep cliffs. Finally, these signatures are related to phase transitions in the spectrum of the feature kernel matrix, and unlike the double descent peak, persist even under optimal regularization.

To solve the imbalance classification, methods of weighting examples have been proposed. Recent work has studied to assign adaptive weights to training examples through learning mechanisms, that is, the weights, similar to classification models, are regarded as parameters that need to be learned. However, the algorithms in recent work use local information to approximately optimize the weights, which may lead to inaccurate learning of the weights. In this work, we first propose a novel mechanism of learning with a constraint, which can accurately train the weights and model. Then, we propose a combined method of our learning mechanism and the work by Hu et al., which can promote each other to perform better. Our proposed method can be applied to any type of deep network model. Experiments show that compared with the state-of-the-art algorithms, our method has significant improvement in varieties of settings, including text and image classification over different imbalance ratios, binary and multi-class classification.

Deep neural network (DNN) model compression for efficient on-device inference is becoming increasingly important to reduce memory requirements and keep user data on-device. To this end, we propose a novel differentiable k-means clustering layer (DKM) and its application to train-time weight clustering-based DNN model compression. DKM casts k-means clustering as an attention problem and enables joint optimization of the DNN parameters and clustering centroids. Unlike prior works that rely on additional regularizers and parameters, DKM-based compression keeps the original loss function and model architecture fixed. We evaluated DKM-based compression on various DNN models for computer vision and natural language processing (NLP) tasks. Our results demonstrate that DKM delivers superior compression and accuracy trade-off on ImageNet1k and GLUE benchmarks. For example, DKM-based compression can offer 74.5% top-1 ImageNet1k accuracy on ResNet50 DNN model with 3.3MB model size (29.4x model compression factor). For MobileNet-v1, which is a challenging DNN to compress, DKM delivers 63.9% top-1 ImageNet1k accuracy with 0.72 MB model size (22.4x model compression factor). This result is 6.8% higher top-1accuracy and 33% relatively smaller model size than the current state-of-the-art DNN compression algorithms. Additionally, DKM enables compression of DistilBERT model by 11.8x with minimal (1.1%) accuracy loss on GLUE NLP benchmarks.

One potential drawback of using aggregated performance measurement in machine learning is that models may learn to accept higher errors on some training cases as compromises for lower errors on others, with the lower errors actually being instances of overfitting. This can lead both to stagnation at local optima and to poor generalization. Lexicase selection is an uncompromising method developed in evolutionary computation, which selects models on the basis of sequences of individual training case errors instead of using aggregated metrics such as loss and accuracy. In this paper, we investigate how the general idea of lexicase selection can fit into the context of deep learning to improve generalization. We propose Gradient Lexicase Selection, an optimization framework that combines gradient descent and lexicase selection in an evolutionary fashion. Experimental results show that the proposed method improves the generalization performance of various popular deep neural network architectures on three image classification benchmarks. Qualitative analysis also indicates that our method helps the networks learn more diverse representations.

Reward hacking---where RL agents exploit gaps in misspecified proxy rewards---has been widely observed, but not yet systematically studied. To understand reward hacking, we construct four RL environments with different misspecified rewards. We investigate reward hacking as a function of agent capabilities: model capacity, action space resolution, and observation space noise. Typically, more capable agents are able to better exploit reward misspecifications, causing them to attain higher proxy reward and lower true reward. Moreover, we find instances of \emph{phase transitions}: capability thresholds at which the agent's behavior qualitatively shifts, leading to a sharp decrease in the true reward. Such phase transitions pose challenges to monitoring the safety of ML systems. To encourage further research on reward misspecification, address this, we propose an anomaly detection task for aberrant policies and offer several baseline detectors.