ICML 2024 Papers

The second-order properties of the training loss have a massive impact on the optimization dynamics of deep learning models. Fort & Scherlis (2019) discovered that a large excess of positive curvature and local convexity of the loss Hessian is associated with highly trainable initial points located in a region coined the "Goldilocks zone". Only a handful of subsequent studies touched upon this relationship, so it remains largely unexplained. In this paper, we present a rigorous and comprehensive analysis of the Goldilocks zone for homogeneous neural networks. In particular, we derive the fundamental condition resulting in excess of positive curvature of the loss, explaining and refining its conventionally accepted connection to the initialization norm. Further, we relate the excess of positive curvature to model confidence, low initial loss, and a previously unknown type of vanishing cross-entropy loss gradient. To understand the importance of excessive positive curvature for trainability of deep networks, we optimize fully-connected and convolutional architectures outside the Goldilocks zone and analyze the emergent behaviors. We find that strong model performance is not perfectly aligned with the Goldilocks zone, calling for further research into this relationship.

Single domain generalization (single DG) aims at learning a robust model generalizable to unseen domains from only one training domain, making it a highly ambitious and challenging task. State-of-the-art approaches have mostly relied on data augmentations, such as adversarial perturbation and style enhancement, to synthesize new data and thus increase robustness. Nevertheless, they have largely overlooked the underlying coherence between the augmented domains, which in turn leads to inferior results in real-world scenarios. In this paper, we propose a simple yet effective scheme, termed as *StyDeSty*, to explicitly account for the alignment of the source and pseudo domains in the process of data augmentation, enabling them to interact with each other in a self-consistent manner and further giving rise to a latent domain with strong generalization power. The heart of StyDeSty lies in the interaction between a *stylization* module for generating novel stylized samples using the source domain, and a *destylization* module for transferring stylized and source samples to a latent domain to learn content-invariant features. The stylization and destylization modules work adversarially and reinforce each other. During inference, the destylization module transforms the input sample with an arbitrary style shift to the latent domain, in which the downstream tasks are carried out. Specifically, the location of the destylization layer within the backbone network is determined by a dedicated neural architecture search (NAS) strategy. We evaluate StyDeSty on multiple benchmarks and demonstrate that it yields encouraging results, outperforming the state of the art by up to 13.44% on classification accuracy. Codes are available https://github.com/Huage001/StyDeSty.

Understanding what makes high-dimensional data learnable is a fundamental question in machine learning. On the one hand, it is believed that the success of deep learning lies in its ability to build a hierarchy of representations that become increasingly more abstract with depth, going from simple features like edges to more complex concepts. On the other hand, learning to be insensitive to invariances of the task, such as smooth transformations for image datasets, has been argued to be important for deep networks and it strongly correlates with their performance. In this work, we aim to explain this correlation and unify these two viewpoints. We show that by introducing sparsity to generative hierarchical models of data, the task acquires insensitivity to spatial transformations that are discrete versions of smooth transformations. In particular, we introduce the Sparse Random Hierarchy Model (SRHM), where we observe and rationalize that a hierarchical representation mirroring the hierarchical model is learnt precisely when such insensitivity is learnt, thereby explaining the strong correlation between the latter and performance. Moreover, we quantify how the sample complexity of CNNs learning the SRHM depends on both the sparsity and hierarchical structure of the task.

As large language models (LLMs) become more capable, fine-tuning techniques for aligning with human intent are increasingly important. A key consideration for aligning these models is how to most effectively use human resources, or model resources in the case where LLMs themselves are used as oracles. Reinforcement learning from Human or AI preferences (RLHF/RLAIF) is the most prominent example of such a technique, but is complex and often unstable. Direct Preference Optimization (DPO) has recently been proposed as a simpler and more stable alternative. In this work, we develop an active learning strategy for DPO to make better use of preference labels. We propose a practical acquisition function for prompt/completion pairs based on the predictive entropy of the language model and a measure of certainty of the implicit preference model optimized by DPO. We demonstrate how our approach improves both the rate of learning and final performance of fine-tuning on pairwise preference data.

Finite-sum optimization has wide applications in machine learning, covering important problems such as support vector machines, regression, etc. In this paper, we initiate the study of solving finite-sum optimization problems by quantum computing. Specifically, let $f_1,\ldots,f_n:\mathbb{R}^d\to\mathbb{R}$ be $\ell$-smooth convex functions and $\psi:\mathbb{R}^d\to\mathbb{R}$ be a $\mu$-strongly convex proximal function. The goal is to find an $\epsilon$-optimal point for $F(\mathbf{x})=\frac{1}{n}\sum_{i=1}^n f_i(\mathbf{x})+\psi(\mathbf{x})$. We give a quantum algorithm with complexity $\tilde{O}\big(n+\sqrt{d}+\sqrt{\ell/\mu}\big(n^{1/3}d^{1/3}+n^{-2/3}d^{5/6}\big)\big)$, improving the classical tight bound $\tilde{\Theta}\big(n+\sqrt{n\ell/\mu}\big)$. We also prove a quantum lower bound $\tilde{\Omega}(n+n^{3/4}(\ell/\mu)^{1/4})$ when $d$ is large enough. Both our quantum upper and lower bounds can extend to the cases where $\psi$ is not necessarily strongly convex, or each $f_i$ is Lipschitz but not necessarily smooth. In addition, when $F$ is nonconvex, our quantum algorithm can find an $\epsilon$-critial point using $\tilde{O}(n+\ell(d^{1/3}n^{1/3}+\sqrt{d})/\epsilon^2)$ queries.

Recent works like Tree-of-Thought (ToT) and Reasoning via Planning (RAP) aim to augment the multi-step reasoning capabilities of LLMs by using tree-search algorithms. These methods rely on prompting a pre-trained model to serve as a value function and focus on problems with low search depth. As a result, these methods cannot benefit from in-domain training and only rely on pretraining process — they will not work in domains where the pre-trained LLM does not have enough knowledge to serve as an effective value function or in domains that require long-horizon planning. To address these limitations, we present an AlphaZero-like tree-search learning framework for LLMs (termed TS-LLM), systematically illustrating how tree-search with a learned value function can guide LLM decoding. TS-LLM distinguishes itself in two key ways. (1) Leveraging a learned value function and AlphaZero-like algorithms, our approach can be generally adaptable to a wide range of tasks, language models of any size, and tasks of varying search depths. (2) Our approach can guide LLMs during both inference and training, iteratively improving the LLMs. Empirical results across reasoning, planning, alignment, and decision-making tasks show that TS-LLM outperforms existing approaches and can handle trees with a depth of 64.

This paper explores the connections between tempering (for Sequential Monte Carlo; SMC) and entropic mirror descent to sample from a target probability distribution whose unnormalized density is known. We establish that tempering SMC corresponds to entropic mirror descent applied to the reverse Kullback-Leibler (KL) divergence and obtain convergence rates for the tempering iterates. Our result motivates the tempering iterates from an optimization point of view, showing that tempering can be seen as a descent scheme of the KL divergence with respect to the Fisher-Rao geometry, in contrast to Langevin dynamics that perform descent of the KL with respect to the Wasserstein-2 geometry. We exploit the connection between tempering and mirror descent iterates to justify common practices in SMC and derive adaptive tempering rules that improve over other alternative benchmarks in the literature.

In light of recent advancements in AI capabilities and the increasingly widespread integration of AI systems into society, governments worldwide are actively seeking to mitigate the potential harms and risks associated with these technologies through regulation and other governance tools. However, there exist significant gaps between governance aspirations and the current state of the technical tooling necessary for their realisation. In this position paper, we survey policy documents published by public-sector institutions in the EU, US, and China to highlight specific areas of disconnect between the technical requirements necessary for enacting proposed policy actions, and the current technical state of the art. Our analysis motivates a call for tighter integration of the AI/ML research community within AI governance in order to i) catalyse technical research aimed at bridging the gap between current and supposed technical underpinnings of regulatory action, as well as ii) increase the level of technical expertise within governing institutions so as to inform and guide effective governance of AI.

Large language models generate high-quality responses with potential misinformation, underscoring the need for regulation by distinguishing AI-generated and human-written texts. Watermarking is pivotal in this context, which involves embedding hidden markers in texts during the LLM inference phase, which is imperceptible to humans. Achieving both the detectability of inserted watermarks and the semantic quality of generated texts is challenging. While current watermarking algorithms have made promising progress in this direction, there remains significant scope for improvement. To address these challenges, we introduce a novel multi-objective optimization (MOO) approach for watermarking that utilizes lightweight networks to generate token-specific watermarking logits and splitting ratios. By leveraging MOO to optimize for both detection and semantic objective functions, our method simultaneously achieves detectability and semantic integrity. Experimental results show that our method outperforms current watermarking techniques in enhancing the detectability of texts generated by LLMs while maintaining their semantic coherence. Our code is available at https://github.com/mignonjia/TS_watermark.

We study the problem of learning Single-Index Models under the $L_2^2$ loss in the agnostic model. We give an efficient learning algorithm, achieving a constant factor approximation to the optimal loss, that succeeds under a range of distributions (including log-concave distributions) and a broad class of monotone and Lipschitz link functions. This is the first efficient constant factor approximate agnostic learner, even for Gaussian data and for any nontrivial class of link functions. Prior work for the case of unknown link function either works in the realizable setting or does not attain constant factor approximation. The main technical ingredient enabling our algorithm and analysis is a novel notion of a local error bound in optimization that we term *alignment sharpness* and that may be of broader interest.

In this paper, we present two key theorems that should have significant implications for machine learning practitioners working with binary classification models. The first theorem provides a formula to calculate the maximum and minimum Precision-Recall AUC ($AUC_{PR}$) for a fixed Receiver Operating Characteristic AUC ($AUC_{ROC}$), demonstrating the variability of $AUC_{PR}$ even with a high $AUC_{ROC}$. This is particularly relevant for imbalanced datasets, where a good $AUC_{ROC}$ does not necessarily imply a high $AUC_{PR}$. The second theorem inversely establishes the bounds of $AUC_{ROC}$ given a fixed $AUC_{PR}$. Our findings highlight that in certain situations, especially for imbalanced datasets, it is more informative to prioritize $AUC_{PR}$ over $AUC_{ROC}$. Additionally, we introduce a method to determine when a higher $AUC_{ROC}$ in one model implies a higher $AUC_{PR}$ in another and vice versa, streamlining the model evaluation process.

This work focuses on the training dynamics of one associative memory module storing outer products of token embeddings. We reduce this problem to the study of a system of particles, which interact according to properties of the data distribution and correlations between embeddings. Through theory and experiments, we provide several insights. In overparameterized regimes, we obtain logarithmic growth of the ``classification margins.'' Yet, we show that imbalance in token frequencies and memory interferences due to correlated embeddings lead to oscillatory transitory regimes. The oscillations are more pronounced with large step sizes, which can create benign loss spikes, although these learning rates speed up the dynamics and accelerate the asymptotic convergence. We also find that underparameterized regimes lead to suboptimal memorization schemes. Finally, we assess the validity of our findings on small Transformer models.

Reinforcement Learning from Human Feedback (RLHF) has shown potential in qualitative tasks where easily defined performance measures are lacking. However, there are drawbacks when RLHF is commonly used to optimize for average human preferences, especially in generative tasks that demand diverse model responses. Meanwhile, Quality Diversity (QD) algorithms excel at identifying diverse and high-quality solutions but often rely on manually crafted diversity metrics. This paper introduces Quality Diversity through Human Feedback (QDHF), a novel approach that progressively infers diversity metrics from human judgments of similarity among solutions, thereby enhancing the applicability and effectiveness of QD algorithms in complex and open-ended domains. Empirical studies show that QDHF significantly outperforms state-of-the-art methods in automatic diversity discovery and matches the efficacy of QD with manually crafted diversity metrics on standard benchmarks in robotics and reinforcement learning. Notably, in open-ended generative tasks, QDHF substantially enhances the diversity of text-to-image generation from a diffusion model and is more favorably received in user studies. We conclude by analyzing QDHF's scalability, robustness, and quality of derived diversity metrics, emphasizing its strength in open-ended optimization tasks. Code and tutorials are available at https://liding.info/qdhf.

Training modern LLMs is extremely resource intensive, and customizing them for various deployment scenarios characterized by limited compute and memory resources through repeated training is impractical. In this paper, we introduce Flextron, a network architecture and post-training model optimization framework supporting flexible model deployment. The Flextron architecture utilizes a nested elastic structure to rapidly adapt to specific user-defined latency and accuracy targets during inference with no additional fine-tuning required. It is also input-adaptive, and can automatically route tokens through its sub-networks for improved performance and efficiency. We present a sample-efficient training method and associated routing algorithms for systematically transforming an existing trained LLM into a Flextron model. We evaluate Flextron on the GPT-3 and LLama-2 family of LLMs, and demonstrate superior performance over multiple end-to-end trained variants and other state-of-the-art elastic networks, all with a single pretraining run that consumes a mere 7.63% tokens compared to original pretraining.

Pretrained deep-learning models are the go-to solution for images or text. However, for tabular data the standard is still to train tree-based models. Indeed, transfer learning on tables hits the challenge of *data integration*: finding correspondences, correspondences in the entries (*entity matching*) where different words may denote the same entity, correspondences across columns (*schema matching*), which may come in different orders, names... We propose a neural architecture that does not need such correspondences. As a result, we can pretrain it on background data that has not been matched. The architecture --CARTE for Context Aware Representation of Table Entries-- uses a graph representation of tabular (or relational) data to process tables with different columns, string embedding of entries and columns names to model an open vocabulary, and a graph-attentional network to contextualize entries with column names and neighboring entries. An extensive benchmark shows that CARTE facilitates learning, outperforming a solid set of baselines including the best tree-based models. CARTE also enables joint learning across tables with unmatched columns, enhancing a small table with bigger ones. CARTE opens the door to large pretrained models for tabular data.

Post-training quantization (PTQ) reduces the memory footprint of LLMs by quantizing their weights to low-precision. In this work, we introduce QuIP#, a weight-only PTQ method that achieves state-of-the-art results in extreme compression regimes ($\le$ 4 bits per weight) using three novel techniques. First, QuIP# improves QuIP's (Chee et al., 2023) incoherence processing by using the randomized Hadamard transform, which is faster and has better theoretical properties. Second, QuIP# uses vector quantization to take advantage of the ball-shaped sub-Gaussian distribution that incoherent weights possess: specifically, we introduce a set of hardware-efficient codebooks based on the highly symmetric $E_8$ lattice, which achieves the optimal 8-dimension unit ball packing. Third, QuIP# uses fine-tuning to improve fidelity to the original model. Our experiments show that QuIP# outperforms existing PTQ methods, enables new behaviors in PTQ scaling, and supports fast inference. Our code can be found at https://github.com/Cornell-RelaxML/quip-sharp.

Large language models (LLMs) are trained to imitate humans to explain human decisions. However, do LLMs explain themselves? Can they help humans build mental models of how LLMs process different inputs? To answer these questions, we propose to evaluate $\textbf{counterfactual simulatability}$ of natural language explanations: whether an explanation can enable humans to precisely infer the model's outputs on diverse counterfactuals of the explained input. For example, if a model answers ''$\textit{yes}$'' to the input question ''$\textit{Can eagles fly?}$'' with the explanation ''$\textit{all birds can fly}$'', then humans would infer from the explanation that it would also answer ''$\textit{yes}$'' to the counterfactual input ''$\textit{Can penguins fly?}$''. If the explanation is precise, then the model's answer should match humans' expectations. We implemented two metrics based on counterfactual simulatability: precision and generality. We generated diverse counterfactuals automatically using LLMs. We then used these metrics to evaluate state-of-the-art LLMs (e.g., GPT-4) on two tasks: multi-hop factual reasoning and reward modeling. We found that LLM's explanations have low precision and that precision does not correlate with plausibility. Therefore, naively optimizing human approvals (e.g., RLHF) may be insufficient.

Reinforcement Learning from Human Feedback (RLHF) aligns language models to human preferences by employing a singular reward model derived from preference data. However, the single reward model overlooks the rich diversity of human preferences inherent in data collected from multiple users. In this work, we first derive an impossibility result of alignment with single reward RLHF, thereby highlighting its insufficiency in representing diverse human preferences. Next, we propose to learn a mixture of reward models via an expectation-maximization algorithm and solve a MaxMin alignment objective inspired by the Egalitarian principle in social choice theory to better honor diverse human preferences. We present comprehensive experimental results on small-scale (GPT-2) and large-scale language (with Tulu2-7B)) and show the efficacy of the proposed approach in the presence of diversity among human preferences. We remark that our findings in this work are not only limited to language models but also extend to reinforcement learning in general.

In the next few years, applications of Generative AI are expected to revolutionize a number of different areas, ranging from science & medicine to education. The potential for these seismic changes has triggered a lively debate about potential risks and resulted in calls for tighter regulation, in particular from some of the major tech companies who are leading in AI development. While regulation is important, it is key that it does not put at risk the budding field of open-source Generative AI. We argue for the responsible open sourcing of generative AI models in the near and medium term. To set the stage, we first introduce an AI openness taxonomy system and apply it to 40 current large language models. We then outline differential benefits and risks of open versus closed source AI and present potential risk mitigation, ranging from best practices to calls for technical and scientific contributions. We hope that this report will add a much needed missing voice to the current public discourse on near to mid-term AI safety and other societal impact.