NeurIPS 2024 Papers

Policy Mirror Descent (PMD) is a powerful and theoretically sound methodology for sequential decision-making. However, it is not directly applicable to Reinforcement Learning (RL) due to the inaccessibility of explicit action-value functions. We address this challenge by introducing a novel approach based on learning a world model of the environment using conditional mean embeddings. Leveraging tools from operator theory we derive a closed-form expression of the action-value function in terms of the world model via simple matrix operations. Combining these estimators with PMD leads to POWR, a new RL algorithm for which we prove convergence rates to the global optimum. Preliminary experiments in finite and infinite state settings support the effectiveness of our method.

When releasing outputs from confidential data, agencies need to balance the analytical usefulness of the released data with the obligation to protect data subjects' confidentiality. For releases satisfying differential privacy, this balance is reflected by the privacy budget, $\varepsilon$. We provide a framework for setting $\varepsilon$ based on its relationship with Bayesian posterior probabilities of disclosure. The agency responsible for the data release decides how much posterior risk it is willing to accept at various levels of prior risk, which implies a unique $\varepsilon$. Agencies can evaluate different risk profiles to determine one that leads to an acceptable trade-off in risk and utility.

Offline Reinforcement Learning (Offline RL) presents challenges of learning effective decision-making policies from static datasets without any online interactions. Data augmentation techniques, such as noise injection and data synthesizing, aim to improve Q-function approximation by smoothing the learned state-action region. However, these methods often fall short of directly improving the quality of offline datasets, leading to suboptimal results. In response, we introduce GTA, Generative Trajectory Augmentation, a novel generative data augmentation approach designed to enrich offline data by augmenting trajectories to be both high-rewarding and dynamically plausible. GTA applies a diffusion model within the data augmentation framework. GTA partially noises original trajectories and then denoises them with classifier-free guidance via conditioning on amplified return value. Our results show that GTA, as a general data augmentation strategy, enhances the performance of widely used offline RL algorithms across various tasks with unique challenges. Furthermore, we conduct a quality analysis of data augmented by GTA and demonstrate that GTA improves the quality of the data. Our code is available at https://github.com/Jaewoopudding/GTA

We introduce a multi-winner reconfiguration model to examine how to transition between subsets of alternatives (aka. committees) through a sequence of minor yet impactful modifications, called reconfiguration path. We analyze this model under four approval-based voting rules: Chamberlin-Courant (CC), Proportional Approval Voting (PAV), Approval Voting (AV), and Satisfaction Approval Voting (SAV). The problem exhibits computational intractability for CC and PAV, and polynomial solvability for AV and SAV. We provide a detailed multivariate complexity analysis for CC and PAV, demonstrating that although the problem remains challenging in many scenarios, there are specific cases that allow for efficient parameterized algorithms.

Reinforcement Learning (RL) encompasses diverse paradigms, including model-based RL, policy-based RL, and value-based RL, each tailored to approximate the model, optimal policy, and optimal value function, respectively. This work investigates the potential hierarchy of representation complexity among these RL paradigms. By utilizing computational complexity measures, including time complexity and circuit complexity, we theoretically unveil a potential representation complexity hierarchy within RL. We find that representing the model emerges as the easiest task, followed by the optimal policy, while representing the optimal value function presents the most intricate challenge. Additionally, we reaffirm this hierarchy from the perspective of the expressiveness of Multi-Layer Perceptrons (MLPs), which align more closely with practical deep RL and contribute to a completely new perspective in theoretical studying representation complexity in RL. Finally, we conduct deep RL experiments to validate our theoretical findings.

Reinforcement Learning from Human Feedback (RLHF) has been pivotal in aligning Large Language Models with human values but often suffers from overoptimization due to its reliance on a proxy reward model. To mitigate this limitation, we first propose a lightweight uncertainty quantification method that assesses the reliability of the proxy reward using only the last layer embeddings of the reward model. Enabled by this efficient uncertainty quantification method, we formulate AdvPO, a distributionally robust optimization procedure to tackle the reward overoptimization problem in RLHF. Through extensive experiments on the Anthropic HH and TL;DR summarization datasets, we verify the effectiveness of AdvPO in mitigating the overoptimization problem, resulting in enhanced RLHF performance as evaluated through human-assisted evaluation.

Even though a variety of methods have been proposed in the literature, efficient and effective latent-space control (i.e., control in a learned low-dimensional space) of physical systems remains an open challenge. We argue that a promising avenue is to leverage powerful and well-understood closed-form strategies from control theory literature in combination with learned dynamics, such as potential-energy shaping. We identify three fundamental shortcomings in existing latent-space models that have so far prevented this powerful combination: (i) they lack the mathematical structure of a physical system, (ii) they do not inherently conserve the stability properties of the real systems, (iii) these methods do not have an invertible mapping between input and latent-space forcing. This work proposes a novel Coupled Oscillator Network (CON) model that simultaneously tackles all these issues. More specifically, (i) we show analytically that CON is a Lagrangian system - i.e., it possesses well-defined potential and kinetic energy terms. Then, (ii) we provide formal proof of global Input-to-State stability using Lyapunov arguments. Moving to the experimental side, we demonstrate that CON reaches SoA performance when learning complex nonlinear dynamics of mechanical systems directly from images. An additional methodological innovation contributing to achieving this third goal is an approximated closed-form solution for efficient integration of network dynamics, which eases efficient training. We tackle (iii) by approximating the forcing-to-input mapping with a decoder that is trained to reconstruct the input based on the encoded latent space force. Finally, we leverage these three properties and show that they enable latent-space control. We use an integral-saturated PID with potential force compensation and demonstrate high-quality performance on a soft robot using raw pixels as the only feedback information.

We introduce a general detection-based approach to text line recognition, be it printed (OCR) or handwritten text (HTR), with latin, chinese or ciphered characters. Detection-based approaches have until now largely been discarded for HTR because reading characters separately is often challenging, and character-level annotation is difficult and expensive. We overcome these challenges thanks to three main insights: (i) synthetic pre-training with diverse enough data to learn reasonable character localization in any script; (ii) modern transformer-based detectors can jointly detect a large number of instances and, if trained with an adequate masking strategy, leverage consistency between the different detections; (iii) once a pre-trained detection model with approximate character localization is available, it is possible to fine-tune it with line-level annotation on real data, even with a different alphabet. Our approach thus builds on a completely different paradigm than most state-of-the-art methods, which rely on autoregressive decoding, predicting character values one by one, while we treat a complete line in parallel. Remarkably, our method demonstrates good performance on range of scripts, usually tackled with specialized approaches: latin script, chinese script, and ciphers, for which we significantly improve state-of-the-art performances. Our code and models are available at [https://github.com/raphael-baena/DTLR](https://github.com/raphael-baena/DTLR).

Tabular data is prevalent in real-world machine learning applications, and new models for supervised learning of tabular data are frequently proposed. Comparative studies assessing performance differences typically have model-centered evaluation setups with overly standardized data preprocessing. This limits the external validity of these studies, as in real-world modeling pipelines, models are typically applied after dataset-specific preprocessing and feature engineering. We address this gap by proposing a data-centric evaluation framework. We select 10 relevant datasets from Kaggle competitions and implement expert-level preprocessing pipelines for each dataset. We conduct experiments with different preprocessing pipelines and hyperparameter optimization (HPO) regimes to quantify the impact of model selection, HPO, feature engineering, and test-time adaptation. Our main findings reveal: 1) After dataset-specific feature engineering, model rankings change considerably, performance differences decrease, and the importance of model selection reduces. 2) Recent models, despite their measurable progress, still significantly benefit from manual feature engineering. This holds true for both tree-based models and neural networks. 3) While tabular data is typically considered static, samples are often collected over time, and adapting to distribution shifts can be important even in supposedly static data. These insights suggest that research efforts should be directed toward a data-centric perspective, acknowledging that tabular data requires feature engineering and often exhibits temporal characteristics.

Designing novel functional proteins crucially depends on accurately modeling their fitness landscape. Given the limited availability of functional annotations from wet-lab experiments, previous methods have primarily relied on self-supervised models trained on vast, unlabeled protein sequence or structure datasets. While initial protein representation learning studies solely focused on either sequence or structural features, recent hybrid architectures have sought to merge these modalities to harness their respective strengths. However, these sequence-structure models have so far achieved only incremental improvements when compared to the leading sequence-only approaches, highlighting unresolved challenges effectively leveraging these modalities together. Moreover, the function of certain proteins is highly dependent on the granular aspects of their surface topology, which have been overlooked by prior models. To address these limitations, we introduce the Sequence-Structure-Surface Fitness (**S3F**) model — a novel multimodal representation learning framework that integrates protein features across several scales. Our approach combines sequence representations from a protein language model with Geometric Vector Perceptron networks encoding protein backbone and detailed surface topology. The proposed method achieves state-of-the-art fitness prediction on the ProteinGym benchmark encompassing 217 substitution deep mutational scanning assays, and provides insights into the determinants of protein function. Our code is at https://github.com/DeepGraphLearning/S3F.

Adversarial Training (AT) has been widely proved to be an effective method to improve the adversarial robustness against adversarial examples for Deep Neural Networks (DNNs). As a variant of AT, Adversarial Robustness Distillation (ARD) has demonstrated its superior performance in improving the robustness of small student models with the guidance of large teacher models. However, both AT and ARD encounter the robust fairness problem: these models exhibit strong robustness when facing part of classes (easy class), but weak robustness when facing others (hard class). In this paper, we give an in-depth analysis of the potential factors and argue that the smoothness degree of samples' soft labels for different classes (i.e., hard class or easy class) will affect the robust fairness of DNNs from both empirical observation and theoretical analysis. Based on the above finding, we propose an Anti-Bias Soft Label Distillation (ABSLD) method to mitigate the adversarial robust fairness problem within the framework of Knowledge Distillation (KD). Specifically, ABSLD adaptively reduces the student's error risk gap between different classes to achieve fairness by adjusting the class-wise smoothness degree of samples' soft labels during the training process, and the smoothness degree of soft labels is controlled by assigning different temperatures in KD to different classes. Extensive experiments demonstrate that ABSLD outperforms state-of-the-art AT, ARD, and robust fairness methods in the comprehensive metric (Normalized Standard Deviation) of robustness and fairness.

Expanding on recent work on scheduling with predicted job sizes, we consider the effect of the cost of predictions in queueing systems, removing the assumption in prior research that predictions are external to the system’s resources and/or cost-free. Additionally, we introduce a novel approach to utilizing predictions, SkipPredict, designed to address their inherent cost. Rather than uniformly applying predictions to all jobs, we propose a tailored approach that categorizes jobs to improve the effectiveness of prediction on performance. To achieve this, we employ one-bit “cheap predictions” to classify jobs as either short or long. SkipPredict prioritizes predicted short jobs over long jobs, and for the long jobs, SkipPredict applies a second round of more detailed “expensive predictions” to approximate Shortest Remaining Processing Time for these jobs. Importantly, our analyses take into account the cost of prediction. We derive closed-form formulas that calculate the mean response time of jobs with size predictions accounting for the prediction cost. We examine the effect of this cost for two distinct models in real-world and synthetic datasets. In the external cost model, predictions are generated by external method without impacting job service times but incur a cost. In the server time cost model, predictions themselves require server processing time and are scheduled on the same server as the jobs.

The goal of data attribution for text-to-image models is to identify the training images that most influence the generation of a new image. Influence is defined such that, for a given output, if a model is retrained from scratch without the most influential images, the model would fail to reproduce the same output. Unfortunately, directly searching for these influential images is computationally infeasible, since it would require repeatedly retraining models from scratch. In our work, we propose an efficient data attribution method by simulating unlearning the synthesized image. We achieve this by increasing the training loss on the output image, without catastrophic forgetting of other, unrelated concepts. We then identify training images with significant loss deviations after the unlearning process and label these as influential. We evaluate our method with a computationally intensive but "gold-standard" retraining from scratch and demonstrate our method's advantages over previous methods.

Text embeddings are essential for tasks such as document retrieval, clustering, and semantic similarity assessment. In this paper, we study how to contrastively train text embedding models in a compute-optimal fashion, given a suite of pretrained decoder-only language models. Our innovation is an algorithm that produces optimal configurations of model sizes, data quantities, and fine-tuning methods for text-embedding models at different computational budget levels. The resulting recipe, which we obtain through extensive experiments, can be used by practitioners to make informed design choices for their embedding models. Specifically, our findings suggest that full fine-tuning and Low-Rank Adaptation fine-tuning produce optimal models at lower and higher computational budgets respectively.

The Passive-Aggressive (PA) method is widely used in online regression problems for handling large-scale streaming data, typically updating model parameters in a passive-aggressive manner based on whether the error exceeds a predefined threshold. However, this approach struggles with determining optimal thresholds and adapting to complex scenarios with side information, where tracking accuracy is not the sole metric in the regression model. To address these challenges, we introduce a novel adaptive framework that allows finer adjustments to the weight vector in PA using side information. This framework adaptively selects the threshold parameter in PA, theoretically ensuring convergence to the optimal setting. Additionally, we present an efficient implementation of our algorithm that significantly reduces computational complexity. Numerical experiments show that our model achieves outstanding performance associated with the side information while maintaining low tracking error, demonstrating marked improvements over traditional PA methods across various scenarios.

Modern diffusion-based image generative models have made significant progress and become promising to enrich training data for the object detection task. However, the generation quality and the controllability for complex scenes containing multi-class objects and dense objects with occlusions remain limited. This paper presents ODGEN, a novel method to generate high-quality images conditioned on bounding boxes, thereby facilitating data synthesis for object detection. Given a domain-specific object detection dataset, we first fine-tune a pre-trained diffusion model on both cropped foreground objects and entire images to fit target distributions. Then we propose to control the diffusion model using synthesized visual prompts with spatial constraints and object-wise textual descriptions. ODGEN exhibits robustness in handling complex scenes and specific domains. Further, we design a dataset synthesis pipeline to evaluate ODGEN on 7 domain-specific benchmarks to demonstrate its effectiveness. Adding training data generated by ODGEN improves up to 25.3% mAP@.50:.95 with object detectors like YOLOv5 and YOLOv7, outperforming prior controllable generative methods. In addition, we design an evaluation protocol based on COCO-2014 to validate ODGEN in general domains and observe an advantage up to 5.6% in mAP@.50:.95 against existing methods.

Zero-Shot Temporal Action Detection (ZSTAD) aims to classify and localize action segments in untrimmed videos for unseen action categories. Most existing ZSTAD methods utilize a foreground-based approach, limiting the integration of text and visual features due to their reliance on pre-extracted proposals. In this paper, we introduce a cross-modal ZSTAD baseline with mutual cross-attention, integrating both text and visual information throughout the detection process. Our simple approach results in superior performance compared to previous methods. Despite this improvement, we further identify a common-action bias issue that the cross-modal baseline over-focus on common sub-actions due to a lack of ability to discriminate text-related visual parts. To address this issue, we propose Text-infused attention and Foreground-aware Action Detection (Ti-FAD), which enhances the ability to focus on text-related sub-actions and distinguish relevant action segments from the background. Our extensive experiments demonstrate that Ti-FAD outperforms the state-of-the-art methods on ZSTAD benchmarks by a large margin: 41.2\% (+ 11.0\%) on THUMOS14 and 32.0\% (+ 5.4\%) on ActivityNet v1.3. Code is available at: https://github.com/YearangLee/Ti-FAD.

We present differentially private algorithms for high-dimensional mean estimation. Previous private estimators on distributions over $\mathbb{R}^d$ suffer from a curse of dimensionality, as they require $\Omega(d^{1/2})$ samples to achieve non-trivial error, even in cases where $O(1)$ samples suffice without privacy. This rate is unavoidable when the distribution is isotropic, namely, when the covariance is a multiple of the identity matrix. Yet, real-world data is often highly anisotropic, with signals concentrated on a small number of principal components. We develop estimators that are appropriate for such signals---our estimators are $(\varepsilon,\delta)$-differentially private and have sample complexity that is dimension-independent for anisotropic subgaussian distributions. Given $n$ samples from a distribution with known covariance-proxy $\Sigma$ and unknown mean $\mu$, we present an estimator $\hat{\mu}$ that achieves error, $\|\hat{\mu}-\mu\|_2\leq \alpha$, as long as $n\gtrsim \text{tr}(\Sigma)/\alpha^2+ \text{tr}(\Sigma^{1/2})/(\alpha\varepsilon)$. We show that this is the optimal sample complexity for this task up to logarithmic factors. Moreover, for the case of unknown covariance, we present an algorithm whose sample complexity has improved dependence on the dimension, from $d^{1/2}$ to $d^{1/4}$.

The application of reinforcement learning (RL) involving interactions with individuals has grown significantly in recent years. These interactions, influenced by factors such as personal preferences and physiological differences, causally influence state transitions, ranging from health conditions in healthcare to learning progress in education. As a result, different individuals may exhibit different state-transition processes. Understanding individualized state-transition processes is essential for optimizing individualized policies. In practice, however, identifying these state-transition processes is challenging, as individual-specific factors often remain latent. In this paper, we establish the identifiability of these latent factors and introduce a practical method that effectively learns these processes from observed state-action trajectories. Experiments on various datasets show that the proposed method can effectively identify latent state-transition processes and facilitate the learning of individualized RL policies.