ICML 2024 Papers

Previous work on fairness in federated learning introduced the notion of *core stability*, which provides utility-based fairness guarantees to any subset of participating agents. However, these guarantees require strong assumptions on agent utilities that render them impractical. To address this shortcoming, we measure the quality of output models in terms of their ordinal *rank* instead of their cardinal utility, and use this insight to adapt the classical notion of *proportional veto core (PVC)* from social choice theory to the federated learning setting. We prove that models that are *PVC-stable* exist in very general learning paradigms, even allowing non-convex model sets, as well as non-convex and non-concave loss functions. We also design Rank-Core-Fed, a distributed federated learning algorithm, to train a PVC-stable model. Finally, we demonstrate that Rank-Core-Fed outperforms baselines in terms of fairness on different datasets.

Reinforcement Learning from Human Feedback (RLHF) is currently the most widely used method to align large language models (LLMs) with human preferences. Existing RLHF methods can be roughly categorized as either reward-based or reward-free. Novel applications such as ChatGPT and Claude leverage reward-based methods that first learn a reward model and apply actor-critic algorithms, such as Proximal Policy Optimization (PPO). However, in academic benchmarks, state-of-the-art results are often achieved via reward-free methods, such as Direct Preference Optimization (DPO). Is DPO truly superior to PPO? Why does PPO perform poorly on these benchmarks? In this paper, we first conduct both theoretical and empirical studies on the algorithmic properties of DPO and show that DPO may have fundamental limitations. Moreover, we also comprehensively examine PPO and reveal the key factors for the best performances of PPO in fine-tuning LLMs. Finally, we benchmark DPO and PPO across a collection of RLHF testbeds, ranging from dialogue to code generation. Experiment results demonstrate that PPO is able to surpass other alignment methods in all cases and achieve state-of-the-art results in challenging code competitions.

This paper proposes the use of "multicalibration": to yield interpretable and reliable confidence scores for outputs generated by large language models (LLMs). Multicalibration asks for calibration not just marginally, but simultaneously across various intersecting groupings of the data. We show how to form groupings for prompt/completion pairs that are correlated with the probability of correctness via two techniques: clustering within an embedding space, and "self-annotation" - querying the LLM by asking it various yes-or-no questions about the prompt. We also develop novel variants of multicalibration algorithms that offer performance improvements by reducing their tendency to overfit. Through systematic benchmarking across various question answering datasets and LLMs, we show how our techniques can yield confidence scores that provide substantial improvements in fine-grained measures of both calibration and accuracy compared to existing methods.

We propose an approach to do learning in Gaussian factor graphs. We treat all relevant quantities (inputs, outputs, parameters, activations) as random variables in a graphical model, and view training and prediction as inference problems with different observed nodes. Our experiments show that these problems can be efficiently solved with belief propagation (BP), whose updates are inherently local, presenting exciting opportunities for distributed and asynchronous training. Our approach can be scaled to deep networks and provides a natural means to do continual learning: use the BP-estimated posterior of the current task as a prior for the next. On a video denoising task we demonstrate the benefit of learnable parameters over a classical factor graph approach and we show encouraging performance of deep factor graphs for continual image classification.

Ensuring privacy-preserving inference on cryptographically secure data is a well-known computational challenge. To alleviate the bottleneck of costly cryptographic computations in non-linear activations, recent methods have suggested linearizing a targeted portion of these activations in neural networks. This technique results in significantly reduced runtimes with often negligible impacts on accuracy. In this paper, we demonstrate that such computational benefits may lead to increased fairness costs. Specifically, we find that reducing the number of ReLU activations disproportionately decreases the accuracy for minority groups compared to majority groups. To explain these observations, we provide a mathematical interpretation under restricted assumptions about the nature of the decision boundary, while also showing the prevalence of this problem across widely used datasets and architectures. Finally, we show how a simple procedure altering the finetuning step for linearized models can serve as an effective mitigation strategy.

We present Langevin Predictive Coding (LPC), a novel algorithm for deep generative model learning that builds upon the predictive coding framework of computational neuroscience. By injecting Gaussian noise into the predictive coding inference procedure and incorporating an encoder network initialization, we reframe the approach as an amortized Langevin sampling method for optimizing a tight variational lower bound. To increase robustness to sampling step size, we present a lightweight preconditioning technique inspired by Riemannian Langevin methods and adaptive SGD. We compare LPC against VAEs by training generative models on benchmark datasets; our experiments demonstrate superior sample quality and faster convergence for LPC in a fraction of SGD training iterations, while matching or exceeding VAE performance across key metrics like FID, diversity and coverage.

Does the stethoscope in the picture make the adjacent person a doctor or a patient? This, of course, depends on the contextual relationship of the two objects. If it’s obvious, why don’t explanation methods for vision models use contextual information? In this paper, we (1) review the most popular methods of explaining computer vision models by pointing out that they do not take into account context information, (2) show examples of failures of popular XAI methods, (3) provide examples of real-world use cases where spatial context plays a significant role, (4) propose new research directions that may lead to better use of context information in explaining computer vision models, (5) argue that a change in approach to explanations is needed from *where* to *how*.

Albeit the success of federated learning (FL) in decentralized training, bolstering the generalization of models by overcoming heterogeneity across clients still remains a huge challenge. To aim at improved generalization of FL, a group of recent works pursues flatter minima of models by employing sharpness-aware minimization in the local training at the client side. However, we observe that the global model, i.e., the aggregated model, does not lie on flat minima of the global objective, even with the effort of flatness searching in local training, which we define as flatness discrepancy. By rethinking and theoretically analyzing flatness searching in FL through the lens of the discrepancy problem, we propose a method called Federated Learning for Global Flatness (FedGF) that explicitly pursues the flatter minima of the global models, leading to the relieved flatness discrepancy and remarkable performance gains in the heterogeneous FL benchmarks.

In coming up with solutions to real-world problems, humans implicitly adhere to constraints that are too numerous and complex to be specified completely. However, reinforcement learning (RL) agents need these constraints to learn the correct optimal policy in these settings. The field of Inverse Constraint Reinforcement Learning (ICRL) deals with this problem and provides algorithms that aim to estimate the constraints from expert demonstrations collected offline. Practitioners prefer to know a measure of confidence in the estimated constraints, before deciding to use these constraints, which allows them to only use the constraints that satisfy a desired level of confidence. However, prior works do not allow users to provide the desired level of confidence for the inferred constraints. This work provides a principled ICRL method that can take a confidence level with a set of expert demonstrations and outputs a constraint that is at least as constraining as the true underlying constraint with the desired level of confidence. Further, unlike previous methods, this method allows a user to know if the number of expert trajectories is insufficient to learn a constraint with a desired level of confidence, and therefore collect more expert trajectories as required to simultaneously learn constraints with the desired level of confidence and a policy that achieves the desired level of performance.

Black-box finetuning is an emerging interface for adapting state-of-the-art language models to user needs. However, such access may also let malicious actors undermine model safety. To demonstrate the challenge of defending finetuning interfaces, we introduce covert malicious finetuning, a method to compromise model safety via finetuning while evading detection. Our method constructs a malicious dataset where every individual datapoint appears innocuous, but finetuning on the dataset teaches the model to respond to encoded harmful requests with encoded harmful responses. Applied to GPT-4, our method produces a finetuned model that acts on harmful instructions 99% of the time and avoids detection by defense mechanisms such as dataset inspection, safety evaluations, and input/output classifiers. Our findings question whether black-box finetuning access can be secured against sophisticated adversaries.

Many existing anomaly detection methods assume the availability of a large-scale normal dataset. But for many applications, limited by resources, removing all anomalous samples from a large un-labeled dataset is unrealistic, resulting in contaminated datasets. To detect anomalies accurately under such scenarios, from the probabilistic perspective, the key question becomes how to learn the normal-data distribution from a contaminated dataset. To this end, we propose to collect two additional small datasets that are comprised of partially-observed normal and anomaly samples, and then use them to help learn the distribution under an adversarial learning scheme. We prove that under some mild conditions, the proposed method is able to learn the correct normal-data distribution. Then, we consider the overfitting issue caused by the small size of the two additional datasets, and a correctness-guaranteed flipping mechanism is further developed to alleviate it. Theoretical results under incomplete observed anomaly types are also presented. Extensive experimental results demonstrate that our method outperforms representative baselines when detecting anomalies under contaminated datasets.

Pre-trained Vision-Language Models (VLMs) are able to understand visual concepts, describe and decompose complex tasks into sub-tasks, and provide feedback on task completion. In this paper, we aim to leverage these capabilities to support the training of reinforcement learning (RL) agents. In principle, VLMs are well suited for this purpose, as they can naturally analyze image-based observations and provide feedback (reward) on learning progress. However, inference in VLMs is computationally expensive, so querying them frequently to compute rewards would significantly slowdown the training of an RL agent. To address this challenge, we propose a framework named Code as Reward (VLM-CaR). VLM-CaR produces dense reward functions from VLMs through code generation, thereby significantly reducing the computational burden of querying the VLM directly. We show that the dense rewards generated through our approach are very accurate across a diverse set of discrete and continuous environments, and can be more effective in training RL policies than the original sparse environment rewards.

Diffusion models have made significant contributions to computer vision, sparking a growing interest in the community recently regarding the application of it to graph generation. The existing discrete graph diffusion models exhibit heightened computational complexity and diminished training efficiency. A preferable and natural way is to directly diffuse the graph within the latent space. However, due to the non-Euclidean structure of graphs is not isotropic in the latent space, the existing latent diffusion models effectively make it difficult to capture and preserve the topological information of graphs. To address the above challenges, we propose a novel geometrically latent diffusion framework HypDiff. Specifically, we first establish a geometrically latent space with interpretability measures based on hyperbolic geometry, to define anisotropic latent diffusion processes for graphs. Then, we propose a geometrically latent diffusion process that is constrained by both radial and angular geometric properties, thereby ensuring the preservation of the original topological properties in the generative graphs. Extensive experimental results demonstrate the superior effectiveness of HypDiff for graph generation with various topologies.

In this paper, we establish a novel connection between total variation (TV) distance estimation and probabilistic inference. In particular, we present an efficient, structure-preserving reduction from relative approximation of TV distance to probabilistic inference over directed graphical models. This reduction leads to a fully polynomial randomized approximation scheme (FPRAS) for estimating TV distances between same-structure distributions over any class of Bayes nets for which there is an efficient probabilistic inference algorithm. In particular, it leads to an FPRAS for estimating TV distances between distributions that are defined over a common Bayes net of small treewidth. Prior to this work, such approximation schemes only existed for estimating TV distances between product distributions. Our approach employs a new notion of *partial* couplings of high-dimensional distributions, which might be of independent interest.

We leverage convex and bilevel optimization techniques to develop a general gradient-based parameter learning framework for neural-symbolic (NeSy) systems. We demonstrate our framework with NeuPSL, a state-of-the-art NeSy architecture. To achieve this, we propose a smooth primal and dual formulation of NeuPSL inference and show learning gradients are functions of the optimal dual variables. Additionally, we develop a dual block coordinate descent algorithm for the new formulation that naturally exploits warm-starts. This leads to over $100 \times$ learning runtime improvements over the current best NeuPSL inference method. Finally, we provide extensive empirical evaluations across $8$ datasets covering a range of tasks and demonstrate our learning framework achieves up to a $16$% point prediction performance improvement over alternative learning methods.

We study the problem of parameter-free stochastic optimization, inquiring whether, and under what conditions, do fully parameter-free methods exist: these are methods that achieve convergence rates competitive with optimally tuned methods, without requiring significant knowledge of the true problem parameters. Existing parameter-free methods can only be considered ``partially'' parameter-free, as they require some non-trivial knowledge of the true problem parameters, such as a bound on the stochastic gradient norms, a bound on the distance to a minimizer, etc. In the non-convex setting, we demonstrate that a simple hyperparameter search technique results in a fully parameter-free method that outperforms more sophisticated state-of-the-art algorithms. We also provide a similar result in the convex setting with access to noisy function values under mild noise assumptions. Finally, assuming only access to stochastic gradients, we establish a lower bound that renders fully parameter-free stochastic convex optimization infeasible, and provide a method which is (partially) parameter-free up to the limit indicated by our lower bound.

This paper introduces a framework for post-processing machine learning models so that their predictions satisfy multi-group fairness guarantees. Based on the celebrated notion of multicalibration, we introduce $(s,g,\alpha)-$GMC (Generalized Multi-Dimensional Multicalibration) for multi-dimensional mappings $s$, constraints $g$, and a pre-specified threshold level $\alpha$. We propose associated algorithms to achieve this notion in general settings. This framework is then applied to diverse scenarios encompassing different fairness concerns, including false negative rate control in image segmentation, prediction set conditional uncertainty quantification in hierarchical classification, and de-biased text generation in language models. We conduct numerical studies on several datasets and tasks.

Modern language models can contain billions of parameters, raising the question of whether they can generalize beyond the training data or simply parrot their training corpora. We provide the first non-vacuous generalization bounds for pretrained large language models (LLMs), indicating that language models are capable of discovering regularities that generalize to unseen data. In particular, we derive a compression bound that is valid for the unbounded log-likelihood loss using prediction smoothing, and we extend the bound to handle subsampling, making bound computation 900 times faster on massive datasets. To achieve the extreme level of compression required for non-vacuous bounds, we devise SubLoRA, a simple low-dimensional nonlinear parameterization that leads to non-vacuous generalization bounds for very large models with up to 849 million parameters. Finally, we use our bounds to understand LLM generalization and find that larger models have better generalization bounds and are more compressible than smaller models.

Instrumental variable (IV) regression can be approached through its formulation in terms of conditional moment restrictions (CMR). Building on variants of the generalized method of moments, most CMR estimators are implicitly based on approximating the population data distribution via reweightings of the empirical sample. While for large sample sizes, in the independent identically distributed (IID) setting, reweightings can provide sufficient flexibility, they might fail to capture the relevant information in presence of corrupted data or data prone to adversarial attacks. To address these shortcomings, we propose the Sinkhorn Method of Moments, an optimal transport-based IV estimator that takes into account the geometry of the data manifold through data-derivative information. We provide a simple plug-and-play implementation of our method that performs on par with related estimators in standard settings but improves robustness against data corruption and adversarial attacks.