ICLR 2023 Papers

Neural Processes (NPs) are popular methods in meta-learning that can estimate predictive uncertainty on target datapoints by conditioning on a context dataset. Previous state-of-the-art method Transformer Neural Processes (TNPs) achieve strong performance but require quadratic computation with respect to the number of context datapoints, significantly limiting its scalability. Conversely, existing sub-quadratic NP variants perform significantly worse than that of TNPs. Tackling this issue, we propose Latent Bottlenecked Attentive Neural Processes (LBANPs), a new computationally efficient sub-quadratic NP variant, that has a querying computational complexity independent of the number of context datapoints. The model encodes the context dataset into a constant number of latent vectors on which self-attention is performed. When making predictions, the model retrieves higher-order information from the context dataset via multiple cross-attention mechanisms on the latent vectors. We empirically show that LBANPs achieve results competitive with the state-of-the-art on meta-regression, image completion, and contextual multi-armed bandits. We demonstrate that LBANPs can trade-off the computational cost and performance according to the number of latent vectors. Finally, we show LBANPs can scale beyond existing attention-based NP variants to larger dataset settings.

Intelligent agents need to remember salient information to reason in partially-observed environments. For example, agents with a first-person view should remember the positions of relevant objects even if they go out of view. Similarly, to effectively navigate through rooms agents need to remember the floor plan of how rooms are connected. However, most benchmark tasks in reinforcement learning do not test long-term memory in agents, slowing down progress in this important research direction. In this paper, we introduce the Memory Maze, a 3D domain of randomized mazes specifically designed for evaluating long-term memory in agents. Unlike existing benchmarks, Memory Maze measures long-term memory separate from confounding agent abilities and requires the agent to localize itself by integrating information over time. With Memory Maze, we propose an online reinforcement learning benchmark, a diverse offline dataset, and an offline probing evaluation. Recording a human player establishes a strong baseline and verifies the need to build up and retain memories, which is reflected in their gradually increasing rewards within each episode. We find that current algorithms benefit from training with truncated backpropagation through time and succeed on small mazes, but fall short of human performance on the large mazes, leaving room for future algorithmic designs to be evaluated on the Memory Maze.

Beyond the success story of adversarial training (AT) in the recent text domain on top of pre-trained language models (PLMs), our empirical study showcases the inconsistent gains from AT on some tasks, e.g. commonsense reasoning, named entity recognition. This paper investigates AT from the perspective of the contextualized language representation outputted by PLM encoders. We find the current AT attacks lean to generate sub-optimal adversarial examples that can fool the decoder part but have a minor effect on the encoder. However, we find it necessary to effectively deviate the latter one to allow AT to gain. Based on the observation, we propose simple yet effective \textit{Contextualized representation-Adversarial Training} (CreAT), in which the attack is explicitly optimized to deviate the contextualized representation of the encoder. It allows a global optimization of adversarial examples that can fool the entire model. We also find CreAT gives rise to a better direction to optimize the adversarial examples, to let them less sensitive to hyperparameters. Compared to AT, CreAT produces consistent performance gains on a wider range of tasks and is proven to be more effective for language pre-training where only the encoder part is kept for downstream tasks. We achieve the new state-of-the-art performances on a series of challenging benchmarks, e.g. AdvGLUE (59.1 $ \rightarrow $ 61.1), HellaSWAG (93.0 $ \rightarrow $ 94.9), ANLI (68.1 $ \rightarrow $ 69.3).

Few-shot classification is the ability to adapt to any new classification task from only a few training examples. The performance of current top-performing few-shot classifiers varies widely across different tasks where they often fail on a subset of `difficult' tasks. This phenomenon has real-world consequences for deployed few-shot systems where safety and reliability are paramount, yet little has been done to understand these failure cases. In this paper, we study these difficult tasks to gain a more nuanced understanding of the limitations of current methods. To this end, we develop a general and computationally efficient algorithm called FastDiffSel to extract difficult tasks from any large-scale vision dataset. Notably, our algorithm can extract tasks at least 20x faster than existing methods enabling its use on large-scale datasets. We use FastDiffSel to extract difficult tasks from Meta-Datasset, a widely-used few-shot classification benchmark, and other challenging large-scale vision datasets including ORBIT, CURE-OR and ObjectNet. These tasks are curated into Hard-MD++, a new few-shot testing benchmark to promote the development of methods that are robust to even the most difficult tasks. We use Hard-MD++ to stress-test an extensive suite of few-shot classification methods and show that state-of-the-art approaches fail catastrophically on difficult tasks. We believe that our extraction algorithm FastDiffSel and Hard-MD++ will aid researchers in further understanding failure modes of few-shot classification models.

Advances in the expressivity of pretrained models have increased interest in the design of adaptation protocols which enable safe and effective transfer learning. Going beyond conventional linear probing (LP) and fine tuning (FT) strategies, protocols that can effectively control feature distortion, i.e., the failure to update features orthogonal to the in-distribution, have been found to achieve improved out-of-distribution generalization (OOD). In order to limit this distortion, the LP+FT protocol, which first learns a linear probe and then uses this initialization for subsequent FT, was proposed. However, in this paper, we find when adaptation protocols (LP, FT, LP+FT) are also evaluated on a variety of safety objectives (e.g., calibration, robustness, etc.), a complementary perspective to feature distortion is helpful to explain protocol behavior. To this end, we study the susceptibility of protocols to simplicity bias (SB), i.e. the well-known propensity of deep neural networks to rely upon simple features, as SB has recently been shown to underlie several problems in robust generalization. Using a synthetic dataset, we demonstrate the susceptibility of existing protocols to SB. Given the strong effectiveness of LP+FT, we then propose modified linear probes that help mitigate SB, and lead to better initializations for subsequent FT. We verify the effectiveness of the proposed LP+FT variants for decreasing SB in a controlled setting, and their ability to improve OOD generalization and safety on three adaptation datasets.

Most graph neural networks follow the message passing mechanism. However, it faces the over-smoothing problem when multiple times of message passing is applied to a graph, causing indistinguishable node representations and prevents the model to effectively learn dependencies between farther-away nodes. On the other hand, features of neighboring nodes with different labels are likely to be falsely mixed, resulting in the heterophily problem. In this work, we propose to order the messages passing into the node representation, with specific blocks of neurons targeted for message passing within specific hops. This is achieved by aligning the hierarchy of the rooted-tree of a central node with the ordered neurons in its node representation. Experimental results on an extensive set of datasets show that our model can simultaneously achieve the state-of-the-art in both homophily and heterophily settings, without any targeted design. Moreover, its performance maintains pretty well while the model becomes really deep, effectively preventing the over-smoothing problem. Finally, visualizing the gating vectors shows that our model learns to behave differently between homophily and heterophily settings, providing an explainable graph neural model.

Despite their remarkable achievement, gigantic transformers encounter significant drawbacks, including exorbitant computational and memory footprints during training, as well as severe collapse evidenced by a high degree of parameter redundancy. Sparsely-activated Mixture-of-Experts (SMoEs) have shown promise to mitigate the issue of training efficiency, yet they are prone to (1) $\textit{redundant experts}$ due to representational collapse; and (2) $\textit{poor expert scalability for inference and downstream fine-tuning}$, primarily due to overfitting of the learned routing policy to the number of activated experts during training. As recent research efforts are predominantly focused on improving routing policies to encourage expert specializations, this work focuses on $\textit{exploring the overlooked scalability bottleneck of SMoEs}$ and leveraging it to effectively $\textbf{scale dense transformers}$. To this end, we propose a new plug-and-play training framework, $\textbf{SMoE-Dropout}$, to enable scaling transformers to better accuracy in their full capacity without collapse. Specifically, SMoE-Dropout consists of a $\textit{randomly initialized and fixed}$ router network to activate experts and gradually increases the activated expert number as training progresses over time. Transformers trained by SMoE-Dropout naturally exhibit a $\textbf{``self-slimmable”}$ property subject to resource availability, offering smooth and consistent performance boosts with an increase in activated experts during inference or fine-tuning. Our extensive experiments across diverse transformer architectures on a variety of tasks demonstrate the superior performance and substantial computation savings of SMoE-Dropout, compared to dense training baselines with equivalent parameter counts. In particular, our trained BERT outperforms its densely trained counterpart with consistent improvements of {$1.03\%$, $0.78\%$, $1.09\%$} on challenging reasoning tasks {$\texttt{ASDiv-A}$, $\texttt{MAWPS}$, $\texttt{SVAMP}$}, respectively. Codes and models are available in https://github.com/VITA-Group/Random-MoE-as-Dropout.

As large language models (LLMs) grow larger and more sophisticated, assessing their "reasoning" capabilities in natural language grows more challenging. Recent question answering (QA) benchmarks that attempt to assess reasoning are often limited by a narrow scope of covered situations and subject matters. We introduce WikiWhy, a QA dataset built around a novel auxiliary task: explaining why an answer is true in natural language. WikiWhy contains over 9,000 "why" question-answer-rationale triples, grounded on Wikipedia facts across a diverse set of topics. Each rationale is a set of supporting statements connecting the question to the answer. WikiWhy serves as a benchmark for the reasoning capabilities of LLMs because it demands rigorous explicit rationales for each answer to demonstrate the acquisition of implicit commonsense knowledge, which is unlikely to be easily memorized. GPT-3 baselines achieve only 38.7% human-evaluated correctness in the end-to-end answer & explain condition, leaving significant room for future improvements.

Many important problems involving molecular property prediction from 3D structures have limited data, posing a generalization challenge for neural networks. In this paper, we describe a pre-training technique based on denoising that achieves a new state-of-the-art in molecular property prediction by utilizing large datasets of 3D molecular structures at equilibrium to learn meaningful representations for downstream tasks. Relying on the well-known link between denoising autoencoders and score-matching, we show that the denoising objective corresponds to learning a molecular force field -- arising from approximating the Boltzmann distribution with a mixture of Gaussians -- directly from equilibrium structures. Our experiments demonstrate that using this pre-training objective significantly improves performance on multiple benchmarks, achieving a new state-of-the-art on the majority of targets in the widely used QM9 dataset. Our analysis then provides practical insights into the effects of different factors -- dataset sizes, model size and architecture, and the choice of upstream and downstream datasets -- on pre-training.

In the scenario-based Question Answering (QA) task, models are asked to find answers that are appropriate to the user scenarios associated with the question and identify information that is missing from the scenarios but is necessary for the answers to hold. Scenarios commonly include multiple properties of users, such as age, employment status, and income level for the question “How much can I claim from this benefit”. The properties relevant to a potential answer are given in a document, which will state conditions necessary for the answer to hold. Documents also may specify how conditions interact with each other, e.g. with text like “one of the conditions below must apply”. Although understanding the relationship between conditions is crucial for solving this challenging QA task, limited work has been done so far in modeling this. In this paper, we propose the T-Reasoner model, which solves this problem with three jointly learned modules: an entailment module which checks whether a condition has been satisfied by the scenario, a decoding module which locates eligible answers from documents, and a reasoning module which infers the relationship between conditions and performs a reasoning step to determine the logically consistent answers and identify missing conditions. T-Reasoner outperforms strong baselines on a synthetic scenario-based QA dataset and achieves a new state-of-the-art on two scenario-based QA benchmarks, outperforming the prior best models by 3-10 points.

Node classification is a classical graph representation learning task on which Graph Neural Networks (GNNs) have recently achieved strong results. However, it is often believed that standard GNNs only work well for homophilous graphs, i.e., graphs where edges tend to connect nodes of the same class. Graphs without this property are called heterophilous, and it is typically assumed that specialized methods are required to achieve strong performance on such graphs. In this work, we challenge this assumption. First, we show that the standard datasets used for evaluating heterophily-specific models have serious drawbacks, making results obtained by using them unreliable. The most significant of these drawbacks is the presence of a large number of duplicate nodes in the datasets Squirrel and Chameleon, which leads to train-test data leakage. We show that removing duplicate nodes strongly affects GNN performance on these datasets. Then, we propose a set of heterophilous graphs of varying properties that we believe can serve as a better benchmark for evaluating the performance of GNNs under heterophily. We show that standard GNNs achieve strong results on these heterophilous graphs, almost always outperforming specialized models. Our datasets and the code for reproducing our experiments are available at https://github.com/yandex-research/heterophilous-graphs

Hyperparameter optimization (HPO) and neural architecture search (NAS) are methods of choice to obtain the best-in-class machine learning models, but in practice they can be costly to run. When models are trained on large datasets, tuning them with HPO or NAS rapidly becomes prohibitively expensive for practitioners, even when efficient multi-fidelity methods are employed. We propose an approach to tackle the challenge of tuning machine learning models trained on large datasets with limited computational resources. Our approach, named PASHA, extends ASHA and is able to dynamically allocate maximum resources for the tuning procedure depending on the need. The experimental comparison shows that PASHA identifies well-performing hyperparameter configurations and architectures while consuming significantly fewer computational resources than ASHA.

We consider the sequential decision making problem of learning from an expert that has access to more information than the learner. For many problems this extra information will enable the expert to achieve greater long term reward than any policy without this privileged information access. We call these experts ``Impossibly Good'' because no learning algorithm will be able to reproduce their behavior. However, in these settings it is reasonable to attempt to recover the best policy possible given the agent's restricted access to information. We provide a set of necessary criteria on the expert that will allow a learner to recover the optimal policy in the reduced information space from the expert's advice alone. We also provide a new approach called Elf Distillation (Explorer Learning from Follower) that can be used in cases where these criteria are not met and environmental rewards must be taken into account. We show that this algorithm performs better than a variety of strong baselines on a challenging suite of Minigrid and Vizdoom environments.

Achieving structured, data-free sparsity of deep neural networks (DNNs) remains an open area of research. In this work, we address the challenge of pruning filters without access to the original training set or loss function. We propose the discriminative filters hypothesis, that well-trained models possess discriminative filters, and any non-discriminative filters can be pruned without impacting the predictive performance of the classifier. Based on this hypothesis, we propose a new paradigm for pruning neural networks: distributional pruning, wherein we only require access to the distributions that generated the original datasets. Our approach to solving the problem of formalising and quantifying the discriminating ability of filters is through the total variation (TV) distance between the class-conditional distributions of the filter outputs. We present empirical results that, using this definition of discriminability, support our hypothesis on a variety of datasets and architectures. Next, we define the LDIFF score, a heuristic to quantify the extent to which a layer possesses a mixture of discriminative and non-discriminative filters. We empirically demonstrate that the LDIFF score is indicative of the performance of random pruning for a given layer, and thereby indicates the extent to which a layer may be pruned. Our main contribution is a novel one-shot pruning algorithm, called TVSPrune, that identifies non-discriminative filters for pruning. We extend this algorithm to IterTVSPrune, wherein we iteratively apply TVSPrune, thereby enabling us to achieve greater sparsity. Last, we demonstrate the efficacy of the TVSPrune on a variety of datasets, and show that in some cases, we can prune up to 60% of parameters with only a 2% loss of accuracy without any fine-tuning of the model, beating the nearest baseline by almost 10%.

Real-world data usually couples the label ambiguity and heavy imbalance, challenging the algorithmic robustness of partial label learning (PLL) and long-tailed learning (LT). The straightforward combination of LT and PLL, i.e., LT-PLL, suffers from a fundamental dilemma: LT methods build upon a given class distribution that is unavailable in PLL, and the performance of PLL is severely influenced in long-tailed context. We show that even with the auxiliary of an oracle class prior, the state-of-the-art methods underperform due to an adverse fact that the constant rebalancing in LT is harsh to the label disambiguation in PLL. To overcome this challenge, we thus propose a dynamic rebalancing method, termed as RECORDS, without assuming any prior knowledge about the class distribution. Based on a parametric decomposition of the biased output, our method constructs a dynamic adjustment that is benign to the label disambiguation process and theoretically converges to the oracle class prior. Extensive experiments on three benchmark datasets demonstrate the significant gain of RECORDS compared with a range of baselines. The code is publicly available.

Heteroscedastic classifiers, which learn a multivariate Gaussian distribution over prediction logits, have been shown to perform well on image classification problems with hundreds to thousands of classes. However, compared to standard classifiers, they introduce extra parameters that scale linearly with the number of classes. This makes them infeasible to apply to larger-scale problems. In addition heteroscedastic classifiers introduce a critical temperature hyperparameter which must be tuned. We propose HET-XL, a heteroscedastic classifier whose parameter count when compared to a standard classifier scales independently of the number of classes. In our large-scale settings, we show that we can remove the need to tune the temperature hyperparameter, by directly learning it on the training data. On large image classification datasets with up to 4B images and 30k classes our method requires 14X fewer additional parameters, does not require tuning the temperature on a held-out set and performs consistently better than the baseline heteroscedastic classifier. HET-XL improves ImageNet 0-shot classification in a multimodal contrastive learning setup which can be viewed as a 3.5 billion class classification problem.

As machine learning models are increasingly being employed to make consequential decisions in real-world settings, it becomes critical to ensure that individuals who are adversely impacted (e.g., loan denied) by the predictions of these models are provided with a means for recourse. While several approaches have been proposed to construct recourses for affected individuals, the recourses output by these methods either achieve low costs (i.e., ease-of-implementation) or robustness to small perturbations (i.e., noisy implementations of recourses), but not both due to the inherent trade-offs between the recourse costs and robustness. Furthermore, prior approaches do not provide end users with any agency over navigating the aforementioned trade-offs. In this work, we address the above challenges by proposing the first algorithmic framework which enables users to effectively manage the recourse cost vs. robustness trade-offs. More specifically, our framework Probabilistically ROBust rEcourse (PROBE) lets users choose the probability with which a recourse could get invalidated (recourse invalidation rate) if small changes are made to the recourse i.e., the recourse is implemented somewhat noisily. To this end, we propose a novel objective function which simultaneously minimizes the gap between the achieved (resulting) and desired recourse invalidation rates, minimizes recourse costs, and also ensures that the resulting recourse achieves a positive model prediction. We develop novel theoretical results to characterize the recourse invalidation rates corresponding to any given instance w.r.t. different classes of underlying models (e.g., linear models, tree based models etc.), and leverage these results to efficiently optimize the proposed objective. Experimental evaluation with multiple real world datasets demonstrates the efficacy of the proposed framework.

Pre-training representations (a.k.a. foundation models) has recently become a prevalent learning paradigm, where one first pre-trains a representation using large-scale unlabeled data, and then learns simple predictors on top of the representation using small labeled data from the downstream tasks. There are two key desiderata for the representation: label efficiency (the ability to learn an accurate classifier on top of the representation with a small amount of labeled data) and universality (usefulness across a wide range of downstream tasks). In this paper, we focus on one of the most popular instantiations of this paradigm: contrastive learning with linear probing, i.e., learning a linear predictor on the representation pre-trained by contrastive learning. We show that there exists a trade-off between the two desiderata so that one may not be able to achieve both simultaneously. Specifically, we provide analysis using a theoretical data model and show that, while more diverse pre-training data result in more diverse features for different tasks (improving universality), it puts less emphasis on task-specific features, giving rise to larger sample complexity for down-stream supervised tasks, and thus worse prediction performance. Guided by this analysis, we propose a contrastive regularization method to improve the trade-off. We validate our analysis and method empirically with systematic experiments using real-world datasets and foundation models.

We consider the estimation of average and counterfactual treatment effects, under two settings: back-door adjustment and front-door adjustment. The goal in both cases is to recover the treatment effect without having an access to a hidden confounder. This objective is attained by first estimating the conditional mean of the desired outcome variable given relevant covariates (the ``first stage" regression), and then taking the (conditional) expectation of this function as a ``second stage" procedure. We propose to compute these conditional expectations directly using a regression function to the learned input features of the first stage, thus avoiding the need for sampling or density estimation. All functions and features (and in particular, the output features in the second stage) are neural networks learned adaptively from data, with the sole requirement that the final layer of the first stage should be linear. The proposed method is shown to converge to the true causal parameter, and outperforms the recent state-of-the-art methods on challenging causal benchmarks, including settings involving high-dimensional image data.