ICLR 2024 Papers

% Self-supervised learning methods have shown promising results across a wide range of tasks in computer vision, natural language processing, and multimodal analysis. However, self-supervised approaches come with a notable limitation, dimensional collapse, where a model doesn't fully utilize its capacity to encode information optimally. Motivated by this, we propose ProSMin, a novel probabilistic self-supervised learning approach that leverages the power of probabilistic models to enhance representation quality and mitigate collapsing representations. Our proposed approach involves two neural networks, the online network and the target network, which collaborate and learn the diverse distribution of representations from each other through probabilistic knowledge distillation. The two networks are trained via our new loss function based on proper scoring rules. We provide a theoretical justification for ProSMin and demonstrate its modified scoring rule. This insight validates the method's optimization process and contributes to its robustness and effectiveness in improving representation quality. We evaluate our probabilistic model on various downstream tasks, such as in-distribution generalization, out-of-distribution detection, dataset corruption, low-shot learning, and transfer learning. Our method achieves superior accuracy and calibration, outperforming the self-supervised baseline in a variety of experiments on large datasets such as ImageNet-O and ImageNet-C. ProSMin thus demonstrates its scalability and real-world applicability. Our code is publicly available: https://github.com/amirvhd/SSL-sore-rule.

Multiple choice questions (MCQs) serve as a common yet important task format in the evaluation of large language models (LLMs). This work shows that modern LLMs are vulnerable to option position changes in MCQs due to their inherent “selection bias”, namely, they prefer to select specific option IDs as answers (like “Option A”). Through extensive empirical analyses with 20 LLMs on three benchmarks, we pinpoint that this behavioral bias primarily stems from LLMs’ token bias, where the model a priori assigns more probabilistic mass to specific option ID tokens (e.g., A/B/C/D) when predicting answers from the option IDs. To mitigate selection bias, we propose a label-free, inference-time debiasing method, called PriDe, which separates the model’s prior bias for option IDs from the overall prediction distribution. PriDe first estimates the prior by permutating option contents on a small number of test samples, and then applies the estimated prior to debias the remaining samples. We demonstrate that it achieves interpretable and transferable debiasing with high computational efficiency. We hope this work can draw broader research attention to the bias and robustness of modern LLMs.

We introduce SocioDojo, an open-ended lifelong learning environment for developing ready-to-deploy autonomous agents capable of performing human-like analysis and decision-making on societal topics such as economics, finance, politics, and culture. It consists of (1) information sources from news, social media, reports, etc., (2) a knowledge base built from books, journals, and encyclopedias, plus a toolbox of Internet and knowledge graph search interfaces, (3) 30K high-quality time series in finance, economy, society, and polls, which support a novel task called "hyperportfolio", that can reliably and scalably evaluate societal analysis and decision-making power of agents, inspired by portfolio optimization with time series as assets to "invest". We also propose a novel Analyst-Assistant-Actuator architecture for the hyperportfolio task, and a Hypothesis & Proof prompting for producing in-depth analyses on input news, articles, etc. to assist decision-making. We perform experiments and ablation studies to explore the factors that impact performance. The results show that our proposed method achieves improvements of 32.4% and 30.4% compared to the state-of-the-art method in the two experimental settings.

Previous literature on policy diversity in reinforcement learning (RL) either focuses on the online setting or ignores the policy performance. In contrast, offline RL, which aims to learn high-quality policies from batched data, has yet to fully leverage the intrinsic diversity of the offline dataset. Addressing this dichotomy and aiming to balance quality and diversity poses a significant challenge to extant methodologies. This paper introduces a novel approach, termed Stylized Offline RL (SORL), which is designed to extract high-performing, stylistically diverse policies from a dataset characterized by distinct behavioral patterns. Drawing inspiration from the venerable Expectation-Maximization (EM) algorithm, SORL innovatively alternates between policy learning and trajectory clustering, a mechanism that promotes policy diversification. To further augment policy performance, we introduce advantage-weighted style learning into the SORL framework. Experimental evaluations across multiple environments demonstrate the significant superiority of SORL over previous methods in extracting high-quality policies with diverse behaviors. A case in point is that SORL successfully learns strong policies with markedly distinct playing patterns from a real-world human dataset of a popular basketball video game "Dunk City Dynasty."

Deep learning (DL) models for tabular data problems (e.g. classification, regression) are currently receiving increasingly more attention from researchers. However, despite the recent efforts, the non-DL algorithms based on gradient-boosted decision trees (GBDT) remain a strong go-to solution for these problems. One of the research directions aimed at improving the position of tabular DL involves designing so-called retrieval-augmented models. For a target object, such models retrieve other objects (e.g. the nearest neighbors) from the available training data and use their features and labels to make a better prediction. In this work, we present TabR -- essentially, a feed-forward network with a custom k-Nearest-Neighbors-like component in the middle. On a set of public benchmarks with datasets up to several million objects, TabR marks a big step forward for tabular DL: it demonstrates the best average performance among tabular DL models, becomes the new state-of-the-art on several datasets, and even outperforms GBDT models on the recently proposed "GBDT-friendly" benchmark (see Figure 1). Among the important findings and technical details powering TabR, the main ones lie in the attention-like mechanism that is responsible for retrieving the nearest neighbors and extracting valuable signal from them. In addition to the higher performance, TabR is simple and significantly more efficient compared to prior retrieval-based tabular DL models.

For more than a decade, researchers have measured progress in object recognition on the ImageNet dataset along with its associated generalization benchmarks such as ImageNet-A, -C, and -R. Recent advances in foundation models, trained on orders of magnitude more data, have begun to saturate performance on these benchmarks. Despite this progress, even today’s best models are brittle in practice. As a step toward more holistic measurement of model reliability, we propose studying performance on crowdsourced, global datasets, which contain natural distribution shifts seen practically in deployment. We perform a comprehensive empirical study on two crowdsourced, globally representative datasets, evaluating nearly 100 vision models to uncover several concerning empirical trends: first, that progress on crowdsourced, global data has significantly lagged behind standard benchmarks, with advances on ImageNet occurring at $2.5x$ the rate of progress on crowdsourced, global data. Second, we find that progress on standard benchmarks has failed to improve or exacerbated geographic disparities: geographic disparities between the least performant models and today's best models have more than tripled. We showcase the promise of using more curated and/or representative training datasets for mitigating these trends, and emphasize curation of web-scale, geographically representative training datasets as a critical open problem for the research community.

The ultimate goal of Dataset Distillation is to synthesize a small synthetic dataset such that a model trained on this synthetic set will perform equally well as a model trained on the full, real dataset. Until now, no method of Dataset Distillation has reached this completely lossless goal, in part due to the fact that previous methods only remain effective when the total number of synthetic samples is extremely small. Since only so much information can be contained in such a small number of samples, it seems that to achieve truly loss dataset distillation, we must develop a distillation method that remains effective as the size of the synthetic dataset grows. In this work, we present such an algorithm and elucidate why existing methods fail to generate larger, high-quality synthetic sets. Current state-of-the-art methods rely on trajectory-matching, or optimizing the synthetic data to induce similar long-term training dynamics as the real data. We empirically find that the training stage of the trajectories we choose to match (i.e., early or late) greatly affects the effectiveness of the distilled dataset. Specifically, early trajectories (where the teacher network learns easy patterns) work well for a low-cardinality synthetic set since there are fewer examples wherein to distribute the necessary information. Conversely, late trajectories (where the teacher network learns hard patterns) provide better signals for larger synthetic sets since there are now enough samples to represent the necessary complex patterns. Based on our findings, we propose to align the difficulty of the generated patterns with the size of the synthetic dataset. In doing so, we successfully scale trajectory matching-based methods to larger synthetic datasets, achieving lossless dataset distillation for the very first time. Code and distilled datasets will be released.

We study the implications of the modeling choice to use a graph, instead of a hypergraph, to represent real-world interconnected systems whose constituent relationships are of higher order by nature. Such a modeling choice typically involves an underlying projection process that maps the original hypergraph onto a graph, and is prevalent in graph-based analysis. While hypergraph projection can potentially lead to loss of higher-order relations, there exists very limited studies on the consequences of doing so, as well as its remediation. This work fills this gap by doing two things: (1) we develop analysis based on graph and set theory, showing two ubiquitous patterns of hyperedges that are root to structural information loss in all hypergraph projections; we also quantify the combinatorial impossibility of recovering the lost higher-order structures if no extra help is provided; (2) we still seek to recover the lost higher-order structures in hypergraph projection, and in light of (1)'s findings we make reasonable assumptions to allow the help of some prior knowledge of the application domain. Under this problem setting, we develop a learning-based hypergraph reconstruction method based on an important statistic of hyperedge distributions that we find. Our reconstruction method is systematically evaluated on 8 real-world datasets under different settings, and exhibits consistently top performance. We also demonstrate benefits of the reconstructed hypergraphs through use cases of protein rankings and link predictions.

Dataset distillation aims to condense large datasets into a small number of synthetic examples that can be used as drop-in replacements when training new models. It has applications to interpretability, neural architecture search, privacy, and continual learning. Despite strong successes in supervised domains, such methods have not yet been extended to reinforcement learning, where the lack of a fixed dataset renders most distillation methods unusable. Filling the gap, we formalize $\textit{behaviour distillation}$, a setting that aims to discover and then condense the information required for training an expert policy into a synthetic dataset of state-action pairs, $\textit{without access to expert data}$. We then introduce Hallucinating Datasets with Evolution Strategies (HaDES), a method for behaviour distillation that can discover datasets of $\textit{just four}$ state-action pairs which, under supervised learning, train agents to competitive performance levels in continuous control tasks. We show that these datasets generalize out of distribution to training policies with a wide range of architectures and hyperparameters. We also demonstrate application to a downstream task, namely training multi-task agents in a zero-shot fashion. Beyond behaviour distillation, HaDES provides significant improvements in neuroevolution for RL over previous approaches and achieves SoTA results on one standard supervised dataset distillation task. Finally, we show that visualizing the synthetic datasets can provide human-interpretable task insights.

Newborn brains rapidly learn to solve challenging object recognition tasks, including segmenting objects from backgrounds and recognizing objects across novel backgrounds and viewpoints. Conversely, modern machine-learning (ML) algorithms are "data hungry," requiring more training data than brains to reach similar performance levels. How do we close this learning gap between brains and machines? Here we introduce a new benchmark—a Newborn Embodied Turing Test (NETT) for object segmentation—in which newborn animals and machines are raised in the same environments and tested with the same tasks, permitting direct comparison of their learning abilities. First, we raised newborn chicks in controlled environments containing a single object rotating on a single background, then tested their ability to recognize that object across new backgrounds and viewpoints. Second, we performed “digital twin” experiments in which we reared and tested artificial chicks in virtual environments that mimicked the rearing and testing conditions of the biological chicks. We inserted a variety of ML “brains” into the artificial chicks and measured whether those algorithms learned common object recognition behavior as biological chicks. All biological chicks solved this one-shot object segmentation task, successfully learning background-invariant object representations that generalized across new backgrounds and viewpoints. In contrast, none of the artificial chicks solved this object segmentation task, instead learning background-dependent representations that failed to generalize across new backgrounds and viewpoints. This digital twin design exposes core limitations in current ML algorithms in achieving brain-like object perception. Our NETT is publicly available for comparing ML algorithms with newborn chicks. Ultimately, we anticipate that NETT benchmarks will allow researchers to build embodied AI systems that learn as efficiently and robustly as newborn brains.

Some reinforcement learning (RL) algorithms have the capability of recombining together pieces of previously seen experience to solve a task never seen before during training. This oft-sought property is one of the few ways in which dynamic programming based RL algorithms are considered different from supervised learning (SL) based RL algorithms. Yet, recent RL methods based on off-the-shelf SL algorithms achieve excellent results without an explicit mechanism for stitching; it remains unclear whether those methods forgo this important stitching property. This paper studies this question in the setting of goal-reaching problems. We show that the desirable stitching property corresponds to a form of generalization: after training on a distribution of (state, goal) pairs, one would like to evaluate on (state, goal) pairs not seen together in the training data. Our analysis shows that this sort of generalization is different from i.i.d. generalization. This connection between stitching and generalization reveals why we should not expect existing RL methods based on SL to perform stitching, even in the limit of large datasets and models. We experimentally validate this result on carefully constructed datasets. This connection suggests a simple remedy, the same remedy for improving generalization in supervised learning: data augmentation. We propose a naive temporal data augmentation approach and demonstrate that adding it to RL methods based on SL enables them to successfully stitch together experience, so that they succeed in navigating between states and goals unseen together during training.

Fairness in artificial intelligence models has gained significantly more attention in recent years, especially in the area of medicine, as fairness in medical models is critical to people's well-being and lives. High-quality medical fairness datasets are needed to promote fairness learning research. Existing medical fairness datasets are all for classification tasks, and no fairness datasets are available for medical segmentation, while medical segmentation is an equally important clinical task as classifications, which can provide detailed spatial information on organ abnormalities ready to be assessed by clinicians. In this paper, we propose the first fairness dataset for medical segmentation named Harvard-FairSeg with 10,000 subject samples. In addition, we propose a fair error-bound scaling approach to reweight the loss function with the upper error-bound in each identity group, using the segment anything model (SAM). We anticipate that the segmentation performance equity can be improved by explicitly tackling the hard cases with high training errors in each identity group. To facilitate fair comparisons, we utilize a novel equity-scaled segmentation performance metric to compare segmentation metrics in the context of fairness, such as the equity-scaled Dice coefficient. Through comprehensive experiments, we demonstrate that our fair error-bound scaling approach either has superior or comparable fairness performance to the state-of-the-art fairness learning models. The dataset and code are publicly accessible via https://ophai.hms.harvard.edu/datasets/harvard-fairseg10k.

Despite the frequent challenges posed by ambiguity when representing meaning via natural language, it is often ignored or deliberately removed in tasks mapping language to formally-designed representations, which generally assume a one-to-one mapping between linguistic and formal representations. We attempt to address this shortcoming by introducing AmP, a framework, dataset, and challenge for translating ambiguous natural language to formal representations like logic and code. We define templates and generate data for five well-documented linguistic ambiguities. Using AmP, we investigate how several few-shot text-to-code systems handle ambiguity, introducing three new metrics. We find that large pre-trained models perform poorly at capturing the distribution of possible meanings without deliberate instruction. However, models are able to capture the distribution well when ambiguity is attested in their inputs. These results motivate a call for including ambiguity explicitly in datasets and promote considering the distribution of possible outputs when evaluating systems. We release our data and code.

Explaining multivariate time series is a compound challenge, as it requires identifying important locations in the time series and matching complex temporal patterns. Although previous saliency-based methods addressed the challenges, their perturbation may not alleviate the distribution shift issue, which is inevitable especially in heterogeneous samples. We present ContraLSP, a locally sparse model that introduces counterfactual samples to build uninformative perturbations but keeps distribution using contrastive learning. Furthermore, we incorporate sample-specific sparse gates to generate more binary-skewed and smooth masks, which easily integrate temporal trends and select the salient features parsimoniously. Empirical studies on both synthetic and real-world datasets show that ContraLSP outperforms state-of-the-art models, demonstrating a substantial improvement in explanation quality for time series data. The source code is available at \url{https://github.com/zichuan-liu/ContraLSP}.

The effect of underrepresentation on the performance of minority groups is known to be a serious problem in supervised learning settings; however, it has been underexplored so far in the context of self-supervised learning (SSL). In this paper, we demonstrate that contrastive learning (CL), a popular variant of SSL, tends to collapse representations of minority groups with certain majority groups. We refer to this phenomenon as representation harm and demonstrate it on image and text datasets using the corresponding popular CL methods. Furthermore, our causal mediation analysis of allocation harm on a downstream classification task reveals that representation harm is partly responsible for it, thus emphasizing the importance of studying and mitigating representation harm. Finally, we provide a theoretical explanation for representation harm using a stochastic block model that leads to a representational neural collapse in a contrastive learning setting.

Transductive node prediction has been a popular learning setting in Graph Neural Networks (GNNs). It has been widely observed that the shortage of information flow between the distant nodes and intra-batch nodes (for large-scale graphs) often hurt the generalization of GNNs which overwhelmingly adopt message-passing. Yet there is still no formal and direct theoretical results to quantitatively capture the underlying mechanism, despite the recent advance in both theoretical and empirical studies for GNN's generalization ability. In this paper, the $L$-hop interplay (i.e., message passing capability with training nodes) for a $L$-layer GNN is successfully incorporated in our derived PAC-Bayesian bound for GNNs in the semi-supervised transductive setting. In other words, we quantitatively show how the interplay between training and testing sets influence the generalization ability which also partly explains the effectiveness of some existing empirical methods for enhancing generalization. Based on this result, we further design a plug-and-play ***Graph** **G**lobal **W**orkspace* module for GNNs (InterpGNN-GW) to enhance the interplay, utilizing the key-value attention mechanism to summarize crucial nodes' embeddings into memory and broadcast the memory to all nodes, in contrast to the pairwise attention scheme in previous graph transformers. Extensive experiments on both small-scale and large-scale graph datasets validate the effectiveness of our theory and approaches.

Analyzing model performance in various unseen environments is a critical research problem in the machine learning community. To study this problem, it is important to construct a testbed with out-of-distribution test sets that have broad coverage of environmental discrepancies. However, existing testbeds typically either have a small number of domains or are synthesized by image corruptions, hindering algorithm design that demonstrates real-world effectiveness. In this paper, we introduce CIFAR-10-Warehouse, consisting of 180 datasets collected by prompting image search engines and diffusion models in various ways. Generally sized between 300 and 8,000 images, the datasets contain natural images, cartoons, certain colors, or objects that do not naturally appear. With CIFAR-10-W, we aim to enhance the evaluation and deepen the understanding of two generalization tasks: domain generalization and model accuracy prediction in various out-of-distribution environments. We conduct extensive benchmarking and comparison experiments and show that CIFAR-10-W offers new and interesting insights inherent to these tasks. We also discuss other fields that would benefit from CIFAR-10-W. Data and code are available at https://sites.google.com/view/CIFAR-10-warehouse/.

Understanding and developing optimal representations has long been foundational in machine learning (ML). While disentangled representations have shown promise in generative modeling and representation learning, their downstream usefulness remains debated. Recent studies re-defined disentanglement through a formal connection to symmetries, emphasizing the ability to reduce latent domains (i.e., ML problem spaces) and consequently enhance data efficiency and generative capabilities. However, from an information theory viewpoint, assigning a complex attribute (i.e., features) to a specific latent variable may be infeasible, limiting the applicability of disentangled representations to simple datasets. In this work, we introduce $\alpha$-TCVAE, a variational autoencoder optimized using a novel total correlation (TC) lower bound that maximizes disentanglement and latent variables informativeness. The proposed TC bound is grounded in information theory constructs, generalizes the $\beta$-VAE lower bound, and can be reduced to a convex combination of the known variational information bottleneck (VIB) and conditional entropy bottleneck (CEB) terms. Moreover, we present quantitative analyses and correlation studies that support the idea that smaller latent domains (i.e., disentangled representations) lead to better generative capabilities and diversity. Additionally, we perform downstream task experiments from both representation and RL domains to assess our questions from a broader ML perspective. Our results demonstrate that $\alpha$-TCVAE consistently learns more disentangled representations than baselines and generates more diverse observations without sacrificing visual fidelity. Notably, $\alpha$-TCVAE exhibits marked improvements on MPI3D-Real, the most realistic disentangled dataset in our study, confirming its ability to represent complex datasets when maximizing the informativeness of individual variables. Finally, testing the proposed model off-the-shelf on a state-of-the-art model-based RL agent, Director, significantly shows $\alpha$-TCVAE downstream usefulness on the loconav Ant Maze task. Implementation available at https://github.com/Cmeo97/Alpha-TCVAE

Large Language Models (LLMs) have greatly advanced code auto-completion systems, with a potential for substantial productivity enhancements for developers. However, current benchmarks mainly focus on single-file tasks, leaving an assessment gap for more complex, real-world, multi-file programming scenarios. To fill this gap, we introduce RepoBench, a new benchmark specifically designed for evaluating repository-level code auto-completion systems. RepoBench consists of three interconnected evaluation tasks: RepoBench-R (Retrieval), RepoBench-C (Code Completion), and RepoBench-P (Pipeline). Each task respectively measures the system's ability to retrieve the most relevant code snippets from other files as cross-file context, predict the next line of code with cross-file and in-file context, and handle complex tasks that require a combination of both retrieval and next-line prediction. RepoBench aims to facilitate a more complete comparison of performance and encouraging continuous improvement in auto-completion systems. RepoBench is actively maintained with the latest code, serving as a live benchmark publicly available at https://github.com/Leolty/repobench.