ICLR 2021 Papers

Autonomous agents need large repertoires of skills to act reasonably on new tasks that they have not seen before. However, acquiring these skills using only a stream of high-dimensional, unstructured, and unlabeled observations is a tricky challenge for any autonomous agent. Previous methods have used variational autoencoders to encode a scene into a low-dimensional vector that can be used as a goal for an agent to discover new skills. Nevertheless, in compositional/multi-object environments it is difficult to disentangle all the factors of variation into such a fixed-length representation of the whole scene. We propose to use object-centric representations as a modular and structured observation space, which is learned with a compositional generative world model. We show that the structure in the representations in combination with goal-conditioned attention policies helps the autonomous agent to discover and learn useful skills. These skills can be further combined to address compositional tasks like the manipulation of several different objects.

We re-evaluate the standard practice of sharing weights between input and output embeddings in state-of-the-art pre-trained language models. We show that decoupled embeddings provide increased modeling flexibility, allowing us to significantly improve the efficiency of parameter allocation in the input embedding of multilingual models. By reallocating the input embedding parameters in the Transformer layers, we achieve dramatically better performance on standard natural language understanding tasks with the same number of parameters during fine-tuning. We also show that allocating additional capacity to the output embedding provides benefits to the model that persist through the fine-tuning stage even though the output embedding is discarded after pre-training. Our analysis shows that larger output embeddings prevent the model's last layers from overspecializing to the pre-training task and encourage Transformer representations to be more general and more transferable to other tasks and languages. Harnessing these findings, we are able to train models that achieve strong performance on the XTREME benchmark without increasing the number of parameters at the fine-tuning stage.

We present a new methodology for detecting out-of-distribution (OOD) images by utilizing norms of the score estimates at multiple noise scales. A score is defined to be the gradient of the log density with respect to the input data. Our methodology is completely unsupervised and follows a straight forward training scheme. First, we train a deep network to estimate scores for $L$ levels of noise. Once trained, we calculate the noisy score estimates for $N$ in-distribution samples and take the L2-norms across the input dimensions (resulting in an $N$x$L$ matrix). Then we train an auxiliary model (such as a Gaussian Mixture Model) to learn the in-distribution spatial regions in this $L$-dimensional space. This auxiliary model can now be used to identify points that reside outside the learned space. Despite its simplicity, our experiments show that this methodology significantly outperforms the state-of-the-art in detecting out-of-distribution images. For example, our method can effectively separate CIFAR-10 (inlier) and SVHN (OOD) images, a setting which has been previously shown to be difficult for deep likelihood models.

Time-series learning is the bread and butter of data-driven *clinical decision support*, and the recent explosion in ML research has demonstrated great potential in various healthcare settings. At the same time, medical time-series problems in the wild are challenging due to their highly *composite* nature: They entail design choices and interactions among components that preprocess data, impute missing values, select features, issue predictions, estimate uncertainty, and interpret models. Despite exponential growth in electronic patient data, there is a remarkable gap between the potential and realized utilization of ML for clinical research and decision support. In particular, orchestrating a real-world project lifecycle poses challenges in engineering (i.e. hard to build), evaluation (i.e. hard to assess), and efficiency (i.e. hard to optimize). Designed to address these issues simultaneously, Clairvoyance proposes a unified, end-to-end, autoML-friendly pipeline that serves as a (i) software toolkit, (ii) empirical standard, and (iii) interface for optimization. Our ultimate goal lies in facilitating transparent and reproducible experimentation with complex inference workflows, providing integrated pathways for (1) personalized prediction, (2) treatment-effect estimation, and (3) information acquisition. Through illustrative examples on real-world data in outpatient, general wards, and intensive-care settings, we illustrate the applicability of the pipeline paradigm on core tasks in the healthcare journey. To the best of our knowledge, Clairvoyance is the first to demonstrate viability of a comprehensive and automatable pipeline for clinical time-series ML.

The properties of flat minima in the empirical risk landscape of neural networks have been debated for some time. Increasing evidence suggests they possess better generalization capabilities with respect to sharp ones. In this work we first discuss the relationship between alternative measures of flatness: The local entropy, which is useful for analysis and algorithm development, and the local energy, which is easier to compute and was shown empirically in extensive tests on state-of-the-art networks to be the best predictor of generalization capabilities. We show semi-analytically in simple controlled scenarios that these two measures correlate strongly with each other and with generalization. Then, we extend the analysis to the deep learning scenario by extensive numerical validations. We study two algorithms, Entropy-SGD and Replicated-SGD, that explicitly include the local entropy in the optimization objective. We devise a training schedule by which we consistently find flatter minima (using both flatness measures), and improve the generalization error for common architectures (e.g. ResNet, EfficientNet).

Disparate impact has raised serious concerns in machine learning applications and its societal impacts. In response to the need of mitigating discrimination, fairness has been regarded as a crucial property in algorithmic design. In this work, we study the problem of disparate impact on graph-structured data. Specifically, we focus on dyadic fairness, which articulates a fairness concept that a predictive relationship between two instances should be independent of the sensitive attributes. Based on this, we theoretically relate the graph connections to dyadic fairness on link predictive scores in learning graph neural networks, and reveal that regulating weights on existing edges in a graph contributes to dyadic fairness conditionally. Subsequently, we propose our algorithm, \textbf{FairAdj}, to empirically learn a fair adjacency matrix with proper graph structural constraints for fair link prediction, and in the meanwhile preserve predictive accuracy as much as possible. Empirical validation demonstrates that our method delivers effective dyadic fairness in terms of various statistics, and at the same time enjoys a favorable fairness-utility tradeoff.

The mushroom body of the fruit fly brain is one of the best studied systems in neuroscience. At its core it consists of a population of Kenyon cells, which receive inputs from multiple sensory modalities. These cells are inhibited by the anterior paired lateral neuron, thus creating a sparse high dimensional representation of the inputs. In this work we study a mathematical formalization of this network motif and apply it to learning the correlational structure between words and their context in a corpus of unstructured text, a common natural language processing (NLP) task. We show that this network can learn semantic representations of words and can generate both static and context-dependent word embeddings. Unlike conventional methods (e.g., BERT, GloVe) that use dense representations for word embedding, our algorithm encodes semantic meaning of words and their context in the form of sparse binary hash codes. The quality of the learned representations is evaluated on word similarity analysis, word-sense disambiguation, and document classification. It is shown that not only can the fruit fly network motif achieve performance comparable to existing methods in NLP, but, additionally, it uses only a fraction of the computational resources (shorter training time and smaller memory footprint).

The geometric properties of contextual embedding spaces for deep language models such as BERT and ERNIE, have attracted considerable attention in recent years. Investigations on the contextual embeddings demonstrate a strong anisotropic space such that most of the vectors fall within a narrow cone, leading to high cosine similarities. It is surprising that these LMs are as successful as they are, given that most of their embedding vectors are as similar to one another as they are. In this paper, we argue that the isotropy indeed exists in the space, from a different but more constructive perspective. We identify isolated clusters and low dimensional manifolds in the contextual embedding space, and introduce tools to both qualitatively and quantitatively analyze them. We hope the study in this paper could provide insights towards a better understanding of the deep language models.

We present lambda layers -- an alternative framework to self-attention -- for capturing long-range interactions between an input and structured contextual information (e.g. a pixel surrounded by other pixels). Lambda layers capture such interactions by transforming available contexts into linear functions, termed lambdas, and applying these linear functions to each input separately. Similar to linear attention, lambda layers bypass expensive attention maps, but in contrast, they model both content and position-based interactions which enables their application to large structured inputs such as images. The resulting neural network architectures, LambdaNetworks, significantly outperform their convolutional and attentional counterparts on ImageNet classification, COCO object detection and instance segmentation, while being more computationally efficient. Additionally, we design LambdaResNets, a family of hybrid architectures across different scales, that considerably improves the speed-accuracy tradeoff of image classification models. LambdaResNets reach excellent accuracies on ImageNet while being 3.2 - 4.4x faster than the popular EfficientNets on modern machine learning accelerators. In large-scale semi-supervised training with an additional 130M pseudo-labeled images, LambdaResNets achieve up to 86.7% ImageNet accuracy while being 9.5x faster than EfficientNet NoisyStudent and 9x faster than a Vision Transformer with comparable accuracies.

Influence functions approximate the effect of training samples in test-time predictions and have a wide variety of applications in machine learning interpretability and uncertainty estimation. A commonly-used (first-order) influence function can be implemented efficiently as a post-hoc method requiring access only to the gradients and Hessian of the model. For linear models, influence functions are well-defined due to the convexity of the underlying loss function and are generally accurate even across difficult settings where model changes are fairly large such as estimating group influences. Influence functions, however, are not well-understood in the context of deep learning with non-convex loss functions. In this paper, we provide a comprehensive and large-scale empirical study of successes and failures of influence functions in neural network models trained on datasets such as Iris, MNIST, CIFAR-10 and ImageNet. Through our extensive experiments, we show that the network architecture, its depth and width, as well as the extent of model parameterization and regularization techniques have strong effects in the accuracy of influence functions. In particular, we find that (i) influence estimates are fairly accurate for shallow networks, while for deeper networks the estimates are often erroneous; (ii) for certain network architectures and datasets, training with weight-decay regularization is important to get high-quality influence estimates; and (iii) the accuracy of influence estimates can vary significantly depending on the examined test points. These results suggest that in general influence functions in deep learning are fragile and call for developing improved influence estimation methods to mitigate these issues in non-convex setups.

Dancing to music is one of human's innate abilities since ancient times. In machine learning research, however, synthesizing dance movements from music is a challenging problem. Recently, researchers synthesize human motion sequences through autoregressive models like recurrent neural network (RNN). Such an approach often generates short sequences due to an accumulation of prediction errors that are fed back into the neural network. This problem becomes even more severe in the long motion sequence generation. Besides, the consistency between dance and music in terms of style, rhythm and beat is yet to be taken into account during modeling. In this paper, we formalize the music-driven dance generation as a sequence-to-sequence learning problem and devise a novel seq2seq architecture to efficiently process long sequences of music features and capture the fine-grained correspondence between music and dance. Furthermore, we propose a novel curriculum learning strategy to alleviate error accumulation of autoregressive models in long motion sequence generation, which gently changes the training process from a fully guided teacher-forcing scheme using the previous ground-truth movements, towards a less guided autoregressive scheme mostly using the generated movements instead. Extensive experiments show that our approach significantly outperforms the existing state-of-the-arts on automatic metrics and human evaluation. We also make a demo video to demonstrate the superior performance of our proposed approach at https://www.youtube.com/watch?v=lmE20MEheZ8.

Structured pruning methods are among the effective strategies for extracting small resource-efficient convolutional neural networks from their dense counterparts with minimal loss in accuracy. However, most existing methods still suffer from one or more limitations, that include 1) the need for training the dense model from scratch with pruning-related parameters embedded in the architecture, 2) requiring model-specific hyperparameter settings, 3) inability to include budget-related constraint in the training process, and 4) instability under scenarios of extreme pruning. In this paper, we present ChipNet, a deterministic pruning strategy that employs continuous Heaviside function and a novel crispness loss to identify a highly sparse network out of an existing dense network. Our choice of continuous Heaviside function is inspired by the field of design optimization, where the material distribution task is posed as a continuous optimization problem, but only discrete values (0 or 1) are practically feasible and expected as final outcomes. Our approach's flexible design facilitates its use with different choices of budget constraints while maintaining stability for very low target budgets. Experimental results show that ChipNet outperforms state-of-the-art structured pruning methods by remarkable margins of up to 16.1% in terms of accuracy. Further, we show that the masks obtained with ChipNet are transferable across datasets. For certain cases, it was observed that masks transferred from a model trained on feature-rich teacher dataset provide better performance on the student dataset than those obtained by directly pruning on the student data itself.

Simulated virtual environments serve as one of the main driving forces behind developing and evaluating skill learning algorithms. However, existing environments typically only simulate rigid body physics. Additionally, the simulation process usually does not provide gradients that might be useful for planning and control optimizations. We introduce a new differentiable physics benchmark called PasticineLab, which includes a diverse collection of soft body manipulation tasks. In each task, the agent uses manipulators to deform the plasticine into a desired configuration. The underlying physics engine supports differentiable elastic and plastic deformation using the DiffTaichi system, posing many under-explored challenges to robotic agents. We evaluate several existing reinforcement learning (RL) methods and gradient-based methods on this benchmark. Experimental results suggest that 1) RL-based approaches struggle to solve most of the tasks efficiently; 2) gradient-based approaches, by optimizing open-loop control sequences with the built-in differentiable physics engine, can rapidly find a solution within tens of iterations, but still fall short on multi-stage tasks that require long-term planning. We expect that PlasticineLab will encourage the development of novel algorithms that combine differentiable physics and RL for more complex physics-based skill learning tasks. PlasticineLab will be made publicly available.

In goal-conditioned Hierarchical Reinforcement Learning (HRL), a high-level policy periodically sets subgoals for a low-level policy, and the low-level policy is trained to reach those subgoals. A proper subgoal representation function, which abstracts a state space to a latent subgoal space, is crucial for effective goal-conditioned HRL, since different low-level behaviors are induced by reaching subgoals in the compressed representation space. Observing that the high-level agent operates at an abstract temporal scale, we propose a slowness objective to effectively learn the subgoal representation (i.e., the high-level action space). We provide a theoretical grounding for the slowness objective. That is, selecting slow features as the subgoal space can achieve efficient hierarchical exploration. As a result of better exploration ability, our approach significantly outperforms state-of-the-art HRL and exploration methods on a number of benchmark continuous-control tasks. Thanks to the generality of the proposed subgoal representation learning method, empirical results also demonstrate that the learned representation and corresponding low-level policies can be transferred between distinct tasks.

Recent work has shown deep learning can accelerate the prediction of physical dynamics relative to numerical solvers. However, limited physical accuracy and an inability to generalize under distributional shift limit its applicability to the real world. We propose to improve accuracy and generalization by incorporating symmetries into convolutional neural networks. Specifically, we employ a variety of methods each tailored to enforce a different symmetry. Our models are both theoretically and experimentally robust to distributional shift by symmetry group transformations and enjoy favorable sample complexity. We demonstrate the advantage of our approach on a variety of physical dynamics including Rayleigh–Bénard convection and real-world ocean currents and temperatures. Compare with image or text applications, our work is a significant step towards applying equivariant neural networks to high-dimensional systems with complex dynamics.

Recent work has shown that large text-based neural language models acquire a surprising propensity for one-shot learning. Here, we show that an agent situated in a simulated 3D world, and endowed with a novel dual-coding external memory, can exhibit similar one-shot word learning when trained with conventional RL algorithms. After a single introduction to a novel object via visual perception and language ("This is a dax"), the agent can manipulate the object as instructed ("Put the dax on the bed"), combining short-term, within-episode knowledge of the nonsense word with long-term lexical and motor knowledge. We find that, under certain training conditions and with a particular memory writing mechanism, the agent's one-shot word-object binding generalizes to novel exemplars within the same ShapeNet category, and is effective in settings with unfamiliar numbers of objects. We further show how dual-coding memory can be exploited as a signal for intrinsic motivation, stimulating the agent to seek names for objects that may be useful later. Together, the results demonstrate that deep neural networks can exploit meta-learning, episodic memory and an explicitly multi-modal environment to account for 'fast-mapping', a fundamental pillar of human cognitive development and a potentially transformative capacity for artificial agents.

Sequential deep learning models such as RNN, causal CNN and attention mechanism do not readily consume continuous-time information. Discretizing the temporal data, as we show, causes inconsistency even for simple continuous-time processes. Current approaches often handle time in a heuristic manner to be consistent with the existing deep learning architectures and implementations. In this paper, we provide a principled way to characterize continuous-time systems using deep learning tools. Notably, the proposed approach applies to all the major deep learning architectures and requires little modifications to the implementation. The critical insight is to represent the continuous-time system by composing neural networks with a temporal kernel, where we gain our intuition from the recent advancements in understanding deep learning with Gaussian process and neural tangent kernel. To represent the temporal kernel, we introduce the random feature approach and convert the kernel learning problem to spectral density estimation under reparameterization. We further prove the convergence and consistency results even when the temporal kernel is non-stationary, and the spectral density is misspecified. The simulations and real-data experiments demonstrate the empirical effectiveness of our temporal kernel approach in a broad range of settings.

We develop an approach to growing deep network architectures over the course of training, driven by a principled combination of accuracy and sparsity objectives. Unlike existing pruning or architecture search techniques that operate on full-sized models or supernet architectures, our method can start from a small, simple seed architecture and dynamically grow and prune both layers and filters. By combining a continuous relaxation of discrete network structure optimization with a scheme for sampling sparse subnetworks, we produce compact, pruned networks, while also drastically reducing the computational expense of training. For example, we achieve $49.7\%$ inference FLOPs and $47.4\%$ training FLOPs savings compared to a baseline ResNet-50 on ImageNet, while maintaining $75.2\%$ top-1 validation accuracy --- all without any dedicated fine-tuning stage. Experiments across CIFAR, ImageNet, PASCAL VOC, and Penn Treebank, with convolutional networks for image classification and semantic segmentation, and recurrent networks for language modeling, demonstrate that we both train faster and produce more efficient networks than competing architecture pruning or search methods.

In this paper, we introduce Watch-And-Help (WAH), a challenge for testing social intelligence in agents. In WAH, an AI agent needs to help a human-like agent perform a complex household task efficiently. To succeed, the AI agent needs to i) understand the underlying goal of the task by watching a single demonstration of the human-like agent performing the same task (social perception), and ii) coordinate with the human-like agent to solve the task in an unseen environment as fast as possible (human-AI collaboration). For this challenge, we build VirtualHome-Social, a multi-agent household environment, and provide a benchmark including both planning and learning based baselines. We evaluate the performance of AI agents with the human-like agent as well as and with real humans using objective metrics and subjective user ratings. Experimental results demonstrate that our challenge and virtual environment enable a systematic evaluation on the important aspects of machine social intelligence at scale.