ICML 2024 Papers

Parameter-efficient finetuning (PEFT) has become ubiquitous to adapt foundation models to downstream task requirements while retaining their generalization ability. However, the amount of additionally introduced parameters and compute for successful adaptation and hyperparameter searches can explode quickly, especially when deployed at scale to serve numerous individual requests. To ensure effective, parameter-efficient, and hyperparameter-robust adaptation, we propose the *ETHER* transformation family, which performs Efficient fineTuning via HypErplane Reflections. By design, *ETHER* transformations require *a minimal number of parameters*, are *less likely to deteriorate model performance*, and exhibit *robustness to hyperparameter and learning rate choices*. In particular, we introduce *ETHER* and its relaxation *ETHER+*, which match or outperform existing PEFT methods with significantly fewer parameters ($\sim$$10$-$100$ times lower than LoRA or OFT) across multiple image synthesis and natural language tasks without *exhaustive hyperparameter tuning*. Finally, we investigate the recent emphasis on Hyperspherical Energy retention for adaptation and raise questions on its practical utility. The code is available at https://github.com/mwbini/ether.

In recent years, there have been significant improvements in various forms of image outlier detection. However, outlier detection performance under adversarial settings lags far behind that in standard settings. This is due to the lack of effective exposure to adversarial scenarios during training, especially on unseen outliers, leading detection models failing to learn robust features. To bridge this gap, we introduce RODEO, a data-centric approach that generates effective outliers for robust outlier detection. More specifically, we show that incorporating outlier exposure (OE) and adversarial training could be an effective strategy for this purpose, as long as the exposed training outliers meet certain characteristics, including diversity, and both conceptual differentiability and analogy to the inlier samples. We leverage a text-to-image model to achieve this goal. We demonstrate both quantitatively and qualitatively that our adaptive OE method effectively generates ''diverse'' and ''near-distribution'' outliers, leveraging information from both text and image domains. Moreover, our experimental results show that utilizing our synthesized outliers significantly enhances the performance of the outlier detector, particularly in adversarial settings.

Nowadays Spiking Transformers have exhibited remarkable performance close to Artificial Neural Networks (ANNs), while enjoying the inherent energy-efficiency of Spiking Neural Networks (SNNs). However, training Spiking Transformers on GPUs is considerably more time-consuming compared to the ANN counterparts, despite the energy-efficient inference through neuromorphic computation. In this paper, we investigate the token sparsification technique for efficient training of Spiking Transformer and find conventional methods suffer from noticeable performance degradation. We analyze the issue and propose our Sparsification with Timestep-wise Anchor Token and dual Alignments (STATA). Timestep-wise Anchor Token enables precise identification of important tokens across timesteps based on standardized criteria. Additionally, dual Alignments incorporate both Intra and Inter Alignment of the attention maps, fostering the learning of inferior attention. Extensive experiments show the effectiveness of STATA thoroughly, which demonstrates up to $\sim$1.53$\times$ training speedup and $\sim$48% energy reduction with comparable performance on various datasets and architectures.

Performative prediction, as introduced by Perdomo et al., is a framework for studying social prediction in which the data distribution itself changes in response to the deployment of a model. Existing work in this field usually hinges on three assumptions that are easily violated in practice: that the performative risk is convex over the deployed model, that the mapping from the model to the data distribution is known to the model designer in advance, and the first-order information of the performative risk is available. In this paper, we initiate the study of performative prediction problems that do not require these assumptions. Specifically, we develop a parameterization framework that parametrizes the performative prediction objective as a function of the induced data distribution. We also develop a two-level zeroth-order optimization procedure, where the first level performs iterative optimization on the distribution parameter space, and the second level learns the model that induced a particular target distribution parameter at each iteration. Under mild conditions, this reparameterization allows us to transform the non-convex objective into a convex one and achieve provable regret guarantees. In particular, we provide a regret bound that is sublinear in the total number of performative samples taken and is only polynomial in the dimension of the model parameter.

Local steps are crucial for Federated Learning (FL) algorithms and have witnessed great empirical success in reducing communication costs and improving the generalization performance of deep neural networks. However, there are limited studies on the effect of local steps on heterogeneous FL. A few works investigate this problem from the optimization perspective. Woodworth et al. (2020a) showed that the iteration complexity of Local SGD, the most popular FL algorithm, is dominated by the baseline mini-batch SGD, which does not show the benefits of local steps. In addition, Levy (2023) proposed a new local update method that provably benefits over mini-batch SGD. However, in the same setting, there is still no work analyzing the effects of local steps to generalization in a heterogeneous FL setting. Motivated by our experimental findings where Local SGD learns more distinguishing features than parallel SGD, this paper studies the generalization benefits of local steps from a feature learning perspective. We propose a novel federated data model that exhibits a new form of data heterogeneity, under which we show that a convolutional neural network (CNN) trained by GD with *global* updates will miss some pattern-related features, while the network trained by GD with *local* updates can learn all features in polynomial time. Consequently, local steps help CNN generalize better in our data model. In a different parameter setting, we also prove that Local GD with *one-shot* model averaging can learn all features and generalize well in all clients. Our experimental results also confirm the benefits of local steps in improving test accuracy on real-world data.

Out-of-distribution (OOD) generalisation is challenging because it involves not only learning from empirical data, but also deciding among various notions of generalisation, e.g. optimise based on the average-case risk, worst-case risk, or interpolations thereof. While this decision should in principle be decided by the model operator like medical doctors in practice, this information might not always be available at training time. This situation leads to arbitrary commitments to specific generalisation strategies by machine learners due to these deployment uncertainties. We introduce the Imprecise Domain Generalisation framework to mitigate this, featuring an imprecise risk optimisation that allows learners to stay imprecise by optimising against a continuous spectrum of generalisation strategies during training, and a model framework that allows operators to specify their generalisation preference at deployment. Our work, supported by theoretical and empirical evidence, showcases the benefits of integrating imprecision into domain generalisation.

Group Equivariant CNNs (G-CNNs) have shown promising efficacy in various tasks, owing to their ability to capture hierarchical features in an equivariant manner. However, their equivariance is fixed to the symmetry of the whole group, limiting adaptability to diverse partial symmetries in real-world datasets, such as limited rotation symmetry of handwritten digit images and limited color-shift symmetry of flower images. Recent efforts address this limitation, one example being Partial G-CNN which restricts the output group space of convolution layers to break full equivariance. However, such an approach still fails to adjust equivariance levels across data. In this paper, we propose a novel approach, Variational Partial G-CNN (VP G-CNN), to capture varying levels of partial equivariance specific to each data instance. VP G-CNN redesigns the distribution of the output group elements to be conditioned on input data, leveraging variational inference to avoid overfitting. This enables the model to adjust its equivariance levels according to the needs of individual data points. Additionally, we address training instability inherent in discrete group equivariance models by redesigning the reparametrizable distribution. We demonstrate the effectiveness of VP G-CNN on both toy and real-world datasets, including MNIST67-180, CIFAR10, ColorMNIST, and Flowers102. Our results show robust performance, even in uncertainty metrics.

In previous literature, backward error analysis was used to find ordinary differential equations (ODEs) approximating the gradient descent trajectory. It was found that finite step sizes implicitly regularize solutions because terms appearing in the ODEs penalize the two-norm of the loss gradients. We prove that the existence of similar implicit regularization in RMSProp and Adam depends on their hyperparameters and the training stage, but with a different "norm" involved: the corresponding ODE terms either penalize the (perturbed) one-norm of the loss gradients or, conversely, impede its reduction (the latter case being typical). We also conduct numerical experiments and discuss how the proven facts can influence generalization.

We study deep neural networks for the multi-label classification (MLab) task through the lens of neural collapse (NC). Previous works have been restricted to the multi-class classification setting and discovered a prevalent NC phenomenon comprising of the following properties for the last-layer features: (i) the variability of features within every class collapses to zero, (ii) the set of feature means form an equi-angular tight frame (ETF), and (iii) the last layer classifiers collapse to the feature mean upon some scaling. We generalize the study to multi-label learning, and prove for the first time that a generalized NC phenomenon holds with the "pick-all-label'' formulation, which we term as MLab NC. While the ETF geometry remains consistent for features with a single label, multi-label scenarios introduce a unique combinatorial aspect we term the "tag-wise average" property, where the means of features with multiple labels are the scaled averages of means for single-label instances. Theoretically, under proper assumptions on the features, we establish that the only global optimizer of the pick-all-label cross-entropy loss satisfy the multi-label NC. In practice, we demonstrate that our findings can lead to better test performance with more efficient training techniques for MLab learning.

We study robust Markov games (RMG) with $s$-rectangular uncertainty. We show a general equivalence between computing a robust Nash equilibrium (RNE) of a $s$-rectangular RMG and computing a Nash equilibrium (NE) of an appropriately constructed regularized MG. The equivalence result yields a planning algorithm for solving $s$-rectangular RMGs, as well as provable robustness guarantees for policies computed using regularized methods. However, we show that even for just reward-uncertain two-player zero-sum matrix games, computing an RNE is PPAD-hard. Consequently, we derive a special uncertainty structure called efficient player-decomposability and show that RNE for two-player zero-sum RMG in this class can be provably solved in polynomial time. This class includes commonly used uncertainty sets such as $L_1$ and $L_\infty$ ball uncertainty sets.

AI-based frameworks for protein engineering use self-supervised learning (SSL) to obtain representations for downstream mutation effect predictions. The most common training objective for these methods is wildtype accuracy: given a sequence or structure where a wildtype residue has been masked, predict the missing amino acid. Wildtype accuracy, however, does not align with the primary goal of protein engineering, which is to suggest a mutation rather than to identify what already appears in nature. Here we present Evolutionary Ranking (EvoRank), a training objective that incorporates evolutionary information derived from multiple sequence alignments (MSAs) to learn more diverse protein representations. EvoRank corresponds to ranking amino-acid likelihoods in the probability distribution induced by an MSA. This objective forces models to learn the underlying evolutionary dynamics of a protein. Across a variety of phenotypes and datasets, we demonstrate that EvoRank leads to dramatic improvements in zero-shot performance and can compete with models fine-tuned on experimental data. This is particularly important in protein engineering, where it is expensive to obtain data for fine-tuning.

Diffusion models have shown remarkable performance on many generative tasks. Despite recent success, most diffusion models are restricted in that they only allow linear transformation of the data distribution. In contrast, broader family of transformations can help train generative distributions more efficiently, simplifying the reverse process and closing the gap between the true negative log-likelihood and the variational approximation. In this paper, we present Neural Diffusion Models (NDMs), a generalization of conventional diffusion models that enables defining and learning time-dependent non-linear transformations of data. We show how to optimise NDMs using a variational bound in a simulation-free setting. Moreover, we derive a time-continuous formulation of NDMs, which allows fast and reliable inference using off-the-shelf numerical ODE and SDE solvers. Finally, we demonstrate the utility of NDMs through experiments on many image generation benchmarks, including MNIST, CIFAR-10, downsampled versions of ImageNet and CelebA-HQ. NDMs outperform conventional diffusion models in terms of likelihood, achieving state-of-the-art results on ImageNet and CelebA-HQ, and produces high-quality samples.

Large-scale recommendation systems are characterized by their reliance on high cardinality, heterogeneous features and the need to handle tens of billions of user actions on a daily basis. Despite being trained on huge volume of data with thousands of features, most Deep Learning Recommendation Models (DLRMs) in industry fail to scale with compute. Inspired by success achieved by Transformers in language and vision domains, we revisit fundamental design choices in recommendation systems. We reformulate recommendation problems as sequential transduction tasks within a generative modeling framework (``Generative Recommenders''), and propose a new architecture, HSTU, designed for high cardinality, non-stationary streaming recommendation data. HSTU outperforms baselines over synthetic and public datasets by up to 65.8% in NDCG, and is 5.3x to 15.2x faster than FlashAttention2-based Transformers on 8192 length sequences. HSTU-based Generative Recommenders, with 1.5 trillion parameters, improve metrics in online A/B tests by 12.4% and have been deployed on multiple surfaces of a large internet platform with billions of users. More importantly, the model quality of Generative Recommenders empirically scales as a power-law of training compute across three orders of magnitude, up to GPT-3/LLaMa-2 scale, which reduces carbon footprint needed for future model developments, and further paves the way for the first foundation models in recommendations.

The intrinsic dimension (ID) represents the minimum dimension needed to describe data on a lower-dimensional manifold within high-dimensional spaces. Network pruning aims to reduce the complexity of high-dimensional networks while minimizing performance trade-offs. This symmetry motivates the exploration of ID as a metric for effective pruning. For vision-language models, we investigate whether different modalities exist on separate manifolds, indicating varying complexity and prunability. We empirically study ID variations in large-scale vision-language pre-trained models and examine the contributions of different modalities to model prunability. We propose a layer importance metric based on ID, which can conveniently integrate with current metrics and enhance performance in vision-language model pruning. The experimental results show a high correlation between ID and modality prunability. Visual representations are more sensitive and crucial to model performance, while language representations are more robust and offer greater prunability. Our findings suggest an asymmetric pruning strategy for vision and language modalities, guided by the ID metric. The code is available at https://github.com/Nofear18/ID_VL_Pruning

Graph embeddings have emerged as a powerful tool for understanding the structure of graphs. Unlike classical spectral methods, recent methods such as DeepWalk, Node2Vec, etc. are based on solving nonlinear optimization problems on the graph, using local information obtained by performing random walks. These techniques have empirically been shown to produce ``better'' embeddings than their classical counterparts. However, due to their reliance on solving a nonconvex optimization problem, obtaining theoretical guarantees on the properties of the solution has remained a challenge, even for simple classes of graphs. In this work, we show convergence properties for the DeepWalk algorithm on graphs obtained from the Stochastic Block Model (SBM). Despite being simplistic, the SBM has proved to be a classic model for analyzing the behavior of algorithms on large graphs. Our results mirror the existing ones for spectral embeddings on SBMs, showing that even in the case of one-dimensional embeddings, the output of the DeepWalk algorithm provably recovers the cluster structure with high probability.

Graph representation learning aims to represent graphs as vectors that can be utilized in downstream tasks such as graph classification. In this work, we focus on learning diverse representations that can capture the graph information as much as possible. We propose quantifying graph information using graph entropy, where we define a probability distribution of a graph based on its nodes' representations and global-graph representation. However, the computation of graph entropy is NP-hard due to the complex vertex-packing polytope involved in its definition. To address this challenge, we provide an approximation method leveraging orthonormal representations for graph entropy maximization. The proposed method is implemented via graph neural networks, resulting in informative node-level and graph-level representations. Experimental results demonstrate the effectiveness of our method in comparison to many baselines in unsupervised learning and semi-supervised learning tasks. The code of our method is available at https://github.com/MathAdventurer/GeMax.

Fragment-based drug discovery is an effective strategy for discovering drug candidates in the vast chemical space, and has been widely employed in molecular generative models. However, many existing fragment extraction methods in such models do not take the target chemical properties into account or rely on heuristic rules. Additionally, the existing fragment-based generative models cannot update the fragment vocabulary with goal-aware fragments newly discovered during the generation. To this end, we propose a molecular generative framework for drug discovery, named *Goal-aware fragment Extraction, Assembly, and Modification* (GEAM). GEAM consists of three modules, each responsible for goal-aware fragment extraction, fragment assembly, and fragment modification. The fragment extraction module identifies important fragments contributing to the desired target properties with the information bottleneck principle, thereby constructing an effective goal-aware fragment vocabulary. Moreover, GEAM can explore beyond the initial vocabulary with the fragment modification module, and the exploration is further enhanced through the dynamic goal-aware vocabulary update. We experimentally demonstrate that GEAM effectively discovers drug candidates through the generative cycle of the three modules in various drug discovery tasks. Our code is available at https://github.com/SeulLee05/GEAM.

Neural Processes (NPs) are popular meta-learning methods for efficiently modelling predictive uncertainty. Recent state-of-the-art methods, however, leverage expensive attention mechanisms, limiting their applications, particularly in low-resource settings. In this work, we propose Constant Memory Attentive Neural Processes (CMANPs), an NP variant that only requires **constant** memory. To do so, we first propose an efficient update operation for Cross Attention. Leveraging the update operation, we propose Constant Memory Attention Block (CMAB), a novel attention block that (i) is permutation invariant, (ii) computes its output in constant memory, and (iii) performs constant computation updates. Finally, building on CMAB, we detail Constant Memory Attentive Neural Processes. Empirically, we show CMANPs achieve state-of-the-art results on popular NP benchmarks while being significantly more memory efficient than prior methods.

In the rapidly changing healthcare landscape, the implementation of offline reinforcement learning (RL) in dynamic treatment regimes (DTRs) presents a mix of unprecedented opportunities and challenges. This position paper offers a critical examination of the current status of offline RL in the context of DTRs. We argue for a reassessment of applying RL in DTRs, citing concerns such as inconsistent and potentially inconclusive evaluation metrics, the absence of naive and supervised learning baselines, and the diverse choice of RL formulation in existing research. Through a case study with more than 17,000 evaluation experiments using a publicly available Sepsis dataset, we demonstrate that the performance of RL algorithms can significantly vary with changes in evaluation metrics and Markov Decision Process (MDP) formulations. Surprisingly, it is observed that in some instances, RL algorithms can be surpassed by random baselines subjected to policy evaluation methods and reward design. This calls for more careful policy evaluation and algorithm development in future DTR works. Additionally, we discussed potential enhancements toward more reliable development of RL-based dynamic treatment regimes and invited further discussion within the community. Code is available at https://github.com/GilesLuo/ReassessDTR.