ICML 2024 Papers

In this paper, we show that Low Rank Adaptation (LoRA) as originally introduced in (Hu et al., 2021) leads to suboptimal finetuning of models with large width. This is due to the fact that adapter matrices A and B in LoRA are updated with the same learning rate in ADAM. Using scaling arguments for large width networks, we demonstrate that the same learning rate does not allow efficient feature learning. We then show that this suboptimality of LoRA can be corrected simply by setting different learning rates for the LoRA adapter matrices A and B with a well-chosen fixed ratio. We call this proposed algorithm LoRA+. In our extensive experiments, LoRA+ improves finetuning speed (up to ∼ 2X SpeedUp) and performance (1% − 2% improvements), at the same computational cost as LoRA. The code is available at https://github.com/nikhil-ghosh-berkeley/loraplus

Traveling waves are a fundamental phenomenon in the brain, playing a crucial role in short-term information storage. In this study, we leverage the concept of traveling wave dynamics within a neural lattice to formulate a theoretical model of neural working memory in Recurrent Neural Networks (RNNs), study its properties, and its real world implications in AI. The proposed model diverges from traditional approaches, which assume information storage in static, register-like locations updated by interference. Instead, the model stores data as waves that is updated by the wave's boundary conditions. We rigorously examine the model's capabilities in representing and learning state histories, which are vital for learning history-dependent dynamical systems. The findings reveal that the model reliably stores external information and enhances the learning process by addressing the diminishing gradient problem of RNNs. To understand the model's real-world applicability, we explore two cases: linear boundary condition and non-linear, self-attention-driven boundary condition. The experiments reveal that the linear scenario is effectively *learned* by RNNs through backpropagation when modeling history-dependent dynamical systems. Conversely, the non-linear scenario parallels an attention-only transformer. Collectively, our findings suggest the broader relevance of traveling waves in AI and its potential in advancing neural network architectures.

While Positive-Unlabeled (PU) learning is vital in many real-world scenarios, its application to graph data still remains under-explored. We unveil that a critical challenge for PU learning on graph lies on the edge heterophily, which directly violates the $\textit{irreducibility assumption}$ for $\textit{Class-Prior Estimation}$ (class prior is essential for building PU learning algorithms) and degenerates the latent label inference on unlabeled nodes during classifier training. In response to this challenge, we introduce a new method, named $\textit{$\underline{G}$raph $\underline{P}$U Learning with $\underline{L}$abel Propagation Loss}$ (GPL). Specifically, GPL considers learning from PU nodes along with an intermediate heterophily reduction, which helps mitigate the negative impact of the heterophilic structure. We formulate this procedure as a bilevel optimization that reduces heterophily in the inner loop and efficiently learns a classifier in the outer loop. Extensive experiments across a variety of datasets have shown that GPL significantly outperforms baseline methods, confirming its effectiveness and superiority.

Vector quantization is a fundamental operation for data compression and vector search. To obtain high accuracy, multi-codebook methods represent each vector using codewords across several codebooks. Residual quantization (RQ) is one such method, which iteratively quantizes the error of the previous step. While the error distribution is dependent on previously-selected codewords, this dependency is not accounted for in conventional RQ as it uses a fixed codebook per quantization step. In this paper, we propose QINCo, a neural RQ variant that constructs specialized codebooks per step that depend on the approximation of the vector from previous steps. Experiments show that QINCo outperforms state-of-the-art methods by a large margin on several datasets and code sizes. For example, QINCo achieves better nearest-neighbor search accuracy using 12-byte codes than the state-of-the-art UNQ using 16 bytes on the BigANN1M and Deep1M datasets.

Turning pass-through network architectures into iterative ones, which use their own output as input, is a well-known approach for boosting performance. In this paper, we argue that such architectures offer an additional benefit: The convergence rate of their successive outputs is highly correlated with the accuracy of the value to which they converge. Thus, we can use the convergence rate as a useful proxy for uncertainty. This results in an approach to uncertainty estimation that provides state-of-the-art estimates at a much lower computational cost than techniques like Ensembles, and without requiring any modifications to the original iterative model. We demonstrate its practical value by embedding it in two application domains: road detection in aerial images and the estimation of aerodynamic properties of 2D and 3D shapes.

Efficient sampling of the Boltzmann distribution of molecular systems is a long-standing challenge. Recently, instead of generating long molecular dynamics simulations, generative machine learning methods such as normalizing flows have been used to learn the Boltzmann distribution directly, without samples. However, this approach is susceptible to mode collapse and thus often does not explore the full configurational space. In this work, we address this challenge by separating the problem into two levels, the fine-grained and coarse-grained degrees of freedom. A normalizing flow conditioned on the coarse-grained space yields a probabilistic connection between the two levels. To explore the configurational space, we employ coarse-grained simulations with active learning which allows us to update the flow and make all-atom potential energy evaluations only when necessary. Using alanine dipeptide as an example, we show that our methods obtain a speedup to molecular dynamics simulations of approximately $15.9$ to $216.2$ compared to the speedup of $4.5$ of the current state-of-the-art machine learning approach.

In this paper, we consider a model called *Online Capacitated Coverage Maximization*, characterized by two features: (1) the dynamic arrival of online agents following a known identical and independent distribution, and (2) each offline agent is associated with a specific coverage valuation over the groundset of online agents. Additionally, both offline and online agents are assigned integer capacities, reflecting finite budgets and operational constraints. We introduce and analyze two matching policies. The first, a non-adaptive policy, utilizes offline statistics derived from solving a benchmark linear program. The second is an enhanced version equipped with real-time boostings and attenuations. We conduct a comprehensive competitive analysis and characterize the competitive ratio for both policies as functions of two crucial parameters: a lower bound on the matching capacity among offline agents and an upper bound on the number of online agents covering any specific feature for offline agents.

Dyna-style model-based reinforcement learning (MBRL) combines model-free agents with predictive transition models through model-based rollouts. This combination raises a critical question: “When to trust your model?”; i.e., which rollout length results in the model providing useful data? Janner et al. (2019) address this question by gradually increasing rollout lengths throughout the training. While theoretically tempting, uniform model accuracy is a fallacy that collapses at the latest when extrapolating. Instead, we propose asking the question “Where to trust your model?”. Using inherent model uncertainty to consider local accuracy, we obtain the Model-Based Actor-Critic with Uncertainty-Aware Rollout Adaption (MACURA) algorithm. We propose an easy-to-tune rollout mechanism and demonstrate substantial improvements in data efficiency and performance compared to state-of-the-art deep MBRL methods on the MuJoCo benchmark.

We examine two types of binary betting markets, whose primary goal is for profit (such as sports gambling) or to gain information (such as prediction markets). We articulate the interplay between belief and price-setting to analyse both types of markets, and show that the goals of maximising bookmaker profit and eliciting information are fundamentally incompatible. A key insight is that profit hinges on the deviation between (the distribution of) bettor and true beliefs, and that heavier tails in bettor belief distribution implies higher profit. Our algorithmic contribution is to introduce online learning methods for price-setting. Traditionally bookmakers update their prices rather infrequently, we present two algorithms that guide price updates upon seeing each bet, assuming very little of bettor belief distributions. The online pricing algorithm achieves stochastic regret of $\mathcal{O}(\sqrt{T})$ against the worst local maximum, or $\mathcal{O}(\sqrt{T \log T})$ with high probability against the global maximum under fair odds. More broadly, the inherent tradeoff between profit and information-seeking in binary betting may inspire new understandings of large-scale multi-agent behaviour.

We study the matroid semi-bandits problem, where at each round the learner plays a subset of $K$ arms from a feasible set, and the goal is to maximize the expected cumulative linear rewards. Existing algorithms have per-round time complexity at least $\Omega(K)$, which becomes expensive when $K$ is large. To address this computational issue, we propose FasterCUCB whose sampling rule takes time sublinear in $K$ for common classes of matroids: $\mathcal{O}(D\text{ polylog}(K)\text{ polylog}(T))$ for uniform matroids, partition matroids, and graphical matroids, and $\mathcal{O}(D\sqrt{K}\text{ polylog}(T))$ for transversal matroids. Here, $D$ is the maximum number of elements in any feasible subset of arms, and $T$ is the horizon. Our technique is based on dynamic maintenance of an approximate maximum-weight basis over inner-product weights. Although the introduction of an approximate maximum-weight basis presents a challenge in regret analysis, we can still guarantee an upper bound on regret as tight as CUCB in the sense that it matches the gap-dependent lower bound by Kveton et al. (2014a) asymptotically.

When the training dataset comprises a 1:1 proportion of dogs to cats, a generative model that produces 1:1 dogs and cats better resembles the training species distribution than another model with 3:1 dogs and cats. Can we capture this phenomenon using existing metrics? Unfortunately, we cannot, because these metrics do not provide any interpretability beyond “diversity". In this context, we propose a new evaluation protocol that measures the divergence of a set of generated images from the training set regarding the distribution of attribute strengths as follows. Singleattribute Divergence (SaD) reveals the attributes that are generated excessively or insufficiently by measuring the divergence of PDFs of individual attributes. Paired-attribute Divergence (PaD) reveals such pairs of attributes by measuring the divergence of joint PDFs of pairs of attributes. For measuring the attribute strengths of an image, we propose Heterogeneous CLIPScore (HCS) which measures the cosine similarity between image and text vectors with heterogeneous initial points. With SaD and PaD, we reveal the following about existing generative models. ProjectedGAN generates implausible attribute relationships such as baby with beard even though it has competitive scores of existing metrics. Diffusion models struggle to capture diverse colors in the datasets. The larger sampling timesteps of the latent diffusion model generate the more minor objects including earrings and necklace. Stable Diffusion v1.5 better captures the attributes than v2.1. Our metrics lay a foundation for explainable evaluations of generative models.

We propose a distributional framework for benchmarking socio-technical risks of foundation models with quantified statistical significance. Our approach hinges on a new statistical relative testing based on first and second order stochastic dominance of real random variables. We show that the second order statistics in this test are linked to mean-risk models commonly used in econometrics and mathematical finance to balance risk and utility when choosing between alternatives. Using this framework, we formally develop a risk-aware approach for foundation model selection given guardrails quantified by specified metrics. Inspired by portfolio optimization and selection theory in mathematical finance, we define a metrics portfolio for each model as a means to aggregate a collection of metrics, and perform model selection based on the stochastic dominance of these portfolios. The statistical significance of our tests is backed theoretically by an asymptotic analysis via central limit theorems instantiated in practice via a bootstrap variance estimate. We use our framework to compare various large language models regarding risks related to drifting from instructions and outputting toxic content.

Subset or core-set selection offers a data-efficient way for training deep learning models. One-shot subset selection poses additional challenges as subset selection is only performed once and full set data become unavailable after the selection. However, most existing methods tend to choose either diverse or difficult data samples, which fail to faithfully represent the joint data distribution that is comprised of both feature and label information. The selection is also performed independently from the subset size, which plays an essential role in choosing what types of samples. To address this critical gap, we propose to conduct Feature similarity and Label variability Balanced One-shot Subset Selection (BOSS), aiming to construct an optimal size-aware subset for data-efficient deep learning. We show that a novel balanced core-set loss bound theoretically justifies the need to simultaneously consider both diversity and difficulty to form an optimal subset. It also reveals how the subset size influences the bound. We further connect the inaccessible bound to a practical surrogate target which is tailored to subset sizes and varying levels of overall difficulty. We design a novel Beta-scoring importance function to delicately control the optimal balance of diversity and difficulty. Comprehensive experiments conducted on both synthetic and real data justify the important theoretical properties and demonstrate the superior performance of BOSS as compared with the competitive baselines.

Recent diffusion models have manifested extraordinary capabilities in generating high-quality, diverse, and innovative images guided by textual prompts. Nevertheless, these state-of-the-art models may encounter the challenge of concept bleeding when generating images with multiple entities or attributes in the prompt, leading to the unanticipated merging or overlapping of distinct objects in the synthesized result. The current work exploits auxiliary networks to produce mask-constrained regions for entities, necessitating the training of an object detection network. In this paper, we investigate the bleeding reason and find that the cross-attention map associated with a specific entity or attribute tends to extend beyond its intended focus, encompassing the background or other unrelated objects and thereby acting as the primary source of concept bleeding. Motivated by this, we propose Entity Localization and Anchoring (ELA) to drive the entity to concentrate on the expected region accurately during inference, eliminating the necessity for training. Specifically, we initially identify the region corresponding to each entity and subsequently employ a tailored loss function to anchor entities within their designated positioning areas. Extensive experiments demonstrate its superior capability in precisely generating multiple objects as specified in the textual prompts.

Local feature selection in machine learning provides instance-specific explanations by focusing on the most relevant features for each prediction, enhancing the interpretability of complex models. However, such methods tend to produce misleading explanations by encoding additional information in their selections. In this work, we attribute the problem of misleading selections by formalizing the concepts of label and feature leakage. We rigorously derive the necessary and sufficient conditions under which we can guarantee no leakage, and show existing methods do not meet these conditions. Furthermore, we propose the first local feature selection method that is proven to have no leakage called SUWR. Our experimental results indicate that SUWR is less prone to overfitting and combines state-of-the-art predictive performance with high feature-selection sparsity. Our generic and easily extendable formal approach provides a strong theoretical basis for future work on interpretability with reliable explanations.

Conformal prediction (CP) can convert any model's output into prediction sets guaranteed to include the true label with any user-specified probability. However, same as the model itself, CP is vulnerable to adversarial test examples (evasion) and perturbed calibration data (poisoning). We derive provably robust sets by bounding the worst-case change in conformity scores. Our tighter bounds lead to more efficient sets. We cover both continuous and discrete (sparse) data and our guarantees work both for evasion and poisoning attacks (on both features and labels).

Vision-language models (VLMs), such as CLIP, have demonstrated impressive zero-shot capabilities for various downstream tasks. Their performance can be further enhanced through few-shot prompt tuning methods. However, current studies evaluate the performance of learned prompts separately on base and new classes. This evaluation lacks practicality for real-world applications since downstream tasks cannot determine whether the data belongs to base or new classes in advance. In this paper, we explore a problem setting called ***O**pen-world **P**rompt **T**uning* (OPT), which involves tuning prompts on base classes and evaluating on a combination of base and new classes. By introducing ***De**composed **P**rompt **T**uning* framework (DePT), we theoretically demonstrate that OPT can be solved by incorporating out-of-distribution detection into prompt tuning, thereby enhancing the base-to-new discriminability. Based on DePT, we present a novel prompt tuning approach, namely, ***De**composed **Co**ntext **Op**timization* (DeCoOp), which introduces new-class detectors and sub-classifiers to further enhance the base-class and new-class discriminability. Experimental results on 11 benchmark datasets validate the effectiveness of DePT and demonstrate that DeCoOp outperforms current state-of-the-art methods, providing a significant 2% average accuracy improvement.

Recent years have witnessed a paradigm shift in deep learning from task-centric model design to task-agnostic representation learning and task-specific fine-tuning. Pretrained model representations are commonly evaluated extensively across various real-world tasks and used as a foundation for different downstream tasks. This paper proposes a solution for assessing the quality of representations in a task-agnostic way. To circumvent the need for real-world data in evaluation, we explore the use of synthetic binary classification tasks with Gaussian mixtures to probe pretrained models and compare the robustness-accuracy performance on pretrained representations with an idealized reference. Our approach offers a holistic evaluation, revealing intrinsic model capabilities and reducing the dependency on real-life data for model evaluation. Evaluated with various pretrained image models, the experimental results confirm that our task-agnostic evaluation correlates with actual linear probing performance on downstream tasks and can also guide parameter choice in robust linear probing to achieve a better robustness-accuracy trade-off.

In recent years, work has gone into developing deep interpretable methods for image classification that clearly attributes a model's output to specific features of the data. One such of these methods is the Prototypical Part Network (ProtoPNet), which attempts to classify images based on meaningful parts of the input. While this architecture is able to produce visually interpretable classifications, it often learns to classify based on parts of the image that are not semantically meaningful. To address this problem, we propose the Reward Reweighing, Reselecting, and Retraining (R3) post-processing framework, which performs three additional corrective updates to a pretrained ProtoPNet in an offline and efficient manner. The first two steps involve learning a reward model based on collected human feedback and then aligning the prototypes with human preferences. The final step is retraining, which realigns the base features and the classifier layer of the original model with the updated prototypes. We find that our R3 framework consistently improves both the interpretability and the predictive accuracy of ProtoPNet and its variants.