ICML 2024 Papers

Cross-modality transfer aims to leverage large pretrained models to complete tasks that may not belong to the modality of pretraining data. Existing works achieve certain success in extending classical finetuning to cross-modal scenarios, yet we still lack understanding about the influence of modality gap on the transfer. In this work, a series of experiments focusing on the source representation quality during transfer are conducted, revealing the connection between larger modality gap and lesser knowledge reuse which means ineffective transfer. We then formalize the gap as the knowledge misalignment between modalities using conditional distribution $P(Y|X)$. Towards this problem, we present **Mo**dality k**N**owledge **A**lignment (MoNA), a meta-learning approach that learns target data transformation to reduce the modality knowledge discrepancy ahead of the transfer. Experiments show that the approach significantly improves upon cross-modal finetuning methods, and most importantly leads to better reuse of source modality knowledge.

Spiking neural networks (SNNs) are gaining popularity in deep learning due to their low energy budget on neuromorphic hardware. However, they still face challenges in lacking sufficient robustness to guard safety-critical applications such as autonomous driving. Many studies have been conducted to defend SNNs from the threat of adversarial attacks. This paper aims to uncover the robustness of SNN through the lens of the stability of nonlinear systems. We are inspired by the fact that searching for parameters altering the leaky integrate-and-fire dynamics can enhance their robustness. Thus, we dive into the dynamics of membrane potential perturbation and simplify the formulation of the dynamics. We present that membrane potential perturbation dynamics can reliably convey the intensity of perturbation. Our theoretical analyses imply that the simplified perturbation dynamics satisfy input-output stability. Thus, we propose a training framework with modified SNN neurons and to reduce the mean square of membrane potential perturbation aiming at enhancing the robustness of SNN. Finally, we experimentally verify the effectiveness of the framework in the setting of Gaussian noise training and adversarial training on the image classification task. Please refer to https://github.com/DingJianhao/stable-snn for our code implementation.

Aligning large language models (LLMs) with human values is imperative to mitigate potential adverse effects resulting from their misuse. Drawing from the sociological insight that acknowledging all parties' concerns is a key factor in shaping human values, this paper proposes a novel direction to align LLMs by themselves: social scene simulation. To achieve this, we present MATRIX, a novel social scene simulator that emulates realistic scenes around a user's input query, enabling the LLM to take social consequences into account before responding. MATRIX serves as a virtual rehearsal space, akin to a Monopolylogue, where the LLM performs diverse roles related to the query and practice by itself. To inject this alignment, we fine-tune the LLM with MATRIX-simulated data, ensuring adherence to human values without compromising inference speed. We theoretically show that the LLM with MATRIX outperforms existing methods under mild assumptions. Finally, extensive experiments validate that our method outperforms over 10 baselines across 4 benchmarks. As evidenced by 875 user ratings, our tuned 13B-size LLM exceeds GPT-4 in aligning with human values. See our project page at https://shuotang123.github.io/MATRIX.

This paper presents a new *bi-Lipschitz* invertible neural network, the BiLipNet, which has the ability to smoothly control both its *Lipschitzness* (output sensitivity to input perturbations) and *inverse Lipschitzness* (input distinguishability from different outputs). The second main contribution is a new scalar-output network, the PLNet, which is a composition of a BiLipNet and a quadratic potential. We show that PLNet satisfies the Polyak-Łojasiewicz condition and can be applied to learn non-convex surrogate losses with a unique and efficiently-computable global minimum. The central technical element in these networks is a novel invertible residual layer with certified strong monotonicity and Lipschitzness, which we compose with orthogonal layers to build the BiLipNet. The certification of these properties is based on incremental quadratic constraints, resulting in much tighter bounds than can be achieved with spectral normalization. Moreover, we formulate the calculation of the inverse of a BiLipNet -- and hence the minimum of a PLNet -- as a series of three-operator splitting problems, for which fast algorithms can be applied.

Cutting plane methods are a fundamental approach for solving integer linear programs (ILPs). In each iteration of such methods, additional linear constraints (cuts) are introduced to the constraint set with the aim of excluding the previous fractional optimal solution while not affecting the optimal integer solution. In this work, we explore a novel approach within cutting plane methods: instead of only adding new cuts, we also consider the removal of previous cuts introduced at any of the preceding iterations of the method under a learnable parametric criteria. We demonstrate that in fundamental combinatorial optimization settings such cut removal policies can lead to significant improvements over both human-based and machine learning-guided cut addition policies even when implemented with simple models.

Discrimination and calibration represent two important properties of survival analysis, with the former assessing the model's ability to accurately rank subjects and the latter evaluating the alignment of predicted outcomes with actual events. With their distinct nature, it is hard for survival models to simultaneously optimize both of them especially as many previous results found improving calibration tends to diminish discrimination performance. This paper introduces a novel approach utilizing *conformal regression* that can improve a model's calibration without degrading discrimination. We provide theoretical guarantees for the above claim, and rigorously validate the efficiency of our approach across 11 real-world datasets, showcasing its practical applicability and robustness in diverse scenarios.

In this paper, we first present an explanation regarding the common occurrence of spikes in the training loss when neural networks are trained with stochastic gradient descent (SGD). We provide evidence that the spikes in the training loss of SGD are "catapults", an optimization phenomenon originally observed in GD with large learning rates in Lewkowycz et al. (2020). We empirically show that these catapults occur in a low-dimensional subspace spanned by the top eigenvectors of the tangent kernel, for both GD and SGD. Second, we posit an explanation for how catapults lead to better generalization by demonstrating that catapults increase feature learning by increasing alignment with the Average Gradient Outer Product (AGOP) of the true predictor. Furthermore, we demonstrate that a smaller batch size in SGD induces a larger number of catapults, thereby improving AGOP alignment and test performance.

Real-valued functions on geometric data---such as node attributes on a graph---can be optimized using descriptors from persistent homology, allowing the user to incorporate topological terms in the loss function. When optimizing a single real-valued function (the one-parameter setting), there is a canonical choice of descriptor for persistent homology: the barcode. The operation mapping a real-valued function to its barcode is differentiable almost everywhere, and the convergence of gradient descent for losses using barcodes is relatively well understood. When optimizing a vector-valued function (the multiparameter setting), there is no unique choice of descriptor for multiparameter persistent homology, and many distinct descriptors have been proposed. This calls for the development of a general framework for differentiability and optimization that applies to a wide range of multiparameter homological descriptors. In this article, we develop such a framework and show that it encompasses well-known descriptors of different flavors, such as signed barcodes and the multiparameter persistence landscape. We complement the theory with numerical experiments supporting the idea that optimizing multiparameter homological descriptors can lead to improved performances compared to optimizing one-parameter descriptors, even when using the simplest and most efficiently computable multiparameter descriptors.

Existing AI alignment approaches assume that preferences are static, which is unrealistic: our preferences change, and may even be influenced by our interactions with AI systems themselves. To clarify the consequences of incorrectly assuming static preferences, we introduce Dynamic Reward Markov Decision Processes (DR-MDPs), which explicitly model preference changes and the AI's influence on them. We show that despite its convenience, the static-preference assumption may undermine the soundness of existing alignment techniques, leading them to implicitly reward AI systems for influencing user preferences in ways users may not truly want. We then explore potential solutions. First, we offer a unifying perspective on how an agent's optimization horizon may partially help reduce undesirable AI influence. Then, we formalize different notions of AI alignment that account for preference change from the outset. Comparing the strengths and limitations of 8 such notions of alignment, we find that they all either err towards causing undesirable AI influence, or are overly risk-averse, suggesting that a straightforward solution to the problems of changing preferences may not exist. As there is no avoiding grappling with changing preferences in real-world settings, this makes it all the more important to handle these issues with care, balancing risks and capabilities. We hope our work can provide conceptual clarity and constitute a first step towards AI alignment practices which explicitly account for (and contend with) the changing and influenceable nature of human preferences.

Deep learning-based image restoration methods generally struggle with faithfully preserving the structures of the original image. In this work, we propose a novel Residual-Conditioned Optimal Transport (RCOT) approach, which models image restoration as an optimal transport (OT) problem for both unpaired and paired settings, introducing the transport residual as a unique degradation-specific cue for both the transport cost and the transport map. Specifically, we first formalize a Fourier residual-guided OT objective by incorporating the degradation-specific information of the residual into the transport cost. We further design the transport map as a two-pass RCOT map that comprises a base model and a refinement process, in which the transport residual is computed by the base model in the first pass and then encoded as a degradation-specific embedding to condition the second-pass restoration. By duality, the RCOT problem is transformed into a minimax optimization problem, which can be solved by adversarially training neural networks. Extensive experiments on multiple restoration tasks show that RCOT achieves competitive performance in terms of both distortion measures and perceptual quality, restoring images with more faithful structures as compared with state-of-the-art methods.

Kahneman & Tversky's $\textit{prospect theory}$ tells us that humans perceive random variables in a biased but well-defined manner (1992); for example, humans are famously loss-averse. We show that objectives for aligning LLMs with human feedback implicitly incorporate many of these biases---the success of these objectives (e.g., DPO) over cross-entropy minimization can partly be ascribed to them belonging to a family of loss functions that we call $\textit{human-aware losses}$ (HALOs). However, the utility functions these methods attribute to humans still differ from those in the prospect theory literature. Using a Kahneman-Tversky model of human utility, we propose a HALO that directly maximizes the utility of generations instead of maximizing the log-likelihood of preferences, as current methods do. We call this approach KTO, and it matches or exceeds the performance of preference-based methods at scales from 1B to 30B, despite only learning from a binary signal of whether an output is desirable. More broadly, our work suggests that there is no one HALO that is universally superior; the best loss depends on the inductive biases most appropriate for a given setting, an oft-overlooked consideration.

Combining the predictions of multiple trained models through ensembling is generally a good way to improve accuracy by leveraging the different learned features of the models, however it comes with high computational and storage costs. Model fusion, the act of merging multiple models into one by combining their parameters reduces these costs but doesn't work as well in practice. Indeed, neural network loss landscapes are high-dimensional and non-convex and the minima found through learning are typically separated by high loss barriers. Numerous recent works have been focused on finding permutations matching one network features to the features of a second one, lowering the loss barrier on the linear path between them in parameter space. However, permutations are restrictive since they assume a one-to-one mapping between the different models' neurons exists. We propose a new model merging algorithm, CCA Merge, which is based on Canonical Correlation Analysis and aims to maximize the correlations between linear combinations of the model features. We show that our alignment method leads to better performances than past methods when averaging models trained on the same, or differing data splits. We also extend this analysis into the harder setting where more than 2 models are merged, and we find that CCA Merge works significantly better than past methods. Our code is publicly available at https://github.com/shoroi/align-n-merge

Reinforcement Learning from Human Feedback (RLHF) has achieved impressive empirical successes while relying on a small amount of human feedback. However, there is limited theoretical justification for this phenomenon. Additionally, most recent studies focus on value-based algorithms despite the recent empirical successes of policy-based algorithms. In this work, we consider an RLHF algorithm based on policy optimization (PO-RLHF). The algorithm is based on the popular Policy Cover-Policy Gradient (PC-PG) algorithm, which assumes knowledge of the reward function. In PO-RLHF, knowledge of the reward function is not assumed and the algorithm relies on trajectory-based comparison feedback to infer the reward function. We provide performance bounds for PO-RLHF with low query complexity, which provides insight into why a small amount of human feedback may be sufficient to get good performance with RLHF. A key novelty is our trajectory-level elliptical potential analysis technique used to infer reward function parameters when comparison queries rather than reward observations are used. We provide and analyze algorithms in two settings: linear and neural function approximation, PG-RLHF and NN-PG-RLHF, respectively.

Offline preference optimization allows fine-tuning large models directly from offline data, and has proved effective in recent alignment practices. We propose generalized preference optimization (GPO), a family of offline losses parameterized by a general class of convex functions. GPO enables a unified view over preference optimization, encompassing existing algorithms such as DPO, IPO and SLiC as special cases, while naturally introducing new variants. The GPO framework also sheds light on how offline algorithms enforce regularization, through the design of the convex function that defines the loss. Our analysis and experiments reveal the connections and subtle differences between the offline regularization and the KL divergence regularization intended by the canonical RLHF formulation. In a controlled setting akin to Gao et al 2023, we also show that different GPO variants achieve similar trade-offs between regularization and performance, though the optimal values of hyper-parameter might differ as predicted by theory. In all, our results present new algorithmic toolkits and empirical insights to alignment practitioners.

Large Language Models (LLMs) drive current AI breakthroughs despite very little being known about their internal representations. In this work, we propose to shed the light on LLMs inner mechanisms through the lens of geometry. In particular, we develop in closed form $(i)$ the intrinsic dimension in which the Multi-Head Attention embeddings are constrained to exist and $(ii)$ the partition and per-region affine mappings of the feedforward (MLP) network of LLMs' layers. Our theoretical findings further enable the design of novel principled solutions applicable to state-of-the-art LLMs. First, we show that, through our geometric understanding, we can bypass LLMs' RLHF protection by controlling the embedding's intrinsic dimension through informed prompt manipulation. Second, we derive interpretable geometrical features that can be extracted from any (pre-trained) LLM, providing a rich abstract representation of their inputs. We observe that these features are sufficient to help solve toxicity detection, and even allow the identification of various types of toxicity. Our results demonstrate how, even in large-scale regimes, exact theoretical results can answer practical questions in LLMs. Code: https://github.com/RandallBalestriero/SplineLLM

How do large language models (LLMs) obtain their answers? The ability to explain and control an LLM’s reasoning process is key for reliability, transparency, and future model developments. We propose SelfIE (Self-Interpretation of Embeddings), a framework that enables LLMs to interpret their own embeddings in natural language by leveraging their ability to respond to inquiries about a given passage. Capable of interpreting open-world concepts in the hidden embeddings, SelfIE reveals LLM internal reasoning in cases such as making ethical decisions, internalizing prompt injection, and recalling harmful knowledge. SelfIE’s text descriptions on hidden embeddings open avenues to control LLM reasoning. We propose Supervised Control, which allows editing open-ended concepts while only requiring gradient computation of individual layer. We extend RLHF to hidden embeddings and propose Reinforcement Control that erases harmful knowledge in LLM without supervision targets.

AI assistants powered by Large Language Models (LLMs) have demonstrated impressive performance in various tasks. However, LLMs still make factual errors in knowledge-intensive tasks such as open-domain question answering. These untruthful responses from AI assistants can pose significant risks in practical applications. Therefore, in this paper, we ask the question **Can AI assistants know what they don't know and express this awareness through natural language?** To investigate this, we construct a model-specific "I don't know" (Idk) dataset. This dataset includes Supervised Fine-tuning data and preference data, categorizing questions based on whether the assistant knows or does not know the answers. Then, we align the assistant with its corresponding Idk dataset using different alignment methods, including Supervised Fine-tuning and preference optimization. Experimental results show that, after alignment with the Idk dataset, the assistant is more capable of declining to answer questions outside its knowledge scope. The assistant aligned with the Idk dataset shows significantly higher truthfulness than the original assistant.

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.

With increased power and prevalence of AI systems, it is ever more critical that AI systems are designed to serve *all*, i.e., people with diverse values and perspectives. However, aligning models to serve *pluralistic* human values remains an open research question. In this piece, we propose a roadmap to pluralistic alignment, specifically using large language models as a test bed. We identify and formalize three possible ways to define and operationalize pluralism in AI systems: 1) *Overton pluralistic* models that present a spectrum of reasonable responses; 2) *Steerably pluralistic* models that can steer to reflect certain perspectives; and 3) *Distributionally pluralistic* models that are well-calibrated to a given population in distribution. We also formalize and discuss three possible classes of *pluralistic benchmarks*: 1) *Multi-objective* benchmarks, 2) *Trade-off steerable* benchmarks that incentivize models to steer to arbitrary trade-offs, and 3) *Jury-pluralistic* benchmarks that explicitly model diverse human ratings. We use this framework to argue that current alignment techniques may be fundamentally limited for pluralistic AI; indeed, we highlight empirical evidence, both from our own experiments and from other work, that standard alignment procedures might *reduce* distributional pluralism in models, motivating the need for further research on pluralistic alignment.