ICML 2024 Papers

What is *agency,* and why does it matter? In this work, we draw from the political science and philosophy literature and give two competing visions of what it means to be an (ethical) agent. The first view, which we term *mechanistic*, is commonly— and implicitly—assumed in AI research, yet it is a fundamentally limited means to understand the ethical characteristics of AI. Under the second view, which we term volitional, AI can no longer be considered an ethical agent. We discuss the implications of each of these views for two critical questions: first, what the ideal system “ought” to look like, and second, how accountability may be achieved. In light of this discussion, we ultimately argue that, in the context of ethically-significant behavior, AI should be viewed not as an agent but as the outcome of political processes.

Graph neural network architectures aligned with the $k$-dimensional Weisfeiler--Leman ($k$-WL) hierarchy offer theoretically well-understood expressive power. However, these architectures often fail to deliver state-of-the-art predictive performance on real-world graphs, limiting their practical utility. While recent works aligning graph transformer architectures with the $k$-WL hierarchy have shown promising empirical results, employing transformers for higher orders of $k$ remains challenging due to a prohibitive runtime and memory complexity of self-attention as well as impractical architectural assumptions, such as an infeasible number of attention heads. Here, we advance the alignment of transformers with the $k$-WL hierarchy, showing stronger expressivity results for each $k$, making them more feasible in practice. In addition, we develop a theoretical framework that allows the study of established positional encodings such as Laplacian PEs and SPE. We evaluate our transformers on the large-scale PCQM4Mv2 dataset, showing competitive predictive performance with the state-of-the-art and demonstrating strong downstream performance when fine-tuning them on small-scale molecular datasets.

Causal inference on networks faces challenges posed in part by violations of standard identification assumptions due to dependencies between treatment units. Although graph geometry fundamentally influences such dependencies, the potential of geometric tools for causal inference on networked treatment units is yet to be unlocked. Moreover, despite significant progress utilizing graph neural networks (GNNs) for causal inference on networks, methods for evaluating their achievable reliability without ground truth are lacking. In this work we establish for the first time a theoretical link between network geometry, the graph Ricci curvature in particular, and causal inference, formalizing the intrinsic challenges that negative curvature poses to estimating causal parameters. The Ricci curvature can then be used to assess the reliability of causal estimates in structured data, as we empirically demonstrate. Informed by this finding, we propose a method using the geometric Ricci flow to reduce causal effect estimation error in networked data, showcasing how this newfound connection between graph geometry and causal inference could improve GNN-based causal inference. Bridging graph geometry and causal inference, this paper opens the door to geometric techniques for improving causal estimation on networks.

Deep neural networks based on linear RNNs interleaved with position-wise MLPs are gaining traction as competitive approaches for sequence modeling. Examples of such architectures include state-space models (SSMs) like S4, LRU, and Mamba: recently proposed models that achieve promising performance on text, genetics, and other data that require long-range reasoning. Despite experimental evidence highlighting these architectures' effectiveness and computational efficiency, their expressive power remains relatively unexplored, especially in connection to specific choices crucial in practice - e.g., carefully designed initialization distribution and potential use of complex numbers. In this paper, we show that combining MLPs with both real or complex linear diagonal recurrences leads to arbitrarily precise approximation of regular causal sequence-to-sequence maps. At the heart of our proof, we rely on a separation of concerns: the linear RNN provides a lossless encoding of the input sequence, and the MLP performs non-linear processing on this encoding. While we show that real diagonal linear recurrences are enough to achieve universality in this architecture, we prove that employing complex eigenvalues near unit disk - i.e., empirically the most successful strategy in S4 - greatly helps the RNN in storing information. We connect this finding with the vanishing gradient issue and provide experiments supporting our claims.

We revisit the objective perturbations framework for differential privacy where noise is added to the input $A\in \mathcal{S}$ and the result is then projected back to the space of admissible datasets $\mathcal{S}$. Through this framework, we first design novel efficient algorithms to privately release pair-wise cosine similarities. Second, we derive a novel algorithm to compute $k$-way marginal queries over $n$ features. Prior work could achieve comparable guarantees only for $k$ even. Furthermore, we extend our results to $t$-sparse datasets, where our efficient algorithms yields novel, stronger guarantees whenever $t\le n^{5/6}/\log n.$ Finally, we provide a theoretical perspective on why *fast* input perturbation algorithms works well in practice. The key technical ingredients behind our results are tight sum-of-squares certificates upper bounding the Gaussian complexity of sets of solutions.

Some neural language models (LMs) exhibit a remarkable capacity for in-context learning (ICL): they can fit predictors to datasets provided as input. While the mechanisms underlying ICL are well-studied in the context of synthetic problems like in-context linear regression, there is still some divergence between these model problems and the “real” ICL exhibited by LMs trained on large text corpora. In this paper, we study ICL through the lens of a new family of model problems we term in context language learning (ICLL). In ICLL, LMs are presented with a set of strings from a formal language, and must generate additional strings from the same language. We focus on in- context learning of regular languages generated by random finite automata. We evaluate a diverse set of neural sequence models on regular ICLL tasks. We first show that Transformers significantly outperform neural sequence models with recurrent or convolutional representations on ICLL tasks. Next, we provide evidence that they do so by computing in-context n-gram statistics using specialized attention heads. Finally, we show that hard-wiring these heads into neural models improves performance not just on synthetic ICLL, but natural language modeling, reducing the perplexity of 340M-parameter Transformers by up to 1.14 points (6.7%) on the SlimPajama dataset. Our results highlight the usefulness of in-context formal language learning as a tool for understanding ICL in models of natural text.

Can we identify the weights of a neural network by probing its input-output mapping? At first glance, this problem seems to have many solutions because of permutation, overparameterisation and activation function symmetries. Yet, we show that the incoming weight vector of each neuron is identifiable up to sign or scaling, depending on the activation function. Our novel method 'Expand-and-Cluster’ can identify layer sizes and weights of a target network for all commonly used activation functions. Expand-and-Cluster consists of two phases: (i) to relax the non-convex optimisation problem, we train multiple overparameterised student networks to best imitate the target function; (ii) to reverse engineer the target network's weights, we employ an ad-hoc clustering procedure that reveals the learnt weight vectors shared between students -- these correspond to the target weight vectors. We demonstrate successful weights and size recovery of trained shallow and deep networks with less than 10% overhead in the layer size and describe an 'ease-of-identifiability' axis by analysing 150 synthetic problems of variable difficulty.

In recent years, the rise of generative self-supervised learning (SSL) paradigms has exhibited impressive performance across visual, language, and multi-modal domains. While the varied designs of generative SSL objectives lead to distinct properties in downstream tasks, a theoretical understanding of these differences remains largely unexplored. In this paper, we establish the first theoretical comparisons between two leading generative SSL paradigms: autoregressive SSL and masked SSL. Through establishing theoretical frameworks, we elucidate the strengths and limitations of autoregressive and masked SSL within the primary evaluation tasks of classification and content generation. Our findings demonstrate that in classification tasks, the flexibility of targeted tokens in masked SSL fosters more inter-sample connections compared to the fixed position of target tokens in autoregressive SSL, which yields superior clustering performance. In content generation tasks, the misalignment between the flexible lengths of test samples and the fixed length of unmasked texts in masked SSL (vs. flexible lengths of conditional texts in autoregressive SSL) hinders its generation performance. To leverage each other's strengths and mitigate weaknesses, we propose diversity-enhanced autoregressive and variable-length masked objectives, which substantially improve the classification performance of autoregressive SSL and the generation performance of masked SSL. Code is available at https://github.com/PKU-ML/LookAheadLookAround.

Ensuring alignment of language model's outputs with human preferences is critical to guarantee a useful, safe, and pleasant user experience. Thus, human alignment has been extensively studied recently and several methods such as Reinforcement Learning from Human Feedback (RLHF), Direct Policy Optimisation (DPO) and Sequence Likelihood Calibration (SLiC) have emerged. In this paper, our contribution is two-fold. First, we show the equivalence between two recent alignment methods, namely Identity Policy Optimisation (IPO) and Nash Mirror Descent (Nash-MD). Second, we introduce a generalisation of IPO, named IPO-MD, that leverages the regularised sampling approach proposed by Nash-MD. This equivalence may seem surprising at first sight, since IPO is an offline method whereas Nash-MD is an online method using a preference model. However, this equivalence can be proven when we consider the online version of IPO, that is when both generations are sampled by the online policy and annotated by a trained preference model. Optimising the IPO loss with such a stream of data becomes then equivalent to finding the Nash equilibrium of the preference model through self-play. Building on this equivalence, we introduce the IPO-MD algorithm that generates data with a mixture policy (between the online and reference policy) similarly as the general Nash-MD algorithm. We compare online-IPO and IPO-MD to different online versions of existing losses on preference data such as DPO and SLiC on a summarisation task.

Adversarial Training (AT), which adversarially perturb the input samples during training, has been acknowledged as one of the most effective defenses against adversarial attacks, yet suffers from inevitably decreased clean accuracy. Instead of perturbing the samples, Sharpness-Aware Minimization (SAM) perturbs the model weights during training to find a more flat loss landscape and improve generalization. However, as SAM is designed for better clean accuracy, its effectiveness in enhancing adversarial robustness remains unexplored. In this work, considering the duality between SAM and AT, we investigate the adversarial robustness derived from SAM. Intriguingly, we find that using SAM alone can improve adversarial robustness. To understand this unexpected property of SAM, we first provide empirical and theoretical insights into how SAM can implicitly learn more robust features, and conduct comprehensive experiments to show that SAM can improve adversarial robustness notably without sacrificing any clean accuracy, shedding light on the potential of SAM to be a substitute for AT when accuracy comes at a higher priority. Code is available at https://github.com/weizeming/SAM_AT.

Multimodal pretraining is an effective strategy for the trinity of goals of representation learning in autonomous robots: $1)$ extracting both local and global task progressions; $2)$ enforcing temporal consistency of visual representation; $3)$ capturing trajectory-level language grounding. Most existing methods approach these via separate objectives, which often reach sub-optimal solutions. In this paper, we propose a universal unified objective that can simultaneously extract meaningful task progression information from image sequences and seamlessly align them with language instructions. We discover that via implicit preferences, where a visual trajectory inherently aligns better with its corresponding language instruction than mismatched pairs, the popular Bradley-Terry model can transform into representation learning through proper reward reparameterizations. The resulted framework, DecisionNCE, mirrors an InfoNCE-style objective but is distinctively tailored for decision-making tasks, providing an embodied representation learning framework that elegantly extracts both local and global task progression features, with temporal consistency enforced through implicit time contrastive learning, while ensuring trajectory-level instruction grounding via multimodal joint encoding. Evaluation on both simulated and real robots demonstrates that DecisionNCE effectively facilitates diverse downstream policy learning tasks, offering a versatile solution for unified representation and reward learning. Project Page: https://2toinf.github.io/DecisionNCE/

Multi-layer perceptron (MLP) is a fundamental component of deep learning, and recent MLP-based architectures, especially the MLP-Mixer, have achieved significant empirical success. Nevertheless, our understanding of why and how the MLP-Mixer outperforms conventional MLPs remains largely unexplored. In this work, we reveal that sparseness is a key mechanism underlying the MLP-Mixers. First, the Mixers have an effective expression as a wider MLP with Kronecker-product weights, clarifying that the Mixers efficiently embody several sparseness properties explored in deep learning. In the case of linear layers, the effective expression elucidates an implicit sparse regularization caused by the model architecture and a hidden relation to Monarch matrices, which is also known as another form of sparse parameterization. Next, for general cases, we empirically demonstrate quantitative similarities between the Mixer and the unstructured sparse-weight MLPs. Following a guiding principle proposed by Golubeva, Neyshabur and Gur-Ari (2021), which fixes the number of connections and increases the width and sparsity, the Mixers can demonstrate improved performance.

While stochastic bilevel optimization methods have been extensively studied for addressing large-scale nested optimization problems in machine learning, it remains an open question whether the optimal complexity bounds for solving bilevel optimization are the same as those in single-level optimization. Our main result resolves this question: SPABA, an adaptation of the PAGE method for nonconvex optimization in (Li et al., 2021) to the bilevel setting, can achieve optimal sample complexity in both the finite-sum and expectation settings. We show the optimality of SPABA by proving that there is no gap in complexity analysis between stochastic bilevel and single-level optimization when implementing PAGE. Notably, as indicated by the results of (Dagréou et al., 2022), there might exist a gap in complexity analysis when implementing other stochastic gradient estimators, like SGD and SAGA. In addition to SPABA, we propose several other single-loop stochastic bilevel algorithms, that either match or improve the state-of-the-art sample complexity results, leveraging our convergence rate and complexity analysis. Numerical experiments demonstrate the superior practical performance of the proposed methods.

Pre-trained language models (PLMs) are commonly used for various downstream natural language processing tasks via fine-tuning. However, recent studies have demonstrated that PLMs are vulnerable to backdoor attacks, which can mislabel poisoned samples to target outputs even after a vanilla fine-tuning process. The key challenge for defending against the backdoored PLMs is that end users who adopt the PLMs for their downstream tasks usually do not have any knowledge about the attacking strategies, such as triggers. To tackle this challenge, in this work, we propose a backdoor mitigation approach, PURE, via head pruning and normalization of attention weights. The idea is to prune the attention heads that are potentially affected by poisoned texts with only clean texts on hand and then further normalize the weights of remaining attention heads to mitigate the backdoor impacts. We conduct experiments to defend against various backdoor attacks on the classification task. The experimental results show the effectiveness of PURE in lowering the attack success rate without sacrificing the performance on clean texts.

Fine-tuning pre-trained Large Language Models (LLMs) is essential to align them with human values and intentions. This process often utilizes methods like pairwise comparisons and KL divergence against a reference LLM, focusing on the evaluation of full answers generated by the models. However, the generation of these responses occurs in a token level, following a sequential, auto-regressive fashion. In this paper, we introduce Token-level Direct Preference Optimization (TDPO), a novel approach to align LLMs with human preferences by optimizing policy at the token level. Unlike previous methods, which face challenges in divergence efficiency, TDPO integrates forward KL divergence constraints for each token, improving alignment and diversity. Utilizing the Bradley-Terry model for a token-based reward system, our method enhances the regulation of KL divergence, while preserving simplicity without the need for explicit reward modeling. Experimental results across various text tasks demonstrate TDPO’s superior performance in balancing alignment with generation diversity. Notably, fine-tuning with TDPO strikes a better balance than DPO in the controlled sentiment generation and single-turn dialogue datasets, and significantly improves the quality of generated responses compared to both DPO and PPO-based RLHF methods.

This paper proposes two different models for equitable resource allocation in online settings. The first one is called *external* equity promotion, where sequentially arriving agents are heterogeneous in their external attributes, namely how many resources they demand, which are drawn from a probability distribution (accessible to the algorithm). The focus is then to devise an allocation policy such that every requester can get a fair share of resources *proportional to their demands*, regardless of their arrival time. The second is called *internal* equity promotion, where arriving requesters can be treated homogeneously in external attributes (demands) but are heterogeneous in internal traits such as demographics. In particular, each requester can be identified as belonging to one or several groups, and an allocation of resources is regarded as equitable when every group of requesters can receive a fair share of resources proportional to the percentage of that group in the whole population. For both models above, we consider as the benchmark a clairvoyant optimal solution that has the privilege to access all random demand realizations in advance. We consider two equity metrics, namely *ex-post* and *ex-ante*, and discuss the challenges under the two metrics in detail. Specifically, we present two linear program (LP)-based policies for external equity promotion under ex-ante with independent demands, each achieving an *optimal* CR of $1/2$ with respect to the benchmark LP. For internal equity promotion, we present optimal policies under both ex-ante and ex-post metrics.

Online linear programming plays an important role in both revenue management and resource allocation, and recent research has focused on developing efficient first-order online learning algorithms. Despite the empirical success of first-order methods, they typically achieve regret no better than $\mathcal{O}(\sqrt{T})$, which is suboptimal compared to the $\mathcal{O}(\log T)$ result guaranteed by the state-of-the-art linear programming (LP)-based online algorithms. This paper establishes several important facts about online linear programming, which unveils the challenge for first-order online algorithms to achieve beyond $\mathcal{O}(\sqrt{T})$ regret. To address this challenge, we introduce a new algorithmic framework which decouples learning from decision-making. For the first time, we show that first-order methods can achieve regret $\mathcal{O}(T^{1/3})$ with this new framework.

Despite the recent advancements in offline RL, no unified algorithm could achieve superior performance across a broad range of tasks. Offline *value function learning*, in particular, struggles with sparse-reward, long-horizon tasks due to the difficulty of solving credit assignment and extrapolation errors that accumulates as the horizon of the task grows. On the other hand, models that can perform well in long-horizon tasks are designed specifically for goal-conditioned tasks, which commonly perform worse than value function learning methods on short-horizon, dense-reward scenarios. To bridge this gap, we propose a hierarchical planner designed for offline RL called PlanDQ. PlanDQ incorporates a diffusion-based planner at the high level, named D-Conductor, which guides the low-level policy through sub-goals. At the low level, we used a Q-learning based approach called the Q-Performer to accomplish these sub-goals. Our experimental results suggest that PlanDQ can achieve superior or competitive performance on D4RL continuous control benchmark tasks as well as AntMaze, Kitchen, and Calvin as long-horizon tasks.

We introduce a novel algorithm for estimating Cumulative Distribution Function (CDF) values under Local Differential Privacy (LDP) by exploiting an unexpected connection between LDP and the current status problem, a classical survival data problem in statistics. This connection leads to the development of tools for constrained isotonic estimation based on binary queries. Through mathematical proofs and extensive numerical testing, we demonstrate that our method achieves uniform and $L_2$ error bounds when estimating the entire CDF curve. By employing increasingly dense grids, the error bound can be improved, exhibiting an asymptotic normal distribution of the proposed estimator. Theoretically, we show that the error bound smoothly changes as the number of grids increases relative to the sample size $n$. Computationally, we demonstrate that our constrained isotonic estimator can be efficiently computed deterministically, eliminating the need for hyperparameters or random optimization.