ICLR 2022 论文列表

We introduce HTLM, a hyper-text language model trained on a large-scale web crawl. Modeling hyper-text has a number of advantages: (1) it is easily gathered at scale, (2) it provides rich document-level and end-task-adjacent supervision (e.g. 'class' and 'id' attributes often encode document category information), and (3) it allows for new structured prompting that follows the established semantics of HTML (e.g. to do zero-shot summarization by infilling '<title>' tags for a webpage that contains the input text). We show that pretraining with a BART-style denoising loss directly on simplified HTML provides highly effective transfer for a wide range of end tasks and supervision levels. HTLM matches or exceeds the performance of comparably sized text-only LMs for zero-shot prompting and fine-tuning for classification benchmarks, while also setting new state-of-the-art performance levels for zero-shot summarization. We also find that hyper-text prompts provide more value to HTLM, in terms of data efficiency, than plain text prompts do for existing LMs, and that HTLM is highly effective at auto-prompting itself, by simply generating the most likely hyper-text formatting for any available training data. We will release all code and models to support future HTLM research.

In this work, we evaluate the effectiveness of representation learning approaches for decision making in visually complex environments. Representation learning is essential for effective reinforcement learning (RL) from high-dimensional in- puts. Unsupervised representation learning approaches based on reconstruction, prediction or contrastive learning have shown substantial learning efficiency gains. Yet, they have mostly been evaluated in clean laboratory or simulated settings. In contrast, real environments are visually complex and contain substantial amounts of clutter and distractors. Unsupervised representations will learn to model such distractors, potentially impairing the agent’s learning efficiency. In contrast, an alternative class of approaches, which we call task-induced representation learning, leverages task information such as rewards or demonstrations from prior tasks to focus on task-relevant parts of the scene and ignore distractors. We investi- gate the effectiveness of unsupervised and task-induced representation learning approaches on four visually complex environments, from Distracting DMControl to the CARLA driving simulator. For both, RL and imitation learning, we find that representation learning generally improves sample efficiency on unseen tasks even in visually complex scenes and that task-induced representations can double learning efficiency compared to unsupervised alternatives.

Background. Commonly, Deep Neural Networks (DNNs) generalize well on samples drawn from a distribution similar to that of the training set. However, DNNs' predictions are brittle and unreliable when the test samples are drawn from a dissimilar distribution. This is a major concern for deployment in real-world applications, where such behavior may come at a considerable cost, such as industrial production lines, autonomous vehicles, or healthcare applications. Contributions. We frame Out Of Distribution (OOD) detection in DNNs as a statistical hypothesis testing problem. Tests generated within our proposed framework combine evidence from the entire network. Unlike previous OOD detection heuristics, this framework returns a $p$-value for each test sample. It is guaranteed to maintain the Type I Error (T1E - incorrectly predicting OOD for an actual in-distribution sample) for test data. Moreover, this allows to combine several detectors while maintaining the T1E. Building on this framework, we suggest a novel OOD procedure based on low-order statistics. Our method achieves comparable or better results than state-of-the-art methods on well-accepted OOD benchmarks, without retraining the network parameters or assuming prior knowledge on the test distribution --- and at a fraction of the computational cost.

This paper proposes a novel method for deep learning based on the analytical convolution of multidimensional Gaussian mixtures. In contrast to tensors, these do not suffer from the curse of dimensionality and allow for a compact representation, as data is only stored where details exist. Convolution kernels and data are Gaussian mixtures with unconstrained weights, positions, and covariance matrices. Similar to discrete convolutional networks, each convolution step produces several feature channels, represented by independent Gaussian mixtures. Since traditional transfer functions like ReLUs do not produce Gaussian mixtures, we propose using a fitting of these functions instead. This fitting step also acts as a pooling layer if the number of Gaussian components is reduced appropriately. We demonstrate that networks based on this architecture reach competitive accuracy on Gaussian mixtures fitted to the MNIST and ModelNet data sets.

Parametric computer-aided design (CAD) tools are the predominant way that engineers specify physical structures, from bicycle pedals to airplanes to printed circuit boards. The key characteristic of parametric CAD is that design intent is encoded not only via geometric primitives, but also by parameterized constraints between the elements. This relational specification can be viewed as the construction of a constraint program, allowing edits to coherently propagate to other parts of the design. Machine learning offers the intriguing possibility of accelerating the design process via generative modeling of these structures, enabling new tools such as autocompletion, constraint inference, and conditional synthesis. In this work, we present such an approach to generative modeling of parametric CAD sketches, which constitute the basic computational building blocks of modern mechanical design. Our model, trained on real-world designs from the SketchGraphs dataset, autoregressively synthesizes sketches as sequences of primitives, with initial coordinates, and constraints that reference back to the sampled primitives. As samples from the model match the constraint graph representation used in standard CAD software, they may be directly imported, solved, and edited according to downstream design tasks. In addition, we condition the model on various contexts, including partial sketches (primers) and images of hand-drawn sketches. Evaluation of the proposed approach demonstrates its ability to synthesize realistic CAD sketches and its potential to aid the mechanical design workflow.

Graph neural networks (GNNs) are widely used in regression and classification problems applied to text, in areas such as sentiment analysis and medical decision-making processes. We propose a novel form for node attributes within a GNN based model that captures node-specific embeddings for every word in the vocabulary. This provides a global representation at each node, coupled with node-level updates according to associations among words in a transcript. We demonstrate the efficacy of the approach by augmenting the accuracy of measuring major depressive disorder (MDD). Prior research has sought to make a diagnostic prediction of depression levels from patient data using several modalities, including audio, video, and text. On the DAIC-WOZ benchmark, our method outperforms state-of-art methods by a substantial margin, including those using multiple modalities. Moreover, we also evaluate the performance of our novel model on a Twitter sentiment dataset. We show that our model outperforms a general GNN model by leveraging our novel 2-D node attributes. These results demonstrate the generality of the proposed method.

Online reinforcement learning (RL) can suffer from poor exploration, sparse reward, insufficient data, and overhead caused by inefficient interactions between an immature policy and a complicated environment. Model-based offline RL instead trains an environment model using a dataset of pre-collected experiences so online RL methods can learn in an offline manner by solely interacting with the model. However, the uncertainty and accuracy of the environment model can drastically vary across different state-action pairs so the RL agent may achieve high model return but perform poorly in the true environment. Unlike previous works that need to carefully tune the trade-off between the model return and uncertainty in a single objective, we study a bi-objective formulation for model-based offline RL that aims at producing a pool of diverse policies on the Pareto front performing different levels of trade-offs, which provides the flexibility to select the best policy for each realistic environment from the pool. Our method, ''Pareto policy pool (P3)'', does not need to tune the trade-off weight but can produce policies allocated at different regions of the Pareto front. For this purpose, we develop an efficient algorithm that solves multiple bi-objective optimization problems with distinct constraints defined by reference vectors targeting diverse regions of the Pareto front. We theoretically prove that our algorithm can converge to the targeted regions. In order to obtain more Pareto optimal policies without linearly increasing the cost, we leverage the achieved policies as initialization to find more Pareto optimal policies in their neighborhoods. On the D4RL benchmark for offline RL, P3 substantially outperforms several recent baseline methods over multiple tasks, especially when the quality of pre-collected experiences is low.

Multiple domains like vision, natural language, and audio are witnessing tremendous progress by leveraging Transformers for large scale pre-training followed by task specific fine tuning. In contrast, in robotics we primarily train a single robot for a single task. However, modular robot systems now allow for the flexible combination of general-purpose building blocks into task optimized morphologies. However, given the exponentially large number of possible robot morphologies, training a controller for each new design is impractical. In this work, we propose MetaMorph, a Transformer based approach to learn a universal controller over a modular robot design space. MetaMorph is based on the insight that robot morphology is just another modality on which we can condition the output of a Transformer. Through extensive experiments we demonstrate that large scale pre-training on a variety of robot morphologies results in policies with combinatorial generalization capabilities, including zero shot generalization to unseen robot morphologies. We further demonstrate that our pre-trained policy can be used for sample-efficient transfer to completely new robot morphologies and tasks.

The most significant barrier to the advancement of Neural Architecture Search (NAS) is its demand for large computational resources, which hinders scientifically sound empirical evaluations of NAS methods. Tabular NAS benchmarks have alleviated this problem substantially, making it possible to properly evaluate NAS methods in seconds on commodity machines. However, an unintended consequence of tabular NAS benchmarks has been a focus on extremely small architectural search spaces since their construction relies on exhaustive evaluations of the space. This leads to unrealistic results that do not transfer to larger spaces. To overcome this fundamental limitation, we propose a methodology to create cheap NAS surrogate benchmarks for arbitrary search spaces. We exemplify this approach by creating surrogate NAS benchmarks on the existing tabular NAS-Bench-101 and on two widely used NAS search spaces with up to $10^{21}$ architectures ($10^{13}$ times larger than any previous tabular NAS benchmark). We show that surrogate NAS benchmarks can model the true performance of architectures better than tabular benchmarks (at a small fraction of the cost), that they lead to faithful estimates of how well different NAS methods work on the original non-surrogate benchmark, and that they can generate new scientific insight. We open-source all our code and believe that surrogate NAS benchmarks are an indispensable tool to extend scientifically sound work on NAS to large and exciting search spaces.

Based on large-scale pre-trained multilingual representations, recent cross-lingual transfer methods have achieved impressive transfer performances. However, the performance of target languages still lags far behind the source language. In this paper, our analyses indicate such a performance gap is strongly associated with the cross-lingual representation discrepancy. To achieve better cross-lingual transfer performance, we propose the cross-lingual manifold mixup (X-Mixup) method, which adaptively calibrates the representation discrepancy and gives a compromised representation for target languages. Experiments on the XTREME benchmark show X-Mixup achieves 1.8% performance gains on multiple text understanding tasks, compared with strong baselines, and significantly reduces the cross-lingual representation discrepancy.

Humans use natural language to compose common concepts from their environment into plausible, day-to-day scene descriptions. However, such generative commonsense reasoning (GCSR) skills are lacking in state-of-the-art text generation methods. Descriptive sentences about arbitrary concepts generated by neural text generation models (e.g., pre-trained text-to-text Transformers) are often grammatically fluent but may not correspond to human common sense, largely due to their lack of mechanisms to capture concept relations, to identify implicit concepts, and to perform generalizable reasoning about unseen concept compositions. In this paper, we propose an Imagine-and-Verbalize (I\&V) method, which learns to imagine a relational scene knowledge graph (SKG) with relations between the input concepts, and leverage the SKG as a constraint when generating a plausible scene description. We collect and harmonize a set of knowledge resources from different domains and modalities, providing a rich auxiliary supervision signal for I\&V. The experiments demonstrate the effectiveness of I\&V in improving language models on both concept-to-sentence and concept-to-story generation tasks, while enabling the model to learn well from fewer task examples and generate SKGs that make common sense to human annotators.

In this work, we study the evolution of the loss Hessian across many classification tasks in order to understand the effect the curvature of the loss has on the training dynamics. Whereas prior work has focused on how different learning rates affect the loss Hessian observed during training, we also analyze the effects of model initialization, architectural choices, and common training heuristics such as gradient clipping and learning rate warmup. Our results demonstrate that successful model and hyperparameter choices allow the early optimization trajectory to either avoid---or navigate out of---regions of high curvature and into flatter regions that tolerate a higher learning rate. Our results suggest a unifying perspective on how disparate mitigation strategies for training instability ultimately address the same underlying failure mode of neural network optimization, namely poor conditioning. Inspired by the conditioning perspective, we show that learning rate warmup can improve training stability just as much as batch normalization, layer normalization, MetaInit, GradInit, and Fixup initialization.

Natural language understanding tasks such as open-domain question answering often require retrieving and assimilating factual information from multiple sources. We propose to address this problem by integrating a semi-parametric representation of a large text corpus into a Transformer model as a source of factual knowledge. Specifically, our method represents knowledge with ``mention memory'', a table of dense vector representations of every entity mention in a corpus. The proposed model - TOME - is a Transformer that accesses the information through internal memory layers in which each entity mention in the input passage attends to the mention memory. This approach enables synthesis of and reasoning over many disparate sources of information within a single Transformer model. In experiments using a memory of 150 million Wikipedia mentions, TOME achieves strong performance on several open-domain knowledge-intensive tasks, including the claim verification benchmarks HoVer and FEVER and several entity-based QA benchmarks. We also show that the model learns to attend to informative mentions without any direct supervision. Finally we demonstrate that the model can generalize to new unseen entities by updating the memory without retraining.

State-of-the-art deep Q-learning methods update Q-values using state transition tuples sampled from the experience replay buffer. This strategy often randomly samples or prioritizes data sampling based on measures such as the temporal difference (TD) error. Such sampling strategies can be inefficient at learning Q-function since a state's correct Q-value preconditions on the accurate successor states' Q-value. Disregarding such a successor's value dependency leads to useless updates and even learning wrong values. To expedite Q-learning, we maintain states' dependency by organizing the agent's experience into a graph. Each edge in the graph represents a transition between two connected states. We perform value backups via a breadth-first search that expands vertices in the graph starting from the set of terminal states successively moving backward. We empirically show that our method is substantially more data-efficient than several baselines on a diverse range of goal-reaching tasks. Notably, the proposed method also outperforms baselines that consume more batches of training experience.

Conveying complex objectives to reinforcement learning (RL) agents often requires meticulous reward engineering. Preference-based RL methods are able to learn a more flexible reward model based on human preferences by actively incorporating human feedback, i.e. teacher's preferences between two clips of behaviors. However, poor feedback-efficiency still remains as a problem in current preference-based RL algorithms, as tailored human feedback is very expensive. To handle this issue, previous methods have mainly focused on improving query selection and policy initialization. At the same time, recent exploration methods have proven to be a recipe for improving sample-efficiency in RL. We present an exploration method specifically for preference-based RL algorithms. Our main idea is to design an intrinsic reward by measuring the novelty based on learned reward. Specifically, we utilize disagreement across ensemble of learned reward models. Our intuition is that disagreement in learned reward model reflects uncertainty in tailored human feedback and could be useful for exploration. Our experiments show that reward uncertainty exploration improves both feedback- and sample-efficiency of preference-based RL algorithms on complex robot manipulation tasks from Meta-World benchmarks, compared with other existing exploration methods that measure the novelty of state visitation.

Definitions of the distance between two machine learning models either characterize the similarity of the models' predictions or of their weights. While similarity of weights is attractive because it implies similarity of predictions in the limit, it suffers from being inapplicable to comparing models with different architectures. On the other hand, the similarity of predictions is broadly applicable but depends heavily on the choice of model inputs during comparison. In this paper, we instead propose to compute distance between black-box models by comparing their Local Interpretable Model-Agnostic Explanations (LIME). To compare two models, we take a reference dataset, and locally approximate the models on each reference point with linear models trained by LIME. We then compute the cosine distance between the concatenated weights of the linear models. This yields an approach that is both architecture-independent and possesses the benefits of comparing models in weight space. We empirically show that our method, which we call Zest, can be applied to two problems that require measurements of model similarity: detecting model stealing and machine unlearning.

Building embodied intelligent agents that can interact with 3D indoor environments has received increasing research attention in recent years. While most works focus on single-object or agent-object visual functionality and affordances, our work proposes to study a novel, underexplored, kind of visual relations that is also important to perceive and model -- inter-object functional relationships (e.g., a switch on the wall turns on or off the light, a remote control operates the TV). Humans often spend no effort or only a little to infer these relationships, even when entering a new room, by using our strong prior knowledge (e.g., we know that buttons control electrical devices) or using only a few exploratory interactions in cases of uncertainty (e.g., multiple switches and lights in the same room). In this paper, we take the first step in building AI system learning inter-object functional relationships in 3D indoor environments with key technical contributions of modeling prior knowledge by training over large-scale scenes and designing interactive policies for effectively exploring the training scenes and quickly adapting to novel test scenes. We create a new dataset based on the AI2Thor and PartNet datasets and perform extensive experiments that prove the effectiveness of our proposed method.

Deep neural networks (DNNs) are proved to be vulnerable against backdoor attacks. A backdoor could be embedded in the target DNNs through injecting a backdoor trigger into the training examples, which can cause the target DNNs misclassify an input attached with the backdoor trigger. Recent backdoor detection methods often require the access to the original poisoned training data, the parameters of the target DNNs, or the predictive confidence for each given input, which are impractical in many real-world applications, e.g., on-device de-ployed DNNs. We address the black-box hard-label backdoor detection problem where the DNN is a fully black-box and only its final output label is accessible. We approach this problem from the optimization perspective and show that the objective of backdoor detection is bounded by an adversarial objective. Further theoretical and empirical studies reveal that this adversarial objective leads to a solution with highly skewed distribution; a singularity is often observed in the adversarial map of a backdoor-infected example, which we call the adversarial singularity phenomenon. Based on this observation, we propose the adversarial extreme value analysis(AEVA) algorithm to detect backdoors in black-box neural networks. The AEVA algorithm is based on an extreme value analysis on the adversarial map, computed from the monte-carlo gradient estimation due to the black-box hard-label constraint. Evidenced by extensive experiments across three popular tasks and backdoor attacks, our approach is shown effective in detecting backdoor attacks under the black-box hard-label scenarios

Despite the empirical success of the deep Q network (DQN) reinforcement learning algorithm and its variants, DQN is still not well understood and it does not guarantee convergence. In this work, we show that DQN can indeed diverge and cease to operate in realistic settings. Although there exist gradient-based convergent methods, we show that they actually have inherent problems in learning dynamics which cause them to fail even for simple tasks. To overcome these problems, we propose a convergent DQN algorithm (C-DQN) that is guaranteed to converge and can work with large discount factors (0.9998). It learns robustly in difficult settings and can learn several difficult games in the Atari 2600 benchmark that DQN fails to solve.