ICML 2024 Papers

In offline reinforcement learning (RL), the performance of the learned policy highly depends on the quality of offline datasets. However, the offline dataset contains very limited optimal trajectories in many cases. This poses a challenge for offline RL algorithms, as agents must acquire the ability to transit to high-reward regions. To address this issue, we introduce Diffusionbased Trajectory Stitching (DiffStitch), a novel diffusion-based data augmentation pipeline that systematically generates stitching transitions between trajectories. DiffStitch effectively connects low-reward trajectories with high-reward trajectories, forming globally optimal trajectories and thereby mitigating the challenges faced by offline RL algorithms in learning trajectory stitching. Empirical experiments conducted on D4RL datasets demonstrate the effectiveness of our pipeline across RL methodologies. Notably, DiffStitch demonstrates substantial enhancements in the performance of one-step methods(IQL), imitation learning methods(TD3+BC) and trajectory optimization methods(DT). Our code is publicly available at https://github.com/guangheli12/DiffStitch

Establishing robust policies is essential to counter attacks or disturbances affecting deep reinforcement learning (DRL) agents. Recent studies explore state-adversarial robustness and suggest the potential lack of an optimal robust policy (ORP), posing challenges in setting strict robustness constraints. This work further investigates ORP: At first, we introduce a consistency assumption of policy (CAP) stating that optimal actions in the Markov decision process remain consistent with minor perturbations, supported by empirical and theoretical evidence. Building upon CAP, we crucially prove the existence of a deterministic and stationary ORP that aligns with the Bellman optimal policy. Furthermore, we illustrate the necessity of $L^{\infty}$-norm when minimizing Bellman error to attain ORP. This finding clarifies the vulnerability of prior DRL algorithms that target the Bellman optimal policy with $L^{1}$-norm and motivates us to train a Consistent Adversarial Robust Deep Q-Network (CAR-DQN) by minimizing a surrogate of Bellman Infinity-error. The top-tier performance of CAR-DQN across various benchmarks validates its practical effectiveness and reinforces the soundness of our theoretical analysis.

Dataset distillation aims to synthesize a small number of images per class (IPC) from a large dataset to approximate full dataset training with minimal performance loss. While effective in very small IPC ranges, many distillation methods become less effective, even underperforming random sample selection, as IPC increases. Our examination of state-of-the-art trajectory-matching based distillation methods across various IPC scales reveals that these methods struggle to incorporate the complex, rare features of harder samples into the synthetic dataset even with the increased IPC, resulting in a persistent coverage gap between easy and hard test samples. Motivated by such observations, we introduce SelMatch, a novel distillation method that effectively scales with IPC. SelMatch uses selection-based initialization and partial updates through trajectory matching to manage the synthetic dataset's desired difficulty level tailored to IPC scales. When tested on CIFAR-10/100 and TinyImageNet, SelMatch consistently outperforms leading selection-only and distillation-only methods across subset ratios from 5% to 30%.

Uncertainty estimation is crucial for deep learning models to detect out-of-distribution (OOD) inputs. However, the naive deep learning classifiers produce uncalibrated uncertainty for OOD data. Improving the uncertainty estimation typically requires external data for OOD-aware training or considerable costs to build an ensemble. In this work, we improve on uncertainty estimation without extra OOD data or additional inference costs using an alternative *Split-Ensemble* method. Specifically, we propose a novel *subtask-splitting* ensemble training objective where a task is split into several complementary subtasks based on feature similarity. Each subtask considers part of the data as in distribution while all the rest as OOD data. Diverse submodels can therefore be trained on each subtask with OOD-aware objectives, learning generalizable uncertainty estimation. To avoid overheads, we enable low-level feature sharing among submodels, building a tree-like Split-Ensemble architecture via iterative splitting and pruning. Empirical study shows Split-Ensemble, without additional computational cost, improves accuracy over a single model by 0.8%, 1.8%, and 25.5% on CIFAR-10, CIFAR-100, and Tiny-ImageNet, respectively. OOD detection for the same backbone and in-distribution datasets surpasses a single model baseline by 2.2%, 8.1%, and 29.6% in mean AUROC, respectively.

In this paper, we utilize hyperspheres and regular $n$-simplexes and propose an approach to learning deep features equivariant under the transformations of $n$D reflections and rotations, encompassed by the powerful group of O$(n)$. Namely, we propose O$(n)$-equivariant neurons with spherical decision surfaces that generalize to any dimension $n$, which we call Deep Equivariant Hyperspheres. We demonstrate how to combine them in a network that directly operates on the basis of the input points and propose an invariant operator based on the relation between two points and a sphere, which as we show, turns out to be a Gram matrix. Using synthetic and real-world data in $n$D, we experimentally verify our theoretical contributions and find that our approach is superior to the competing methods for O$(n)$-equivariant benchmark datasets (classification and regression), demonstrating a favorable speed/performance trade-off. The code is available on [GitHub](https://github.com/pavlo-melnyk/equivariant-hyperspheres).

Existing Large Language Models (LLMs) usually remain static after deployment, which might make it hard to inject new knowledge into the model. We aim to build models containing a considerable portion of self-updatable parameters, enabling the model to integrate new knowledge effectively and efficiently. To this end, we introduce MEMORYLLM, a model that comprises a transformer and a fixed-size memory pool within the latent space of the transformer. MEMORYLLM can self-update with text knowledge and memorize the knowledge injected earlier. Our evaluations demonstrate the ability of MEMORYLLM to effectively incorporate new knowledge, as evidenced by its performance on model editing benchmarks. Meanwhile, the model exhibits long-term information retention capacity, which is validated through our custom-designed evaluations and long-context benchmarks. MEMORYLLM also shows operational integrity without any sign of performance degradation even after nearly a million memory updates. Our code and model are open-sourced at https://github.com/wangyu-ustc/MemoryLLM.

This paper presents FedType, a simple yet pioneering framework designed to fill research gaps in heterogeneous model aggregation within federated learning (FL). FedType introduces small identical proxy models for clients, serving as agents for information exchange, ensuring model security, and achieving efficient communication simultaneously. To transfer knowledge between large private and small proxy models on clients, we propose a novel uncertainty-based asymmetrical reciprocity learning method, eliminating the need for any public data. Comprehensive experiments conducted on benchmark datasets demonstrate the efficacy and generalization ability of FedType across diverse settings. Our approach redefines federated learning paradigms by bridging model heterogeneity, eliminating reliance on public data, prioritizing client privacy, and reducing communication costs (The codes are available in the supplementation materials).

Use of machine learning to perform database operations, such as indexing, cardinality estimation, and sorting, is shown to provide substantial performance benefits. However, when datasets change and data distribution shifts, empirical results also show performance degradation for learned models, possibly to worse than non-learned alternatives. This, together with a lack of theoretical understanding of learned methods undermines their practical applicability, since there are no guarantees on how well the models will perform after deployment. In this paper, we present the first known theoretical characterization of the performance of learned models in dynamic datasets, for the aforementioned operations. Our results show novel theoretical characteristics achievable by learned models and provide bounds on the performance of the models that characterize their advantages over non-learned methods, showing why and when learned models can outperform the alternatives. Our analysis develops the *distribution learnability* framework and novel theoretical tools which build the foundation for the analysis of learned database operations in the future.

Data subset selection aims to find a smaller yet informative subset of a large dataset that can approximate the full-dataset training, addressing challenges associated with training neural networks on large-scale datasets. However, existing methods tend to specialize in either high or low selection ratio regimes, lacking a universal approach that consistently achieves competitive performance across a broad range of selection ratios. We introduce a universal and efficient data subset selection method, Best Window Selection (BWS), by proposing a method to choose the best window subset from samples ordered based on their difficulty scores. This approach offers flexibility by allowing the choice of window intervals that span from easy to difficult samples. Furthermore, we provide an efficient mechanism for selecting the best window subset by evaluating its quality using kernel ridge regression. Our experimental results demonstrate the superior performance of BWS compared to other baselines across a broad range of selection ratios over datasets, including CIFAR-10/100 and ImageNet, and the scenarios involving training from random initialization or fine-tuning of pre-trained models.

The *Rashomon Effect*, coined by Leo Breiman, describes the phenomenon that there exist many equally good predictive models for the same dataset. This phenomenon happens for many real datasets and when it does, it sparks both magic and consternation, but mostly magic. In light of the Rashomon Effect, this perspective piece proposes reshaping the way we think about machine learning, particularly for tabular data problems in the nondeterministic (noisy) setting. We address how the Rashomon Effect impacts (1) the existence of simple-yet-accurate models, (2) flexibility to address user preferences, such as fairness and monotonicity, without losing performance, (3) uncertainty in predictions, fairness, and explanations, (4) reliable variable importance, (5) algorithm choice, specifically, providing advanced knowledge of which algorithms might be suitable for a given problem, and (6) public policy. We also discuss a theory of when the Rashomon Effect occurs and why. Our goal is to illustrate how the Rashomon Effect can have a massive impact on the use of machine learning for complex problems in society.

Robotic manipulation tasks often span over long horizons and encapsulate multiple subtasks with different skills. Learning policies directly from long-horizon demonstrations is challenging without intermediate keyframes guidance and corresponding skill annotations. Existing approaches for keyframe identification often struggle to offer reliable decomposition for low accuracy and fail to provide semantic relevance between keyframes and skills. For this, we propose a unified **K**eyframe **I**dentifier and **S**kill **A**notator (**KISA**) that utilizes pretrained visual-language representations for precise and interpretable decomposition of unlabeled demonstrations. Specifically, we develop a simple yet effective temporal enhancement module that enriches frame-level representations with expanded receptive fields to capture semantic dynamics at the video level. We further propose coarse contrastive learning and fine-grained monotonic encouragement to enhance the alignment between visual representations from keyframes and language representations from skills. The experimental results across three benchmarks demonstrate that KISA outperforms competitive baselines in terms of accuracy and interpretability of keyframe identification. Moreover, KISA exhibits robust generalization capabilities and the flexibility to incorporate various pretrained representations.

Deep learning models are known to overfit and memorize spurious features in the training dataset. While numerous empirical studies have aimed at understanding this phenomenon, a rigorous theoretical framework to quantify it is still missing. In this paper, we consider spurious features that are uncorrelated with the learning task, and we provide a precise characterization of how they are memorized via two separate terms: _(i)_ the _stability_ of the model with respect to individual training samples, and _(ii)_ the _feature alignment_ between the spurious pattern and the full sample. While the first term is well established in learning theory and it is connected to the generalization error in classical work, the second one is, to the best of our knowledge, novel. Our key technical result gives a precise characterization of the feature alignment for the two prototypical settings of random features (RF) and neural tangent kernel (NTK) regression. We prove that the memorization of spurious features weakens as the generalization capability increases and, through the analysis of the feature alignment, we unveil the role of the model and of its activation function. Numerical experiments show the predictive power of our theory on standard datasets (MNIST, CIFAR-10).

Given a dataset of individuals each described by a covariate vector, a treatment, and an observed outcome on the treatment, the goal of the individual treatment effect (ITE) estimation task is to predict outcome changes resulting from a change in treatment. A fundamental challenge is that in the observational data, a covariate’s outcome is observed only under one treatment, whereas we need to infer the difference in outcomes under two different treatments. Several existing approaches address this issue through training with inferred pseudo-outcomes, but their success relies on the quality of these pseudo-outcomes. We propose PairNet, a novel ITE estimation training strategy that minimizes losses over pairs of examples based on their factual observed outcomes. Theoretical analysis for binary treatments reveals that PairNet is a consistent estimator of ITE risk, and achieves smaller generalization error than baseline models. Empirical comparison with thirteen existing methods across eight benchmarks, covering both discrete and continuous treatments, shows that PairNet achieves significantly lower ITE error compared to the baselines. Also, it is model-agnostic and easy to implement.

Denoising is intuitively related to projection. Indeed, under the manifold hypothesis, adding random noise is approximately equivalent to orthogonal perturbation. Hence, learning to denoise is approximately learning to project. In this paper, we use this observation to interpret denoising diffusion models as approximate gradient descent applied to the Euclidean distance function. We then provide straight-forward convergence analysis of the DDIM sampler under simple assumptions on the projection error of the denoiser. Finally, we propose a new gradient-estimation sampler, generalizing DDIM using insights from our theoretical results. In as few as 5-10 function evaluations, our sampler achieves state-of-the-art FID scores on pretrained CIFAR-10 and CelebA models and can generate high quality samples on latent diffusion models.

The diversity of contexts in which large language models (LLMs) are deployed requires the ability to modify or customize default model behaviors to incorporate nuanced requirements and preferences. A convenient interface to specify such model adjustments is high-level verbal feedback, such as “Don’t use emojis when drafting emails to my boss.” However, while writing high-level feedback is far simpler than collecting annotations for reinforcement learning from human feedback (RLHF), we find that simply prompting a model with such feedback leads to $\textbf{overgeneralization}$–applying feedback in contexts where it is not relevant. We propose a new method Contextualized Critiques with Constrained Preference Optimization (C3PO) to learn from high-level verbal feedback while reducing overgeneralization compared to current work. C3PO uses a piece of high-level feedback to generate a small synthetic preference dataset to specify when and how the feedback should (and should not) be applied. It then fine-tunes the model in accordance with the synthetic preference data while minimizing the divergence from the original model for prompts where the feedback does not apply. Our experimental results indicate that our approach effectively applies verbal feedback to relevant scenarios while preserving existing behaviors for other contexts more than current methods. For both human- and GPT-4-generated high-level feedback, C3PO effectively adheres to the given feedback comparably to in-context baselines while reducing overgeneralization by 30%.

On a variety of tasks, the performance of neural networks predictably improves with training time, dataset size and model size across many orders of magnitude. This phenomenon is known as a neural scaling law. Of fundamental importance is the compute-optimal scaling law, which reports the performance as a function of units of compute when choosing model sizes optimally. We analyze a random feature model trained with gradient descent as a solvable model of network training and generalization. This reproduces many observations about neural scaling laws. First, our model makes a prediction about why the scaling of performance with training time and with model size have different power law exponents. Consequently, the theory predicts an asymmetric compute-optimal scaling rule where the number of training steps are increased faster than model parameters, consistent with recent empirical observations. Second, it has been observed that early in training, networks converge to their infinite-width dynamics at a rate $1/\text{width}$ but at late time exhibit a rate $\text{width}^{-c}$, where $c$ depends on the structure of the architecture and task. We show that our model exhibits this behavior. Lastly, our theory shows how the gap between training and test loss can gradually build up over time due to repeated reuse of data.

Diffusion models (DMs) have established themselves as the state-of-the-art generative modeling approach in the visual domain and beyond. A crucial drawback of DMs is their slow sampling speed, relying on many sequential function evaluations through large neural networks. Sampling from DMs can be seen as solving a differential equation through a discretized set of noise levels known as the sampling schedule. While past works primarily focused on deriving efficient solvers, little attention has been given to finding optimal sampling schedules, and the entire literature relies on hand-crafted heuristics. In this work, for the first time, we propose a general and principled approach to optimizing the sampling schedules of DMs for high-quality outputs, called Align Your Steps. We leverage methods from stochastic calculus and find optimal schedules specific to different solvers, trained DMs and datasets. We evaluate our novel approach on several image, video as well as 2D toy data synthesis benchmarks, using a variety of different samplers, and observe that our optimized schedules outperform previous hand-crafted schedules in almost all experiments. Our method demonstrates the untapped potential of sampling schedule optimization, especially in the few-step synthesis regime.

This paper studies policy evaluation with multiple data sources, especially in scenarios that involve one experimental dataset with two arms, complemented by a historical dataset generated under a single control arm. We propose novel data integration methods that linearly integrate base policy value estimators constructed based on the experimental and historical data, with weights optimized to minimize the mean square error (MSE) of the resulting combined estimator. We further apply the pessimistic principle to obtain more robust estimators, and extend these developments to sequential decision making. Theoretically, we establish non-asymptotic error bounds for the MSEs of our proposed estimators, and derive their oracle, efficiency and robustness properties across a broad spectrum of reward shift scenarios. Numerical experiments and real-data-based analyses from a ridesharing company demonstrate the superior performance of the proposed estimators.

We study how to make decisions that minimize Bayesian regret in offline linear bandits. Prior work suggests that one must take actions with maximum lower confidence bound (LCB) on their reward. We argue that reliance on LCB is inherently flawed in this setting and propose a new algorithm that directly minimizes upper-bounds on the Bayesian regret using efficient conic optimization solvers. Our bounds build heavily on new connections to monetary risk measures. Proving a matching lower-bound, we show that our upper-bounds are tight, and by minimizing them we are guaranteed to outperform the LCB approach. Our numerical results on synthetic domains confirm that our approach is superior to maximizing LCB.