论文检索

Reinforcement Learning · Inverse

Thomas Kleine Buening, Victor Villin, Christos Dimitrakakis

Learning a reward function from demonstrations suffers from low sample-efficiency. Even with abundant data, current inverse reinforcement learning methods that focus on learning from a single environment can fail to handle slight changes in the environment dynamics. We tackle these challenges through adaptive environment design. In our framework, the learner repeatedly interacts with the expert, with the former selecting environments to identify the reward function as quickly as possible from the expert’s demonstrations in said environments. This results in improvements in both sample-efficiency and robustness, as we show experimentally, for both exact and approximate inference.

Sucheng Ren, Zeyu Wang, Hongru Zhu, Junfei Xiao, Alan Yuille, Cihang Xie

This paper enhances image-GPT (iGPT), one of the pioneering works that introduce autoregressive pretraining to predict the next pixels for visual representation learning. Two simple yet essential changes are made. First, we shift the prediction target from raw pixels to semantic tokens, enabling a higher-level understanding of visual content. Second, we supplement the autoregressive modeling by instructing the model to predict not only the next tokens but also the visible tokens. This pipeline is particularly effective when semantic tokens are encoded by discriminatively trained models, such as CLIP. We introduce this novel approach as D-iGPT. Extensive experiments showcase that D-iGPT excels as a strong learner of visual representations: A notable achievement is its compelling performance on the ImageNet-1K dataset --- by training on publicly available datasets, D-iGPT unprecedentedly achieves **90.0%** top-1 accuracy with a vanilla ViT-H. Additionally, D-iGPT shows strong generalization on the downstream task. Code is available at https://github.com/OliverRensu/D-iGPT.

Deep Learning · Generative Models and Autoencoders

Katherine Crowson, Stefan Baumann, Alex Birch, Tanishq Abraham, Daniel Kaplan, Enrico Shippole

We present the Hourglass Diffusion Transformer (HDiT), an image-generative model that exhibits linear scaling with pixel count, supporting training at high resolution (e.g. $1024 \times 1024$) directly in pixel-space. Building on the Transformer architecture, which is known to scale to billions of parameters, it bridges the gap between the efficiency of convolutional U-Nets and the scalability of Transformers. HDiT trains successfully without typical high-resolution training techniques such as multiscale architectures, latent autoencoders or self-conditioning. We demonstrate that HDiT performs competitively with existing models on ImageNet $256^2$, and sets a new state-of-the-art for diffusion models on FFHQ-$1024^2$. Code is available at https://github.com/crowsonkb/k-diffusion.

Deep Learning · Graph Neural Networks

Caterina Graziani, Tamara Drucks, Fabian Jogl, Monica Bianchini, franco scarselli, Thomas Gärtner

We systematically investigate the expressive power of path-based graph neural networks. While it has been shown that path-based graph neural networks can achieve strong empirical results, an investigation into their expressive power is lacking. Therefore, we propose PATH-WL, a general class of color refinement algorithms based on paths and shortest path distance information. We show that PATH-WL is incomparable to a wide range of expressive graph neural networks, can count cycles, and achieves strong empirical results on the notoriously difficult family of strongly regular graphs. Our theoretical results indicate that PATH-WL forms a new hierarchy of highly expressive graph neural networks.

Applications · Computer Vision

Nathaniel Cohen, Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, Tomer Michaeli

Text-to-image (T2I) diffusion models achieve state-of-the-art results in image synthesis and editing. However, leveraging such pre-trained models for video editing is considered a major challenge. Many existing works attempt to enforce temporal consistency in the edited video through explicit correspondence mechanisms, either in pixel space or between deep features. These methods, however, struggle with strong nonrigid motion. In this paper, we introduce a fundamentally different approach, which is based on the observation that spatiotemporal slices of natural videos exhibit similar characteristics to natural images. Thus, the same T2I diffusion model that is normally used only as a prior on video frames, can also serve as a strong prior for enhancing temporal consistency by applying it on spatiotemporal slices. Based on this observation, we present Slicedit, a method for text-based video editing that utilizes a pre-trained T2I diffusion model to process both spatial and spatiotemporal slices. Our method generates videos that retain the structure and motion of the original video while adhering to the target text. Through extensive experiments, we demonstrate Slicedit's ability to edit a wide range of real-world videos, confirming its clear advantages compared to existing baselines.

Mohammad Pezeshki, Diane Bouchacourt, Mark Ibrahim, Nicolas Ballas, Pascal Vincent, David Lopez-Paz

Environment annotations are essential for the success of many out-of-distribution (OOD) generalization methods. Unfortunately, these are costly to obtain and often limited by human annotators' biases. To achieve robust generalization, it is essential to develop algorithms for automatic environment discovery within datasets. Current proposals, which divide examples based on their training error, suffer from one fundamental problem. These methods introduce hyper-parameters and early-stopping criteria, which require a validation set with human-annotated environments, the very information subject to discovery. In this paper, we propose Cross-Risk Minimization (XRM) to address this issue. XRM trains twin networks, each learning from one random half of the training data, while imitating confident held-out mistakes made by its sibling. XRM provides a recipe for hyper-parameter tuning, does not require early-stopping, and can discover environments for all training and validation data. Algorithms built on top of XRM environments achieve oracle worst-group-accuracy, addressing a long-standing challenge in OOD generalization.

Theory · Active Learning and Interactive Learning

Ben Adcock, Juan Cardenas, Nick Dexter

This work considers the fundamental problem of learning an unknown object from training data using a given model class. We introduce a framework that allows for objects in arbitrary Hilbert spaces, general types of (random) linear measurements as training data and general types of nonlinear model classes. We establish a series of learning guarantees for this framework, which provide explicit relations between the amount of training data and the model class to ensure near-best generalization bounds. In doing so, we introduce the key notion of the *variation* of a model class with respect to a distribution of sampling operators. We show that this framework can accommodate many different types of well-known problems of interest, such as matrix sketching by random sampling, compressed sensing with isotropic vectors, active learning in regression and compressed sensing with generative models. In all cases, known results become straightforward corollaries of our general theory. Hence, this work provides a powerful framework for studying and analyzing many different types of learning problems.

Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, Wynne Hsu

Existing research of video understanding still struggles to achieve in-depth comprehension and reasoning in complex videos, primarily due to the under-exploration of two key bottlenecks: fine-grained spatial-temporal perceptive understanding and cognitive-level video scene comprehension. This paper bridges the gap by presenting a novel solution. We first introduce a novel video Multimodal Large Language Model (MLLM), MotionEpic, which achieves fine-grained pixel-level spatial-temporal video grounding by integrating video spatial-temporal scene graph (STSG) representation. Building upon MotionEpic, we then develop a Video-of-Thought (VoT) reasoning framework. VoT inherits the Chain-of-Thought (CoT) core, breaking down a complex task into simpler and manageable sub-problems, and addressing them step-by-step from a low-level pixel perception to high-level cognitive interpretation. Extensive experiments across various complex video QA benchmarks demonstrate that our overall framework strikingly boosts existing state-of-the-art. To our knowledge, this is the first attempt at successfully implementing the CoT technique for achieving human-level video reasoning, where we show great potential in extending it to a wider range of video understanding scenarios. Systems and codes will be open later.

Deep Learning · Large Language Models

Neil Band, Xuechen Li, Tengyu Ma, Tatsunori Hashimoto

Language models (LMs) may lead their users to make suboptimal downstream decisions when they confidently hallucinate. This issue can be mitigated by having the LM verbally convey the probability that its claims are correct, but existing models cannot produce long-form text with calibrated confidence statements. Through the lens of decision-making, we define linguistic calibration for long-form generations: an LM is linguistically calibrated if its generations enable its users to make calibrated probabilistic predictions. This definition enables a training framework where a supervised finetuning step bootstraps an LM to emit long-form generations with confidence statements such as "I estimate a 30% chance of..." or "I am certain that...", followed by a reinforcement learning step which rewards generations that enable a user to provide calibrated answers to related questions. We linguistically calibrate Llama 2 7B and find in automated and human evaluations of long-form generations that it is significantly more calibrated than strong finetuned factuality baselines with comparable accuracy. These findings generalize under significant domain shifts to scientific and biomedical questions and to an entirely held-out person biography generation task. Our results demonstrate that long-form generations may be calibrated end-to-end by constructing an objective in the space of the predictions that users make in downstream decision-making.

Deep Learning · Attention Mechanisms

Piotr Nawrot, Adrian Łańcucki, Marcin Chochowski, David Tarjan, Edoardo Ponti

Transformers have emerged as the backbone of large language models (LLMs). However, generation remains inefficient due to the need to store in memory a cache of key–value representations for past tokens, whose size scales linearly with the input sequence length and batch size. As a solution, we propose Dynamic Memory Compression (DMC), a method for on-line key–value cache compression at inference time. Most importantly, the model learns to apply different compression ratios in different heads and layers. We retrofit pre-trained LLMs such as Llama 2 (7B, 13B and 70B) into DMC Transformers, achieving up to $\sim 3.7 \times$ throughput increase during auto-regressive inference on an NVIDIA H100 GPU. DMC is applied via continued pre-training on a negligible percentage of the original data without adding any extra parameters. We find that DMC preserves the original downstream performance with up to 4$\times$ cache compression, outperforming up-trained grouped-query attention (GQA) and key–value eviction policies (H$_2$O, TOVA). GQA and DMC can be even combined to obtain compounded gains. As a result DMC fits longer contexts and larger batches within any given memory budget. We release the DMC code and models at https://github.com/NVIDIA/Megatron-LM/tree/DMC.

Miscellaneous Aspects of Machine Learning · Causality

Inwoo Hwang, Yesong Choe, Yeahoon Kwon, Sanghack Lee

Identifying and estimating a causal effect is a fundamental task when researchers want to infer a causal effect using an observational study without experiments. A conventional assumption is the strict positivity of the given distribution, or so called positivity (or overlap) under the unconfounded assumption that the probabilities of treatments are positive. However, there exist many environments where neither observational data exhibits strict positivity nor unconfounded assumption holds. Against this background, we examine the graphical counterpart of the conventional positivity condition so as to license the use of identification formula without strict positivity. In particular, we explore various approaches, including analysis in a post-hoc manner, do-calculus, $Q$-decomposition, and algorithmic, to yielding a positivity condition for an identification formula, where we relate them, providing a comprehensive view. We further discuss the design of a positivity-aware identification algorithm based on the theoretical characterization of identification formulas.

Reinforcement Learning · Deep RL

Matthias Weissenbacher, Rishabh Agarwal, Yoshinobu Kawahara

An open challenge in reinforcement learning (RL) is the effective deployment of a trained policy to new or slightly different situations as well as semantically-similar environments. We introduce **S**ymmetry-**I**nvariant **T**ransformer (**SiT**), a scalable vision transformer (ViT) that leverages both local and global data patterns in a self-supervised manner to improve generalisation. Central to our approach is Graph Symmetric Attention, which refines the traditional self-attention mechanism to preserve graph symmetries, resulting in invariant and equivariant latent representations. We showcase SiT's superior generalization over ViTs on MiniGrid and Procgen RL benchmarks, and its sample efficiency on Atari 100k and CIFAR10.

Optimization · Zero-order and Black-box Optimization

Wenjie Xu, Wenbin Wang, Yuning Jiang, Bratislav Svetozarevic, Colin Jones

We study the problem of preferential Bayesian optimization (BO), where we aim to optimize a black-box function with only preference feedback over a pair of candidate solutions. Inspired by the likelihood ratio idea, we construct a confidence set of the black-box function using only the preference feedback. An optimistic algorithm with an efficient computational method is then developed to solve the problem, which enjoys an information-theoretic bound on the total cumulative regret, a first-of-its-kind for preferential BO. This bound further allows us to design a scheme to report an estimated best solution, with a guaranteed convergence rate. Experimental results on sampled instances from Gaussian processes, standard test functions, and a thermal comfort optimization problem all show that our method stably achieves better or competitive performance as compared to the existing state-of-the-art heuristics, which, however, do not have theoretical guarantees on regret bounds or convergence.

Uri Stemmer

Private everlasting prediction (PEP), recently introduced by Naor et al. [2023], is a model for differentially private learning in which the learner never publicly releases a hypothesis. Instead, it provides black-box access to a "prediction oracle" that can predict the labels of an *endless stream* of unlabeled examples drawn from the underlying distribution. Importantly, PEP provides privacy both for the initial training set and for the endless stream of classification queries. We present two conceptual modifications to the definition of PEP, as well as new constructions exhibiting significant improvements over prior work. Specifically, we incorporate robustness against poisoning attacks into the definition of PEP; we present a relaxed privacy definition, suitable for PEP, that allows us to disconnect the privacy parameter $\delta$ from the number of total time steps $T$; and we present new constructions for axis-aligned rectangles and decision-stumps exhibiting improved sample complexity and runtime.

Deep Learning · Attention Mechanisms

Simran Arora, Sabri Eyuboglu, Michael Zhang, Aman Timalsina, Silas Alberti, James Zou, Atri Rudra, Christopher Re

Recent work has shown that attention-based language models excel at "recall", the ability to ground generations in tokens previously seen in context. However, the efficiency of attention-based models is bottle-necked during inference by the KV-cache's aggressive memory consumption. In this work, we explore whether we can improve language model efficiency (e.g. by reducing memory consumption) without compromising on recall. By applying experiments and theory to a broad set of architectures, we identify a key tradeoff between a model's recurrent state size and recall ability. We show that efficient alternatives to attention (e.g. H3, Mamba, RWKV) maintain a fixed-size recurrent state, but struggle at recall. We propose BASED a simple architecture combining linear and sliding window attention. By varying BASED window size and linear attention feature dimension, we can dial the state size and traverse the Pareto frontier of the recall-memory tradeoff curve, recovering the full quality of attention on one end and the small state size of attention-alternatives on the other. We train language models up to $1.3$b parameters and show that BASED matches the strongest sub-quadratic models (e.g. Mamba) in perplexity and outperforms them on real-world recall-intensive tasks by 10.36 accuracy points. We further develop IO-aware algorithms that enable BASED to provide 24× higher throughput on language generation than FlashAttention-2, when generating 1024 tokens using 1.3b parameter models. Overall, BASED expands the Pareto frontier of the throughput-recall tradeoff space beyond prior architectures.

Reinforcement Learning

Younghyo Park, Gabriel Margolis, Pulkit Agrawal

Many roboticists dream of presenting a robot with a task in the evening and returning the next morning to find the robot capable of solving the task. What is preventing us from achieving this? Sim-to-real reinforcement learning (RL) has achieved impressive performance on challenging robotics tasks, but requires substantial human effort to set up the task in a way that is amenable to RL. It's our position that algorithmic improvements in policy optimization and other ideas should be guided towards resolving the primary bottleneck of shaping the training environment, i.e., designing observations, actions, rewards and simulation dynamics. Most practitioners don't tune the RL algorithm, but other environment parameters to obtain a desirable controller. We posit that scaling RL to diverse robotic tasks will only be achieved if the community focuses on automating environment shaping procedures.

Collin Burns, Pavel Izmailov, Jan Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike 等

Widely used alignment techniques, such as reinforcement learning from human feedback (RLHF), rely on the ability of humans to supervise model behavior---for example, to evaluate whether a model faithfully followed instructions or generated safe outputs. However, future superhuman models will behave in complex ways too difficult for humans to reliably evaluate; humans will only be able to *weakly supervise* superhuman models. We study an analogy to this problem: can weak model supervision elicit the full capabilities of a much stronger model? We test this using a range of pretrained language models in the GPT-4 family on natural language processing (NLP), chess, and reward modeling tasks. We find that when we naively finetune strong pretrained models on labels generated by a weak model, they consistently perform better than their weak supervisors, a phenomenon we call *weak-to-strong generalization*. However, we are still far from recovering the full capabilities of strong models with naive finetuning alone, suggesting that techniques like RLHF may scale poorly to superhuman models without further work. We find that simple methods can often significantly improve weak-to-strong generalization: for example, when finetuning GPT-4 with a GPT-2-level supervisor and an auxiliary confidence loss, we can recover close to GPT-3.5-level performance on NLP tasks. Our results suggest that it is feasible to make empirical progress today on a fundamental challenge of aligning superhuman models.

Guanjie Chen, Xinyu Zhao, Tianlong Chen, Yu Cheng

Mixture-of-Experts (MoE) has gained increasing popularity as a promising framework for scaling up large language models (LLMs). However, the reliability assessment of MoE lags behind its surging applications. Moreover, when transferred to new domains such as in fine-tuning MoE models sometimes underperform their dense counterparts. Motivated by the research gap and counter-intuitive phenomenon, we propose $\texttt{MoE-RBench}$, the first comprehensive assessment of SMoE reliability from three aspects: $\textit{(i)}$ safety and hallucination, $\textit{(ii)}$ resilience to adversarial attacks, and $\textit{(iii)}$ out-of-distribution robustness. Extensive models and datasets are tested to compare the MoE to dense networks from these reliability dimensions. Our empirical observations suggest that with appropriate hyperparameters, training recipes, and inference techniques, we can build the MoE model more reliably than the dense LLM. In particular, we find that the robustness of SMoE is sensitive to the basic training settings. We hope that this study can provide deeper insights into how to adapt the pre-trained MoE model to other tasks with higher-generation security, quality, and stability. Codes are available at https://github.com/UNITES-Lab/MoE-RBench.

Theory · Active Learning and Interactive Learning

Shahrzad Haddadan, Cheng Xin, Jie Gao

We consider a cooperative learning scenario where a collection of networked agents with individually owned classifiers dynamically update their predictions, for the same classification task, through communication or observations of each other’s predictions. Clearly if highly influential vertices use erroneous classifiers, there will be a negative effect on the accuracy of all the agents in the network. We ask the following question: how can we optimally fix the prediction of a few classifiers so as maximize the overall accuracy in the entire network. To this end we consider an aggregate and an egalitarian objective function. We show a polynomial time algorithm for optimizing the aggregate objective function, and show that optimizing the egalitarian objective function is NP-hard. Furthermore, we develop approximation algorithms for the egalitarian improvement. The performance of all of our algorithms are guaranteed by mathematical analysis and backed by experiments on synthetic and real data.

Deep Learning · Generative Models and Autoencoders

Giannis Daras, Alexandros Dimakis, Constantinos Daskalakis

Ambient diffusion is a recently proposed framework for training diffusion models using corrupted data. Both Ambient Diffusion and alternative SURE-based approaches for learning diffusion models from corrupted data resort to approximations which deteriorate performance. We present the first framework for training diffusion models that provably sample from the uncorrupted distribution given only noisy training data, solving an open problem in Ambient diffusion. Our key technical contribution is a method that uses a double application of Tweedie's formula and a consistency loss function that allows us to extend sampling at noise levels below the observed data noise. We also provide further evidence that diffusion models memorize from their training sets by identifying extremely corrupted images that are almost perfectly reconstructed, raising copyright and privacy concerns. Our method for training using corrupted samples can be used to mitigate this problem. We demonstrate this by fine-tuning Stable Diffusion XL to generate samples from a distribution using only noisy samples. Our framework reduces the amount of memorization of the fine-tuning dataset, while maintaining competitive performance.