ICML 2024 Papers

The Bures-Wasserstein (BW) gradient descent method has gained considerable attention in various domains, including Gaussian barycenter, matrix recovery and variational inference problems, due to its alignment with the Wasserstein geometry of normal distributions. Despite its popularity, existing convergence analysis are often contingent upon specific loss functions, and the exploration of constrained settings within this framework remains limited. In this work, we make an attempt to bridge this gap by providing a general convergence rate guarantee for BW gradient descent when the Euclidean strong convexity of the loss and the constraints is assumed. In an effort to advance practical implementations, we also derive a closed-form solution for the projection onto BW distance-constrained sets, which enables the fast implementation of projected BW gradient descent for problems that arise in the constrained barycenter and distributionally robust optimization literature. Experimental results demonstrate significant improvements in computational efficiency and convergence speed, underscoring the efficacy of our method in practical scenarios.

We introduce a neural-preconditioned iterative solver for Poisson equations with mixed boundary conditions. Typical Poisson discretizations yield large, ill-conditioned linear systems. Iterative solvers can be effective for these problems, but only when equipped with powerful preconditioners. Unfortunately, effective preconditioners like multigrid require costly setup phases that must be re-executed every time domain shapes or boundary conditions change, forming a severe bottleneck for problems with evolving boundaries. In contrast, we present a neural preconditioner trained to efficiently approximate the inverse of the discrete Laplacian in the presence of such changes. Our approach generalizes to domain shapes, boundary conditions, and grid sizes outside the training set. The key to our preconditioner's success is a novel, lightweight neural network architecture featuring spatially varying convolution kernels and supporting fast inference. We demonstrate that our solver outperforms state-of-the-art methods like algebraic multigrid as well as recently proposed neural preconditioners on challenging test cases arising from incompressible fluid simulations.

We address the challenge of actively ranking a set of items/players with varying values/strengths. The comparison outcomes are random, with a greater noise the closer the values. A crucial requirement is that, at each iteration of the algorithm, all items must be compared once, i.e., an iteration is a perfect matching. Furthermore, we presume that comparing two players with closely matched strengths incurs no cost and, in contrast, a unit cost is associated with comparing players whose strength difference is more substantial. Our secondary objective is to determine an optimal matching between players based on this cost function: we propose and analyze an algorithm that draws on concepts from both AKS sorting networks and bandit theory. Our algorithm achieves both objectives with high probability, and the total cost is optimal (up to logarithmic terms).

Informally, the "linear representation hypothesis" is the idea that high-level concepts are represented linearly as directions in some representation space. In this paper, we address two closely related questions: What does "linear representation" actually mean? And, how do we make sense of geometric notions (e.g., cosine similarity and projection) in the representation space? To answer these, we use the language of counterfactuals to give two formalizations of linear representation, one in the output (word) representation space, and one in the input (context) space. We then prove that these connect to linear probing and model steering, respectively. To make sense of geometric notions, we use the formalization to identify a particular (non-Euclidean) inner product that respects language structure in a sense we make precise. Using this *causal inner product*, we show how to unify all notions of linear representation. In particular, this allows the construction of probes and steering vectors using counterfactual pairs. Experiments with LLaMA-2 demonstrate the existence of linear representations of concepts, the connection to interpretation and control, and the fundamental role of the choice of inner product.

In this work, we investigate the problem of adapting to the presence or absence of causal structure in multi-armed bandit problems. In addition to the usual reward signal, we assume the learner has access to additional variables, observed in each round after acting. When these variables $d$-separate the action from the reward, existing work in causal bandits demonstrates that one can achieve strictly better (minimax) rates of regret (Lu et al., 2020). Our goal is to adapt to this favorable ``conditionally benign'' structure, if it is present in the environment, while simultaneously recovering worst-case minimax regret, if it is not. Notably, the learner has no prior knowledge of whether the favorable structure holds. In this paper, we establish the Pareto optimal frontier of adaptive rates. We prove upper and matching lower bounds on the possible trade-offs in the performance of learning in conditionally benign and arbitrary environments, resolving an open question raised by Bilodeau et al. (2022). Furthermore, we are the first to obtain instance-dependent bounds for causal bandits, by reducing the problem to the linear bandit setting. Finally, we examine the common assumption that the marginal distributions of the post-action contexts are known and show that a nontrivial estimate is necessary for better-than-worst-case minimax rates.

When there exists uncertainty, AI machines are designed to make decisions so as to reach the best expected outcomes. Expectations are based on true facts about the objective environment the machines interact with, and those facts can be encoded into AI models in the form of true objective probability functions. Accordingly, AI models involve probabilistic machine learning in which the probabilities should be objectively interpreted. We prove under some basic assumptions when machines can learn the true objective probabilities, if any, and when machines cannot learn them.

Backdoor defense is crucial to ensure the safety and robustness of machine learning models when under attack. However, most existing methods specialize in either the detection or removal of backdoors, but seldom both. While few works have addressed both, these methods rely on strong assumptions or entail significant overhead costs, such as the need of task-specific samples for detection and model retraining for removal. Hence, the key challenge is how to reduce overhead and relax unrealistic assumptions. In this work, we propose two Energy-Based BAckdoor defense methods, called EBBA and EBBA+, that can achieve both backdoored model detection and backdoor removal with low overhead. Our contributions are twofold: First, we offer theoretical analysis for our observation that a predefined target label is more likely to occur among the top results for various samples. Inspired by this, we develop an enhanced energy-based technique, called EBBA, to detect backdoored models without task-specific samples (i.e., samples from any tasks). Secondly, we theoretically analyze that after data corruption, the original clean label of a poisoned sample is more likely to be predicted as a top output by the model, a sharp contrast to clean samples. Accordingly, we extend EBBA to develop EBBA+, a new transferred energy approach to efficiently detect poisoned images and remove backdoors without model retraining. Extensive experiments on multiple benchmark datasets demonstrate the superior performance of our methods over baselines in both backdoor detection and removal. Notably, the proposed methods can effectively detect backdoored model and poisoned images as well as remove backdoors at the same time.

What do different contrastive learning (CL) losses actually optimize for? Although multiple CL methods have demonstrated remarkable representation learning capabilities, the differences in their inner workings remain largely opaque. In this work, we analyse several CL families and prove that, under certain conditions, they admit the same minimisers when optimizing either their batch-level objectives or their expectations asymptotically. In both cases, an intimate connection with the hyperspherical energy minimisation (HEM) problem resurfaces. Drawing inspiration from this, we introduce a novel CL objective, coined Decoupled Hyperspherical Energy Loss (DHEL). DHEL simplifies the problem by decoupling the target hyperspherical energy from the alignment of positive examples while preserving the same theoretical guarantees. Going one step further, we show the same results hold for another relevant CL family, namely kernel contrastive learning (KCL), with the additional advantage of the expected loss being independent of batch size, thus identifying the minimisers in the non-asymptotic regime. Empirical results demonstrate improved downstream performance and robustness across combinations of different batch sizes and hyperparameters and reduced dimensionality collapse, on several computer vision datasets.

Monotonic Linear Interpolation (MLI) refers to the peculiar phenomenon that the error between the initial and converged model monotonically decreases along the linear interpolation, i.e., $(1-\alpha)\boldsymbol{\theta}_0 + \alpha \boldsymbol{\theta}_F$. Previous works focus on paired initial and converged points, relating MLI to the smoothness of the optimization trajectory. In this paper, we find a shocking fact that the error curves still exhibit a monotonic decrease when $\boldsymbol{\theta}_0$ is replaced with noise or even zero values, implying that the decreasing curve may be primarily related to the property of the converged model rather than the optimization trajectory. We further explore the relationship between $\alpha\boldsymbol{\theta}_F$ and $\boldsymbol{\theta}_F$ and propose scale invariance properties in various cases, including Generalized Scale Invariance (GSI), Rectified Scale Invariance (RSI), and Normalized Scale Invariance (NSI). From an inverse perspective, the MLI formula is essentially an equation that adds varying levels of noise (i.e., $(1-\alpha)\boldsymbol{\epsilon}$) to a nearly scale-invariant network (i.e., $\alpha \boldsymbol{\theta}_F$), resulting in a monotonically increasing error as the noise level rises. MLI is a special case where $\boldsymbol{\epsilon}$ is equal to $\boldsymbol{\theta}_0$.

Reinforcement Learning from Human Feedback (RLHF) is key to aligning Large Language Models (LLMs), typically paired with the Proximal Policy Optimization (PPO) algorithm. While PPO is a powerful method designed for general reinforcement learning tasks, it is overly sophisticated for LLMs, leading to laborious hyper-parameter tuning and significant computation burdens. To make RLHF efficient, we present ReMax, which leverages 3 properties of RLHF: fast simulation, deterministic transitions, and trajectory-level rewards. These properties are not exploited in PPO, making it less suitable for RLHF. Building on the renowned REINFORCE algorithm, ReMax does not require training an additional value model as in PPO and is further enhanced with a new variance reduction technique. ReMax offers several benefits over PPO: it is simpler to implement, eliminates more than 4 hyper-parameters in PPO, reduces GPU memory usage, and shortens training time. ReMax can save about 46% GPU memory than PPO when training a 7B model and enables training on A800-80GB GPUs without the memory-saving offloading technique needed by PPO. Applying ReMax to a Mistral-7B model resulted in a 94.78% win rate on the AlpacaEval leaderboard and a 7.739 score on MT-bench, setting a new SOTA for open-source 7B models. These results show the effectiveness of ReMax while addressing the limitations of PPO in LLMs.

Submodular functions, as well as the sub-class of decomposable submodular functions, and their optimization appear in a wide range of applications in machine learning, recommendation systems, and welfare maximization. However, optimization of decomposable submodular functions with millions of component functions is computationally prohibitive. Furthermore, the component functions may be private (they might represent user preference function, for example) and cannot be widely shared. To address these issues, we propose a *federated optimization* setting for decomposable submodular optimization. In this setting, clients have their own preference functions, and a weighted sum of these preferences needs to be maximized. We implement the popular *continuous greedy* algorithm in this setting where clients take parallel small local steps towards the local solution and then the local changes are aggregated at a central server. To address the large number of clients, the aggregation is performed only on a subsampled set. Further, the aggregation is performed only intermittently between stretches of parallel local steps, which reduces communication cost significantly. We show that our federated algorithm is guaranteed to provide a good approximate solution, even in the presence of above cost-cutting measures. Finally, we show how the federated setting can be incorporated in solving fundamental discrete submodular optimization problems such as Maximum Coverage and Facility Location.

With the rapid development of Artificial Intelligence of Things (AIoT), customizing and optimizing operating system (OS) kernel configurations for various AIoT application scenarios is crucial for maximizing system performance. However, existing approaches falter due to the overwhelming problem complexity (i.e., over 15,000 configuration options in the Linux kernel), together with the huge evaluation costs and error-prone options that may result in OS boot-up failure, which all make it an unresolved problem to optimize the Linux kernel automatically. In this paper, we introduce AutoOS, a novel framework exploiting Large Language Models for customizing and optimizing OS kernel configurations automatically for various AIoT application scenarios.Inspired by the inherently directory-structured kernel configuration process, we first formulate our research problem as optimizing on a dynamic tree. We then propose a novel framework integrating a state machine-based traversal algorithm as the observe-prune-propose-act-correct loop, which can effectively refine the optimization space and ensure a successful OS boot-up.Experimental results show that AutoOS can automatically customize and optimize the OS kernel configurations without human effort. More importantly, AutoOS even achieves better performance by up to 25% than vendor-provided configuration.

As the number of large language models (LLMs) released to the public grows, there is a pressing need to understand the safety implications associated with these models learning from third-party custom finetuning data. We explore the behavior of LLMs finetuned on noisy custom data containing unsafe content, represented by datasets that contain biases, toxicity, and harmfulness, finding that while aligned LLMs can readily learn this unsafe content, they also tend to forget it more significantly than other examples when subsequently finetuned on safer content. Drawing inspiration from the discrepancies in forgetting, we introduce the “ForgetFilter” algorithm, which filters unsafe data based on how strong the model's forgetting signal is for that data. We demonstrate that the ForgetFilter algorithm ensures safety in customized finetuning without compromising downstream task performance, unlike sequential safety finetuning. ForgetFilter outperforms alternative strategies like replay and moral self-correction in curbing LLMs’ ability to assimilate unsafe content during custom finetuning, e.g. 75% lower than not applying any safety measures and 62% lower than using self-correction in toxicity score.

An open scientific challenge is how to classify events with reliable measures of uncertainty, when we have a mechanistic model of the data-generating process but the distribution over both labels and latent nuisance parameters is different between train and target data. We refer to this type of distributional shift as generalized label shift (GLS). Direct classification using observed data $\mathbf{X}$ as covariates leads to biased predictions and invalid uncertainty estimates of labels $Y$. We overcome these biases by proposing a new method for robust uncertainty quantification that casts classification as a hypothesis testing problem under nuisance parameters. The key idea is to estimate the classifier's receiver operating characteristic (ROC) across the entire nuisance parameter space, which allows us to devise cutoffs that are invariant under GLS. Our method effectively endows a pre-trained classifier with domain adaptation capabilities and returns valid prediction sets while maintaining high power. We demonstrate its performance on two challenging scientific problems in biology and astroparticle physics with data from realistic mechanistic models.

We prove that the combination of a target network and over-parameterized linear function approximation establishes a weaker convergence condition for bootstrapped value estimation in certain cases, even with off-policy data. Our condition is naturally satisfied for expected updates over the entire state-action space or learning with a batch of complete trajectories from episodic Markov decision processes. Notably, using only a target network or an over-parameterized model does not provide such a convergence guarantee. Additionally, we extend our results to learning with truncated trajectories, showing that convergence is achievable for all tasks with minor modifications, akin to value truncation for the final states in trajectories. Our primary result focuses on temporal difference estimation for prediction, providing high-probability value estimation error bounds and empirical analysis on Baird's counterexample and a Four-room task. Furthermore, we explore the control setting, demonstrating that similar convergence conditions apply to Q-learning.

We study the certifiable robustness of ML classifiers on dirty datasets that could contain missing values. A test point is certifiably robust for an ML classifier if the classifier returns the same prediction for that test point, regardless of which cleaned version (among exponentially many) of the dirty dataset the classifier is trained on. In this paper, we show theoretically that for Naive Bayes Classifiers (NBC) over dirty datasets with missing values: (i) there exists an efficient polynomial time algorithm to decide whether multiple input test points are all certifiably robust over a dirty dataset; and (ii) the data poisoning attack, which aims to make all input test points certifiably non-robust by inserting missing cells to the clean dataset, is in polynomial time for single test points but NP-complete for multiple test points. Extensive experiments demonstrate that our algorithms are efficient and outperform existing baselines.

The remarkable generalization of large-scale models has recently gained significant attention in multimodal research. However, deploying heterogeneous large-scale models with different modalities under Federated Learning (FL) to protect data privacy imposes tremendous challenges on clients' limited computation and storage. In this work, we propose M$^2$FedSA to address the above issue. We realize modularized decomposition of large-scale models via Split Learning (SL) and only retain privacy-sensitive modules on clients, alleviating storage overhead. By freezing large-scale models and introducing two specialized lightweight adapters, the models can better focus on task-specific knowledge and enhance modality-specific knowledge, improving the model's adaptability to different tasks while balancing efficiency. In addition, M$^2$FedSA further improves performance by transferring multimodal knowledge to unimodal clients at both the feature and decision levels, which leverages the complementarity of different modalities. Extensive experiments on various multimodal classification tasks validate the effectiveness of our proposed M$^2$FedSA. The code is made available publicly at https://github.com/M2FedSA/M-2FedSA.

We consider the problem of multi-objective alignment of foundation models with human preferences, which is a critical step towards helpful and harmless AI systems. However, it is generally costly and unstable to fine-tune large foundation models using reinforcement learning (RL), and the multi-dimensionality, heterogeneity, and conflicting nature of human preferences further complicate the alignment process. In this paper, we introduce Rewards-in-Context (RiC), which conditions the response of a foundation model on multiple rewards in its prompt context and applies supervised fine-tuning for alignment. The salient features of RiC are simplicity and adaptivity, as it only requires supervised fine-tuning of a single foundation model and supports dynamic adjustment for user preferences during inference time. Inspired by the analytical solution of an abstracted convex optimization problem, our dynamic inference-time adjustment method approaches the Pareto-optimal solution for multiple objectives. Empirical evidence demonstrates the efficacy of our method in aligning both Large Language Models (LLMs) and diffusion models to accommodate diverse rewards with only around 10% GPU hours compared with multi-objective RL baseline.

We show that a constant number of self-attention layers can efficiently simulate—and be simulated by—a constant number of communication rounds of *Massively Parallel Computation*. As a consequence, we show that logarithmic-depth is sufficient for transformers to solve basic computational tasks that cannot be efficiently solved by several other neural sequence models and sub-quadratic transformer approximations. We thus establish parallelism as a key distinguishing property of transformers.