ICML 2024 Papers

In offline reinforcement learning, the challenge of out-of-distribution (OOD) is pronounced. To address this, existing methods often constrain the learned policy through policy regularization. However, these methods often suffer from the issue of unnecessary conservativeness, hampering policy improvement. This occurs due to the indiscriminate use of all actions from the behavior policy that generates the offline dataset as constraints. The problem becomes particularly noticeable when the quality of the dataset is suboptimal. Thus, we propose Adaptive Advantage-guided Policy Regularization (A2PR), obtaining high-advantage actions from an augmented behavior policy combined with VAE to guide the learned policy. A2PR can select high-advantage actions that differ from those present in the dataset, while still effectively maintaining conservatism from OOD actions. This is achieved by harnessing the VAE capacity to generate samples matching the distribution of the data points. We theoretically prove that the improvement of the behavior policy is guaranteed. Besides, it effectively mitigates value overestimation with a bounded performance gap. Empirically, we conduct a series of experiments on the D4RL benchmark, where A2PR demonstrates state-of-the-art performance. Furthermore, experimental results on additional suboptimal mixed datasets reveal that A2PR exhibits superior performance. Code is available at https://github.com/ltlhuuu/A2PR.

De novo protein design aims to create novel protein structures and sequences unseen in nature. Recent structure-oriented design methods typically employ a two-stage strategy, where structure design and sequence design modules are trained separately, and the backbone structures and sequences are generated sequentially in inference. While diffusion-based generative models like RFdiffusion show great promise in structure design, they face inherent limitations within the two-stage framework. First, the sequence design module risks overfitting, as the accuracy of the generated structures may not align with that of the crystal structures used for training. Second, the sequence design module lacks interaction with the structure design module to further optimize the generated structures. To address these challenges, we propose CarbonNovo, a unified energy-based model for jointly generating protein structure and sequence. Specifically, we leverage a score-based generative model and Markov Random Fields for describing the energy landscape of protein structure and sequence. In CarbonNovo, the structure and sequence design module communicates at each diffusion step, encouraging the generation of more coherent structure-sequence pairs. Moreover, the unified framework allows for incorporating the protein language models as evolutionary constraints for generated proteins. The rigorous evaluation demonstrates that CarbonNovo outperforms two-stage methods across various metrics, including designability, novelty, sequence plausibility, and Rosetta Energy.

Recent advances in reinforcement learning (RL) heavily rely on a variety of well-designed benchmarks, which provide environmental platforms and consistent criteria to evaluate existing and novel algorithms. Specifically, in multi-agent RL (MARL), a plethora of benchmarks based on cooperative games have spurred the development of algorithms that improve the scalability of cooperative multi-agent systems. However, for the competitive setting, a lightweight and open-sourced benchmark with challenging gaming dynamics and visual inputs has not yet been established. In this work, we present FightLadder, a real-time fighting game platform, to empower competitive MARL research. Along with the platform, we provide implementations of state-of-the-art MARL algorithms for competitive games, as well as a set of evaluation metrics to characterize the performance and exploitability of agents. We demonstrate the feasibility of this platform by training a general agent that consistently defeats 12 built-in characters in single-player mode, and expose the difficulty of training a non-exploitable agent without human knowledge and demonstrations in two-player mode. FightLadder provides meticulously designed environments to address critical challenges in competitive MARL research, aiming to catalyze a new era of discovery and advancement in the field. Videos and code at https://sites.google.com/view/fightladder/home.

Diffusion models create data from noise by inverting the forward paths of data towards noise and have emerged as a powerful generative modeling technique for high-dimensional, perceptual data such as images and videos. Rectified flow is a recent generative model formulation that connects data and noise in a straight line. Despite its better theoretical properties and conceptual simplicity, it is not yet decisively established as standard practice. In this work, we improve existing noise sampling techniques for training rectified flow models by biasing them towards perceptually relevant scales. Through a large-scale study, we demonstrate the superior performance of this approach compared to established diffusion formulations for high-resolution text-to-image synthesis. Additionally, we present a novel transformer-based architecture for text-to-image generation that uses separate weights for the two modalities and enables a bidirectional flow of information between image and text tokens, improving text comprehension, typography, and human preference ratings. We demonstrate that this architecture follows predictable scaling trends and correlates lower validation loss to improved text-to-image synthesis as measured by various metrics and human evaluations. Our largest models outperform state-of-the-art models. Stability AI is considering making experimental data, code, and model weights publicly available.

A major risk of using language models in practical applications is their tendency to hallucinate incorrect statements. Hallucinations are often attributed to knowledge gaps in LMs, but we show that LMs sometimes produce hallucinations that they can separately recognize as incorrect. To do this, we construct three question-answering datasets where LMs often state an incorrect answer which is followed by an explanation with at least one incorrect claim. Crucially, we find that GPT-3.5, GPT-4, and LLaMA2-70B-chat can identify 67%, 87%, and 94% of these incorrect claims, respectively. We show that this phenomenon doesn't disappear under higher temperatures sampling, beam search, and zero-shot chain-of-thought prompting. These findings reveal that LM hallucinations can snowball: early mistakes by an LM can lead to more mistakes that otherwise would not be made.

This work explores the feature learning capabilities of deep reinforcement learning algorithms in the pursuit of optimal policy determination. We particularly examine an over-parameterized neural actor-critic framework within the mean-field regime, where both actor and critic components undergo updates via policy gradient and temporal-difference (TD) learning, respectively. We introduce the *mean-field Langevin TD learning* (MFLTD) method, enhancing mean-field Langevin dynamics with proximal TD updates for critic policy evaluation, and assess its performance against conventional approaches through numerical analysis. Additionally, for actor policy updates, we present the *mean-field Langevin policy gradient* (MFLPG), employing policy gradient techniques augmented by Wasserstein gradient flows for parameter space exploration. Our findings demonstrate that MFLTD accurately identifies the true value function, while MFLPG ensures linear convergence of actor sequences towards the globally optimal policy, considering a Kullback-Leibler divergence regularized framework. Through both time particle and discretized analysis, we substantiate the linear convergence guarantees of our neural actor-critic algorithms, representing a notable contribution to neural reinforcement learning focusing on *global optimality* and *feature learning*, extending the existing understanding beyond the conventional scope of lazy training.

Surrogate neural network-based partial differential equation (PDE) solvers have the potential to solve PDEs in an accelerated manner, but they are largely limited to systems featuring fixed domain sizes, geometric layouts, and boundary conditions. We propose Specialized Neural Accelerator-Powered Domain Decomposition Methods (SNAP-DDM), a DDM-based approach to PDE solving in which subdomain problems containing arbitrary boundary conditions and geometric parameters are accurately solved using an ensemble of specialized neural operators. We tailor SNAP-DDM to 2D electromagnetics and fluidic flow problems and show how innovations in network architecture and loss function engineering can produce specialized surrogate subdomain solvers with near unity accuracy. We utilize these solvers with standard DDM algorithms to accurately solve freeform electromagnetics and fluids problems featuring a wide range of domain sizes.

Despite the impressive capabilities of large language models (LLMs) across diverse applications, they still suffer from trustworthiness issues, such as hallucinations and misalignments. Retrieval-augmented language models (RAG) have been proposed to enhance the credibility of generations by grounding external knowledge, but the theoretical understandings of their generation risks remains unexplored. In this paper, we answer: 1) whether RAG can indeed lead to low generation risks, 2) how to provide provable guarantees on the generation risks of RAG and vanilla LLMs, and 3) what sufficient conditions enable RAG models to reduce generation risks. We propose C-RAG, the first framework to certify generation risks for RAG models. Specifically, we provide conformal risk analysis for RAG models and certify an upper confidence bound of generation risks, which we refer to as conformal generation risk. We also provide theoretical guarantees on conformal generation risks for general bounded risk functions under test distribution shifts. We prove that RAG achieves a lower conformal generation risk than that of a single LLM when the quality of the retrieval model and transformer is non-trivial. Our intensive empirical results demonstrate the soundness and tightness of our conformal generation risk guarantees across four widely-used NLP datasets on four state-of-the-art retrieval models.

Federated Learning (FL) enables multiple devices to collaboratively train a shared model while ensuring data privacy. The selection of participating devices in each training round critically affects both the model performance and training efficiency, especially given the vast heterogeneity in training capabilities and data distribution across devices. To deal with these challenges, we introduce a novel device selection solution called FedRank, which is based on an end-to-end, ranking-based model that is pre-trained by imitation learning against state-of-the-art analytical approaches. It not only considers data and system heterogeneity at runtime but also adaptively and efficiently chooses the most suitable clients for model training. Specifically, FedRank views client selection in FL as a ranking problem and employs a pairwise training strategy for the smart selection process. Additionally, an imitation learning-based approach is designed to counteract the cold-start issues often seen in state-of-the-art learning-based approaches. Experimental results reveal that FedRank boosts model accuracy by 5.2% to 56.9%, accelerates the training convergence up to $2.01 \times$ and saves the energy consumption up to 40.1%.

Recent research has made significant progress in designing fusion modules for audio-visual speech separation. However, they predominantly focus on multi-modal fusion at a single temporal scale of auditory and visual features without employing selective attention mechanisms, which is in sharp contrast with the brain. To address this, We propose a novel model called intra- and inter-attention network (IIANet), which leverages the attention mechanism for efficient audio-visual feature fusion. IIANet consists of two types of attention blocks: intra-attention (IntraA) and inter-attention (InterA) blocks, where the InterA blocks are distributed at the top, middle and bottom of IIANet. Heavily inspired by the way how human brain selectively focuses on relevant content at various temporal scales, these blocks maintain the ability to learn modality-specific features and enable the extraction of different semantics from audio-visual features. Comprehensive experiments on three standard audio-visual separation benchmarks (LRS2, LRS3, and VoxCeleb2) demonstrate the effectiveness of IIANet, outperforming previous state-of-the-art methods while maintaining comparable inference time. In particular, the fast version of IIANet (IIANet-fast) has only 7% of CTCNet’s MACs and is 40% faster than CTCNet on CPUs while achieving better separation quality, showing the great potential of attention mechanism for efficient and effective multimodal fusion.

Maximum Inner Product Search (MIPS) is a ubiquitous task in machine learning applications. Given a query vector and $n$ other vectors in $d$ dimensions, the MIPS problem is to find the atom that has the highest inner product with the query vector. Existing MIPS algorithms scale at least as $O(\sqrt{d})$ with respect to $d$, which becomes computationally prohibitive in high-dimensional settings. In this work, we present BanditMIPS, a novel randomized algorithm that provably improves the state-of-the-art complexity from $O(\sqrt{d})$ to $O(1)$ with respect to $d$. We validate the scaling of BanditMIPS and demonstrate that BanditMIPS outperforms prior state-of-the-art MIPS algorithms in sample complexity, wall-clock time, and precision/speedup tradeoff across a variety of experimental settings. Furthermore, we propose a variant of our algorithm, named BanditMIPS-$\alpha$, which improves upon BanditMIPS by employing non-uniform sampling across coordinates. We also demonstrate the usefulness of BanditMIPS in problems for which MIPS is a subroutine, including Matching Pursuit and Fourier analysis. Finally, we demonstrate that BanditMIPS can be used in conjunction with preprocessing techniques to improve its complexity with respect to $n$. All of our experimental results are reproducible via a 1-line script at github.com/ThrunGroup/BanditMIPS.

Current solutions for efficiently constructing large vision-language (VL) models follow a two-step paradigm: projecting the output of pre-trained vision encoders to the input space of pre-trained language models as visual prompts; and then transferring the models to downstream VL tasks via end-to-end parameter-efficient fine-tuning (PEFT). However, this paradigm still exhibits inefficiency since it significantly increases the input length of the language models. In this paper, in contrast to integrating visual prompts into inputs, we regard visual prompts as additional knowledge that facilitates language models in addressing tasks associated with visual information. Motivated by the finding that Feed-Forward Network (FFN) of language models acts as "key-value memory", we introduce a novel approach termed memory-space visual prompting (MemVP), wherein visual prompts are concatenated with the weights of FFN for visual knowledge injection. Experimental results across various VL tasks and language models reveal that MemVP significantly reduces the training time and inference latency of the finetuned VL models and surpasses the performance of previous PEFT methods.

With the advancement of neural networks, diverse methods for neural Granger causality have emerged, which demonstrate proficiency in handling complex data, and nonlinear relationships. However, the existing framework of neural Granger causality has several limitations. It requires the construction of separate predictive models for each target variable, and the relationship depends on the sparsity on the weights of the first layer, resulting in challenges in effectively modeling complex relationships between variables as well as unsatisfied estimation accuracy of Granger causality. Moreover, most of them cannot grasp full-time Granger causality. To address these drawbacks, we propose a **J**acobian **R**egularizer-based **N**eural **G**ranger **C**ausality (**JRNGC**) approach, a straightforward yet highly effective method for learning multivariate summary Granger causality and full-time Granger causality by constructing a single model for all target variables. Specifically, our method eliminates the sparsity constraints of weights by leveraging an input-output Jacobian matrix regularizer, which can be subsequently represented as the weighted causal matrix in the post-hoc analysis. Extensive experiments show that our proposed approach achieves competitive performance with the state-of-the-art methods for learning summary Granger causality and full-time Granger causality while maintaining lower model complexity and high scalability.

Generative models inspired by dynamical transport of measure -- such as flows and diffusions -- construct a continuous-time map between two probability densities. Conventionally, one of these is the target density, only accessible through samples, while the other is taken as a simple base density that is data-agnostic. In this work, using the framework of stochastic interpolants, we formalize how to *couple* the base and the target densities, whereby samples from the base are computed conditionally given samples from the target in a way that is different from (but does not preclude) incorporating information about class labels or continuous embeddings. This enables us to construct dynamical transport maps that serve as conditional generative models. We show that these transport maps can be learned by solving a simple square loss regression problem analogous to the standard independent setting. We demonstrate the usefulness of constructing dependent couplings in practice through experiments in super-resolution and in-painting. The code is available at [https://github.com/interpolants/couplings](https://github.com/interpolants/couplings).

This paper introduces Bespoke Non-Stationary (BNS) Solvers, a solver distillation approach to improve sample efficiency of Diffusion and Flow models. BNS solvers are based on a family of non-stationary solvers that provably subsumes existing numerical ODE solvers and consequently demonstrate considerable improvement in sample approximation (PSNR) over these baselines. Compared to model distillation, BNS solvers benefit from a tiny parameter space ($<$200 parameters), fast optimization (two orders of magnitude faster), maintain diversity of samples, and in contrast to previous solver distillation approaches nearly close the gap from standard distillation methods such as Progressive Distillation in the low-medium NFE regime. For example, BNS solver achieves 45 PSNR / 1.76 FID using 16 NFE in class-conditional ImageNet-64. We experimented with BNS solvers for conditional image generation, text-to-image generation, and text-2-audio generation showing significant improvement in sample approximation (PSNR) in all.

Predicting the future trajectory of surrounding objects is inherently uncertain and vital in the safe and reliable planning of autonomous systems such as in self-driving cars. Although trajectory prediction models have become increasingly sophisticated in dealing with the complexities of spatiotemporal data, the evaluation methods used to assess these models have not kept pace. "Minimum of N" is a common family of metrics used to assess the rich outputs of such models. We critically examine the Minimum of N within the proper scoring rules framework to show that it is not strictly proper and demonstrate how that could lead to a misleading assessment of multimodal trajectory predictions. As an alternative, we propose using Energy Score-based evaluation measures, leveraging their proven propriety for a more reliable evaluation of trajectory distribution predictions.

Textural Inversion, a prompt learning method, learns a singular text embedding for a new "word" to represent image style and appearance, allowing it to be integrated into natural language sentences to generate novel synthesised images. However, identifying multiple unknown object-level concepts within one scene remains a complex challenge. While recent methods have resorted to cropping or masking individual images to learn multiple concepts, these techniques often require prior knowledge of new concepts and are labour-intensive. To address this challenge, we introduce *Multi-Concept Prompt Learning (MCPL)*, where multiple unknown "words" are simultaneously learned from a single sentence-image pair, without any imagery annotations. To enhance the accuracy of word-concept correlation and refine attention mask boundaries, we propose three regularisation techniques: *Attention Masking*, *Prompts Contrastive Loss*, and *Bind Adjective*. Extensive quantitative comparisons with both real-world categories and biomedical images demonstrate that our method can learn new semantically disentangled concepts. Our approach emphasises learning solely from textual embeddings, using less than 10% of the storage space compared to others. The project page, code, and data are available at [https://astrazeneca.github.io/mcpl.github.io](https://astrazeneca.github.io/mcpl.github.io).

Given the black box nature of machine learning models, a plethora of explainability methods have been developed to decipher the factors behind individual decisions. In this paper, we introduce a novel problem of black box (probabilistic) explanation certification. We ask the question: Given a black box model with only query access, an explanation for an example and a quality metric (viz. fidelity, stability), can we find the largest hypercube (i.e., $\ell_{\infty}$ ball) centered at the example such that when the explanation is applied to all examples within the hypercube, (with high probability) a quality criterion is met (viz. fidelity greater than some value)? Being able to efficiently find such a *trust region* has multiple benefits: i) insight into model behavior in a *region*, with a *guarantee*; ii) ascertained *stability* of the explanation; iii) *explanation reuse*, which can save time, energy and money by not having to find explanations for every example; and iv) a possible *meta-metric* to compare explanation methods. Our contributions include formalizing this problem, proposing solutions, providing theoretical guarantees for these solutions that are computable, and experimentally showing their efficacy on synthetic and real data.

Interpretable graph neural networks (XGNNs ) are widely adopted in various scientific applications involving graph-structured data. Existing XGNNs predominantly adopt the attention-based mechanism to learn edge or node importance for extracting and making predictions with the interpretable subgraph. However, the representational properties and limitations of these methods remain inadequately explored. In this work, we present a theoretical framework that formulates interpretable subgraph learning with the multilinear extension of the subgraph distribution, coined as subgraph multilinear extension (SubMT). Extracting the desired interpretable subgraph requires an accurate approximation of SubMT, yet we find that the existing XGNNs can have a huge gap in fitting SubMT. Consequently, the SubMT approximation failure will lead to the degenerated interpretability of the extracted subgraphs. To mitigate the issue, we design a new XGNN architecture called Graph Multilinear neT (GMT), which is provably more powerful in approximating SubMT. We empirically validate our theoretical findings on a number of graph classification benchmarks. The results demonstrate that GMT outperforms the state-of-the-art up to 10% in terms of both interpretability and generalizability across 12 regular and geometric graph benchmarks.