论文检索

Deep Learning · Theory

Hugo Cui, Luca Pesce, Yatin Dandi, FLORENT KRZAKALA, Yue Lu, Lenka Zdeborova, Bruno Loureiro

In this manuscript, we investigate the problem of how two-layer neural networks learn features from data, and improve over the kernel regime, after being trained with a single gradient descent step. Leveraging the insight from (Ba et al., 2022), we model the trained network by a spiked Random Features (sRF) model. Further building on recent progress on Gaussian universality (Dandi et al., 2023), we provide an exact asymptotic description of the generalization error of the sRF in the high-dimensional limit where the number of samples, the width, and the input dimension grow at a proportional rate. The resulting characterization for sRFs also captures closely the learning curves of the original network model. This enables us to understand how adapting to the data is crucial for the network to efficiently learn non-linear functions in the direction of the gradient - where at initialization it can only express linear functions in this regime.

Theory

Luca Arnaboldi, Yatin Dandi, FLORENT KRZAKALA, Bruno Loureiro, Luca Pesce, Ludovic Stephan

We study the impact of the batch size $n_b$ on the iteration time $T$ of training two-layer neural networks with one-pass stochastic gradient descent (SGD) on multi-index target functions of isotropic covariates. We characterize the optimal batch size minimizing the iteration time as a function of the hardness of the target, as characterized by the information exponents. We show that performing gradient updates with large batches $n_b \lesssim d^{\frac{\ell}{2}}$ minimizes the training time without changing the total sample complexity, where $\ell$ is the information exponent of the target to be learned and $d$ is the input dimension. However, larger batch sizes than $n_b \gg d^{\frac{\ell}{2}}$ are detrimental for improving the time complexity of SGD. We provably overcome this fundamental limitation via a different training protocol, *Correlation loss SGD*, which suppresses the auto-correlation terms in the loss function. We show that one can track the training progress by a system of low-dimensional ordinary differential equations (ODEs). Finally, we validate our theoretical results with numerical experiments.

Deep Learning · Theory

Yatin Dandi, Emanuele Troiani, Luca Arnaboldi, Luca Pesce, Lenka Zdeborova, FLORENT KRZAKALA

We investigate the training dynamics of two-layer neural networks when learning multi-index target functions. We focus on multi-pass gradient descent (GD) that reuses the batches multiple times and show that it significantly changes the conclusion about which functions are learnable compared to single-pass gradient descent. In particular, multi-pass GD with finite stepsize is found to overcome the limitations of gradient flow and single-pass GD given by the information exponent (Ben Arous et al., 2021) and leap exponent (Abbe et al., 2023) of the target function. We show that upon re-using batches, the network achieves in just two time steps an overlap with the target subspace even for functions not satisfying the staircase property (Abbe et al., 2021). We characterize the (broad) class of functions efficiently learned in finite time. The proof of our results is based on the analysis of the Dynamical Mean-Field Theory (DMFT). We further provide a closed-form description of the dynamical process of the low-dimensional projections of the weights, and numerical experiments illustrating the theory.

Social Aspects · Trustworthy Machine Learning

Harshay Shah, Andrew Ilyas, Aleksander Madry

*How does the internal computation of a machine learning model transform inputs into predictions?* To tackle this question, we introduce a framework called *component modeling* for decomposing a model prediction in terms of its components---architectural "building blocks" such as convolution filters or attention heads. We focus on a special case of this framework, *component attribution*, where the goal is to estimate the counterfactual impact of individual components on a given prediction. We then present COAR, a scalable algorithm for estimating component attributions, and demonstrate its effectiveness across models, datasets and modalities. Finally, we show that COAR directly enables effective model editing. Our code is available at [github.com/MadryLab/modelcomponents]([https://github.com/MadryLab/modelcomponents]).

Deep Learning · Graph Neural Networks

Dominik Fuchsgruber, Tom Wollschläger, Bertrand Charpentier, Antonio Oroz, Stephan Günnemann

Uncertainty Sampling is an Active Learning strategy that aims to improve the data efficiency of machine learning models by iteratively acquiring labels of data points with the highest uncertainty. While it has proven effective for independent data its applicability to graphs remains under-explored. We propose the first extensive study of Uncertainty Sampling for node classification: **(1)** We benchmark Uncertainty Sampling beyond predictive uncertainty and highlight a significant performance gap to other Active Learning strategies. **(2)** We develop ground-truth Bayesian uncertainty estimates in terms of the data generating process and prove their effectiveness in guiding Uncertainty Sampling toward optimal queries. We confirm our results on synthetic data and design an approximate approach that consistently outperforms other uncertainty estimators on real datasets. **(3)** Based on this analysis, we relate pitfalls in modeling uncertainty to existing methods. Our analysis enables and informs the development of principled uncertainty estimation on graphs.

Applications

Xiao Luo, Yiyang Gu, Huiyu Jiang, Hang Zhou, Jinsheng Huang, Wei Ju, Zhiping Xiao, Ming Zhang, Yizhou Sun

This paper studies the problem of modeling multi-agent dynamical systems, where agents could interact mutually to influence their behaviors. Recent research predominantly uses geometric graphs to depict these mutual interactions, which are then captured by powerful graph neural networks (GNNs). However, predicting interacting dynamics in challenging scenarios such as out-of-distribution shift and complicated underlying rules remains unsolved. In this paper, we propose a new approach named Prototypical Graph ODE (PGODE) to address the problem. The core of PGODE is to incorporate prototype decomposition from contextual knowledge into a continuous graph ODE framework. Specifically, PGODE employs representation disentanglement and system parameters to extract both object-level and system-level contexts from historical trajectories, which allows us to explicitly model their independent influence and thus enhances the generalization capability under system changes. Then, we integrate these disentangled latent representations into a graph ODE model, which determines a combination of various interacting prototypes for enhanced model expressivity. The entire model is optimized using an end-to-end variational inference framework to maximize the likelihood. Extensive experiments in both in-distribution and out-of-distribution settings validate the superiority of PGODE compared to various baselines.

Theory · Online Learning and Bandits

Andrew Jacobsen, Ashok Cutkosky

We develop algorithms for online linear regression which achieve optimal static and dynamic regret guarantees *even in the complete absence of prior knowledge*. We present a novel analysis showing that a discounted variant of the Vovk-Azoury-Warmuth forecaster achieves dynamic regret of the form $R_{T}(\vec{u})\le O\Big(d\log(T)\vee \sqrt{dP_{T}^{\gamma}(\vec{u})T}\Big)$, where $P_{T}^{\gamma}(\vec{u})$ is a measure of variability of the comparator sequence, and show that the discount factor achieving this result can be learned on-the-fly. We show that this result is optimal by providing a matching lower bound. We also extend our results to *strongly-adaptive* guarantees which hold over every sub-interval $[a,b]\subseteq[1,T]$ simultaneously.

Deep Learning · Attention Mechanisms

Muhammed Emrullah Ildiz, Yixiao HUANG, Yingcong Li, Ankit Singh Rawat, Samet Oymak

Modern language models rely on the transformer architecture and attention mechanism to perform language understanding and text generation. In this work, we study learning a 1-layer self-attention model from a set of prompts and the associated outputs sampled from the model. We first establish a formal link between the self-attention mechanism and Markov models under suitable conditions: Inputting a prompt to the self-attention model samples the output token according to a *context-conditioned Markov chain* (CCMC). *CCMC* is obtained by weighing the transition matrix of a standard Markov chain according to the sufficient statistics of the prompt/context. Building on this formalism, we develop identifiability/coverage conditions for the data distribution that guarantee consistent estimation of the latent model under a teacher-student setting and establish sample complexity guarantees under IID data. Finally, we study the problem of learning from a single output trajectory generated in response to an initial prompt. We characterize a *winner-takes-all* phenomenon where the generative process of self-attention evolves to sampling from a small set of *winner tokens* that dominate the context window. This provides a mathematical explanation to the tendency of modern LLMs to generate repetitive text.

Deep Learning · Graph Neural Networks

Wei Ju, Zhengyang Mao, Siyu Yi, Yifang Qin, Yiyang Gu, Zhiping Xiao, Yifan Wang, Xiao Luo, Ming Zhang

In this paper, we study semi-supervised graph classification, which aims at accurately predicting the categories of graphs in scenarios with limited labeled graphs and abundant unlabeled graphs. Despite the promising capability of graph neural networks (GNNs), they typically require a large number of costly labeled graphs, while a wealth of unlabeled graphs fail to be effectively utilized. Moreover, GNNs are inherently limited to encoding local neighborhood information using message-passing mechanisms, thus lacking the ability to model higher-order dependencies among nodes. To tackle these challenges, we propose a Hypergraph-Enhanced DuAL framework named HEAL for semi-supervised graph classification, which captures graph semantics from the perspective of the hypergraph and the line graph, respectively. Specifically, to better explore the higher-order relationships among nodes, we design a hypergraph structure learning to adaptively learn complex node dependencies beyond pairwise relations. Meanwhile, based on the learned hypergraph, we introduce a line graph to capture the interaction between hyperedges, thereby better mining the underlying semantic structures. Finally, we develop a relational consistency learning to facilitate knowledge transfer between the two branches and provide better mutual guidance. Extensive experiments on real-world graph datasets verify the effectiveness of the proposed method against existing state-of-the-art methods.

Deep Learning · Graph Neural Networks

MoonJeong Park, Jaeseung Heo, Dongwoo Kim

Graph Neural Network (GNN) resembles the diffusion process, leading to the over-smoothing of learned representations when stacking many layers. Hence, the reverse process of message passing can produce the distinguishable node representations by inverting the forward message propagation. The distinguishable representations can help us to better classify neighboring nodes with different labels, such as in heterophilic graphs. In this work, we apply the design principle of the reverse process to the three variants of the GNNs. Through the experiments on heterophilic graph data, where adjacent nodes need to have different representations for successful classification, we show that the reverse process significantly improves the prediction performance in many cases. Additional analysis reveals that the reverse mechanism can mitigate the over-smoothing over hundreds of layers. Our code is available at https://github.com/ml-postech/reverse-gnn.

Miscellaneous Aspects of Machine Learning · Causality

Jonas Schweisthal, Dennis Frauen, M van der Schaar, Stefan Feuerriegel

Estimating the conditional average treatment effect (CATE) from observational data is relevant for many applications such as personalized medicine. Here, we focus on the widespread setting where the observational data come from multiple environments, such as different hospitals, physicians, or countries. Furthermore, we allow for violations of standard causal assumptions, namely, overlap within the environments and unconfoundedness. To this end, we move away from point identification and focus on partial identification. Specifically, we show that current assumptions from the literature on multiple environments allow us to interpret the environment as an instrumental variable (IV). This allows us to adapt bounds from the IV literature for partial identification of CATE by leveraging treatment assignment mechanisms across environments. Then, we propose different model-agnostic learners (so-called meta-learners) to estimate the bounds that can be used in combination with arbitrary machine learning models. We further demonstrate the effectiveness of our meta-learners across various experiments using both simulated and real-world data. Finally, we discuss the applicability of our meta-learners to partial identification in instrumental variable settings, such as randomized controlled trials with non-compliance.

Jost Tobias Springenberg, Abbas Abdolmaleki, Jingwei Zhang, Oliver M Groth, Michael Bloesch, Thomas Lampe, Philemon Brakel, Sarah Bechtle, Steven Kapturowski, Roland Hafner 等

We show that offline actor-critic reinforcement learning can scale to large models - such as transformers - and follows similar scaling laws as supervised learning. We find that offline actor-critic algorithms can outperform strong, supervised, behavioral cloning baselines for multi-task training on a large dataset; containing both sub-optimal and expert behavior on 132 continuous control tasks. We introduce a Perceiver-based actor-critic model and elucidate the key features needed to make offline RL work with self- and cross-attention modules. Overall, we find that: i) simple offline actor critic algorithms are a natural choice for gradually moving away from the currently predominant paradigm of behavioral cloning, and ii) via offline RL it is possible to learn multi-task policies that master many domains simultaneously, including real robotics tasks, from sub-optimal demonstrations or self-generated data.

Jost Tobias Springenberg, Abbas Abdolmaleki, Jingwei Zhang, Oliver M Groth, Michael Bloesch, Thomas Lampe, Philemon Brakel, Sarah Bechtle, Steven Kapturowski, Roland Hafner 等

We show that offline actor-critic reinforcement learning can scale to large models - such as transformers - and follows similar scaling laws as supervised learning. We find that offline actor-critic algorithms can outperform strong, supervised, behavioral cloning baselines for multi-task training on a large dataset; containing both sub-optimal and expert behavior on 132 continuous control tasks. We introduce a Perceiver-based actor-critic model and elucidate the key features needed to make offline RL work with self- and cross-attention modules. Overall, we find that: i) simple offline actor critic algorithms are a natural choice for gradually moving away from the currently predominant paradigm of behavioral cloning, and ii) via offline RL it is possible to learn multi-task policies that master many domains simultaneously, including real robotics tasks, from sub-optimal demonstrations or self-generated data.

Miscellaneous Aspects of Machine Learning · General Machine Learning Techniques

Jongha (Jon) Ryu, Xiangxiang Xu, Hasan Sabri Melihcan Erol, Yuheng Bu, Lizhong Zheng, Gregory Wornell

Computing eigenvalue decomposition (EVD) of a given linear operator, or finding its leading eigenvalues and eigenfunctions, is a fundamental task in many machine learning and scientific simulation problems. For high-dimensional eigenvalue problems, training neural networks to parameterize the eigenfunctions is considered as a promising alternative to the classical numerical linear algebra techniques. This paper proposes a new optimization framework based on the low-rank approximation characterization of a truncated singular value decomposition, accompanied by new techniques called *nesting* for learning the top-$L$ singular values and singular functions in the correct order. The proposed method promotes the desired orthogonality in the learned functions implicitly and efficiently via an unconstrained optimization formulation, which is easy to solve with off-the-shelf gradient-based optimization algorithms. We demonstrate the effectiveness of the proposed optimization framework for use cases in computational physics and machine learning.

Theory · Online Learning and Bandits

Jongha (Jon) Ryu, Gregory Wornell

A confidence sequence (CS) is a sequence of confidence sets that contains a target parameter of an underlying stochastic process at any time step with high probability. This paper proposes a new approach to constructing CSs for means of bounded multivariate stochastic processes using a general gambling framework, extending the recently established coin toss framework for bounded random processes. The proposed gambling framework provides a general recipe for constructing CSs for categorical and probability-vector-valued observations, as well as for general bounded multidimensional observations through a simple reduction. This paper specifically explores the use of the mixture portfolio, akin to Cover's universal portfolio, in the proposed framework and investigates the properties of the resulting CSs. Simulations demonstrate the tightness of these confidence sequences compared to existing methods. When applied to the sampling without-replacement setting for finite categorical data, it is shown that the resulting CS based on a universal gambling strategy is provably tighter than that of the posterior-prior ratio martingale proposed by Waudby-Smith and Ramdas.

Deep Learning · Robustness

Aaron Havens, Alexandre Araujo, Huan Zhang, Bin Hu

Self-attention has been widely used in various machine learning models, such as vision transformers. The standard dot-product self-attention is arguably the most popular structure, and there is a growing interest in understanding the mathematical properties of such attention mechanisms. This paper presents a fine-grained local sensitivity analysis of the standard dot-product self-attention, leading to new non-vacuous certified robustness results for vision transformers. Despite the well-known fact that dot-product self-attention is not (globally) Lipschitz, we develop new theoretical analysis of Local Fine-grained Attention Sensitivity (LoFAST) quantifying the effect of input feature perturbations on the attention output. Our analysis reveals that the local sensitivity of dot-product self-attention to $\ell_2$ perturbations can actually be controlled by several key quantities associated with the attention weight matrices and the unperturbed input. We empirically validate our theoretical findings by computing non-vacuous certified $\ell_2$-robustness for vision transformers on CIFAR-10 and SVHN datasets. The code for LoFAST is available at https://github.com/AaronHavens/LoFAST.

Social Aspects · Trustworthy Machine Learning

Zhi-Yi Chin, Chieh Ming Jiang, Ching-Chun Huang, Pin-Yu Chen, Wei-Chen Chiu

Text-to-image diffusion models, e.g. Stable Diffusion (SD), lately have shown remarkable ability in high-quality content generation, and become one of the representatives for the recent wave of transformative AI. Nevertheless, such advance comes with an intensifying concern about the misuse of this generative technology, especially for producing copyrighted or NSFW (i.e. not safe for work) images. Although efforts have been made to filter inappropriate images/prompts or remove undesirable concepts/styles via model fine-tuning, the reliability of these safety mechanisms against diversified problematic prompts remains largely unexplored. In this work, we propose **Prompting4Debugging (P4D)** as a debugging and red-teaming tool that automatically finds problematic prompts for diffusion models to test the reliability of a deployed safety mechanism. We demonstrate the efficacy of our P4D tool in uncovering new vulnerabilities of SD models with safety mechanisms. Particularly, our result shows that around half of prompts in existing safe prompting benchmarks which were originally considered "safe" can actually be manipulated to bypass many deployed safety mechanisms, including concept removal, negative prompt, and safety guidance. Our findings suggest that, without comprehensive testing, the evaluations on limited safe prompting benchmarks can lead to a false sense of safety for text-to-image models.

Probabilistic Methods · Variational Inference

Brooks(Ruijia) Niu, Dongxia Wu, Kai Kim, Yian Ma, Duncan Watson-Parris, Rose Yu

Multi-fidelity surrogate modeling aims to learn an accurate surrogate at the highest fidelity level by combining data from multiple sources. Traditional methods relying on Gaussian processes can hardly scale to high-dimensional data. Deep learning approaches utilize neural network based encoders and decoders to improve scalability. These approaches share encoded representations across fidelities without including corresponding decoder parameters. This hinders inference performance, especially in out-of-distribution scenarios when the highest fidelity data has limited domain coverage. To address these limitations, we propose Multi-fidelity Residual Neural Processes (MFRNP), a novel multi-fidelity surrogate modeling framework. MFRNP explicitly models the residual between the aggregated output from lower fidelities and ground truth at the highest fidelity. The aggregation introduces decoders into the information sharing step and optimizes lower fidelity decoders to accurately capture both in-fidelity and cross-fidelity information. We show that MFRNP significantly outperforms state-of-the-art in learning partial differential equations and a real-world climate modeling task. Our code is published at: https://github.com/Rose-STL-Lab/MFRNP

Optimization

Fabian Schaipp, Ruben Ohana, Michael Eickenberg, Aaron Defazio, Robert Gower

Training a modern machine learning architecture on a new task requires extensive learning-rate tuning, which comes at a high computational cost. Here we develop new Polyak-type adaptive learning rates that can be used on top of any momentum method, and require less tuning to perform well. We first develop MoMo, a **Mo**mentum **Mo**del based adaptive learning rate for SGD-M (stochastic gradient descent with momentum). MoMo uses momentum estimates of the batch losses and gradients sampled at each iteration to build a model of the loss function. Our model also makes use of any known lower bound of the loss function by using truncation, e.g. most losses are lower-bounded by zero. The models is then approximately minimized at each iteration to compute the next step. We show how MoMo can be used in combination with any momentum-based method, and showcase this by developing MoMo-Adam - which is Adam with our new model-based adaptive learning rate. We show that MoMo attains a $\mathcal{O}(1/\sqrt{K})$ convergence rate for convex problems with interpolation, needing knowledge of no problem-specific quantities other than the optimal value. Additionally, for losses with unknown lower bounds, we develop on-the-fly estimates of a lower bound, that are incorporated in our model. We demonstrate that MoMo and MoMo-Adam improve over SGD-M and Adam in terms of robustness to hyperparameter tuning for training image classifiers on MNIST, CIFAR, and Imagenet, for recommender systems on the Criteo dataset, for a transformer model on the translation task IWSLT14, and for a diffusion model.

Reinforcement Learning · Deep RL

Théo Cachet, Christopher Dance, Olivier Sigaud

Vision-language models (VLMs) have tremendous potential for *grounding* language, and thus enabling *language-conditioned agents (LCAs)* to perform diverse tasks specified with text. This has motivated the study of LCAs based on reinforcement learning (RL) with rewards given by rendering images of an environment and evaluating those images with VLMs. If single-task RL is employed, such approaches are limited by the cost and time required to train a policy for each new task. Multi-task RL (MTRL) is a natural alternative, but requires a carefully designed corpus of training tasks and does not always generalize reliably to new tasks. Therefore, this paper introduces a novel decomposition of the problem of building an LCA: first find an *environment configuration* that has a high VLM score for text describing a task; then use a (pretrained) goal-conditioned policy to reach that configuration. We also explore several enhancements to the speed and quality of VLM-based LCAs, notably, the use of distilled models, and the evaluation of configurations from multiple viewpoints to resolve the ambiguities inherent in a single 2D view. We demonstrate our approach on the Humanoid environment, showing that it results in LCAs that outperform MTRL baselines in zero-shot generalization, without requiring any textual task descriptions or other forms of environment-specific annotation during training.