ICML 2024 Papers

Grokking, or delayed generalization, is a phenomenon where generalization in a deep neural network (DNN) occurs long after achieving near zero training error. Previous studies have reported the occurrence of grokking in specific controlled settings, such as DNNs initialized with large-norm parameters or transformers trained on algorithmic datasets. We demonstrate that grokking is actually much more widespread and materializes in a wide range of practical settings, such as training of a convolutional neural network (CNN) on CIFAR10 or a Resnet on Imagenette. We introduce the new concept of delayed robustness, whereby a DNN groks adversarial examples and becomes robust, long after interpolation and/or generalization. We develop an analytical explanation for the emergence of both delayed generalization and delayed robustness based on the local complexity of a DNN's input-output mapping. Our local complexity measures the density of so-called ``linear regions’’ (aka, spline partition regions) that tile the DNN input space and serves as a utile progress measure for training. We provide the first evidence that, for classification problems, the linear regions undergo a phase transition during training whereafter they migrate away from the training samples (making the DNN mapping smoother there) and towards the decision boundary (making the DNN mapping less smooth there). Grokking occurs post phase transition as a robust partition of the input space thanks to the linearization of the DNN mapping around the training points. Web: https://bit.ly/grok-adversarial.

When deployed in the world, a learning agent such as a recommender system or a chatbot often repeatedly interacts with another learning agent (such as a user) over time. In many such two-agent systems, each agent learns separately and the rewards of the two agents are not perfectly aligned. To better understand such cases, we examine the learning dynamics of the two-agent system and the implications for each agent's objective. We model these systems as Stackelberg games with decentralized learning and show that standard regret benchmarks (such as Stackelberg equilibrium payoffs) result in worst-case linear regret for at least one player. To better capture these systems, we construct a relaxed regret benchmark that is tolerant to small learning errors by agents. We show that standard learning algorithms fail to provide sublinear regret, and we develop algorithms to achieve near-optimal $\mathcal{O}(T^{2/3})$ regret for both players with respect to these benchmarks. We further design relaxed environments under which faster learning ($\mathcal{O}(\sqrt{T})$) is possible. Altogether, our results take a step towards assessing how two-agent interactions in sequential and decentralized learning environments affect the utility of both agents.

In supervised learning, understanding an input’s proximity to the training data can help a model decide whether it has sufficient evidence for reaching a reliable prediction. While powerful probabilistic models such as Gaussian Processes naturally have this property, deep neural networks often lack it. In this paper, we introduce Distance Aware Bottleneck (DAB), i.e., a new method for enriching deep neural networks with this property. Building on prior information bottleneck approaches, our method learns a codebook that stores a compressed representation of all inputs seen during training. The distance of a new example from this codebook can serve as an uncertainty estimate for the example. The resulting model is simple to train and provides deterministic uncertainty estimates by a single forward pass. Finally, our method achieves better out-of-distribution (OOD) detection and misclassification prediction than prior methods, including expensive ensemble methods, deep kernel Gaussian Processes, and approaches based on the standard information bottleneck.

We present GALA3D, generative 3D GAussians with LAyout-guided control, for effective compositional text-to-3D generation. We first utilize large language models (LLMs) to generate the initial layout and introduce a layout-guided 3D Gaussian representation for 3D content generation with adaptive geometric constraints. We then propose an instance-scene compositional optimization mechanism with conditioned diffusion to collaboratively generate realistic 3D scenes with consistent geometry, texture, scale, and accurate interactions among multiple objects while simultaneously adjusting the coarse layout priors extracted from the LLMs to align with the generated scene. Experiments show that GALA3D is a user-friendly, end-to-end framework for state-of-the-art scene-level 3D content generation and controllable editing while ensuring the high fidelity of object-level entities within the scene. The source codes and models will be available at gala3d.github.io.

In this paper, we discuss feature engineering for single-pass uncertainty estimation. For accurate uncertainty estimates, neural networks must extract differences in the feature space that quantify uncertainty. This could be achieved by current single-pass approaches that maintain feature distances between data points as they traverse the network. While initial results are promising, maintaining feature distances within the network representations frequently inhibits information compression and opposes the learning objective. We study this effect theoretically and empirically to arrive at a simple conclusion: preserving feature distances in the output is beneficial when the preserved features contribute to learning the label distribution and act in opposition otherwise. We then propose Transitional Uncertainty with Layered Intermediate Predictions (TULIP) as a simple approach to address the shortcomings of current single-pass estimators. Specifically, we implement feature preservation by extracting features from intermediate representations before information is collapsed by subsequent layers. We refer to the underlying preservation mechanism as transitional feature preservation. We show that TULIP matches or outperforms current single-pass methods on standard benchmarks and in practical settings where these methods are less reliable (imbalances, complex architectures, medical modalities).

A driving force behind the diverse applicability of modern machine learning is the ability to extract meaningful features across many sources. However, many practical domains involve data that are non-identically distributed across sources, and possibly statistically dependent within its source, violating vital assumptions in existing theoretical studies of representation learning. Toward addressing these issues, we establish statistical guarantees for learning general *nonlinear* representations from multiple data sources that admit different input distributions and possibly dependent data. Specifically, we study the sample-complexity of learning $T+1$ functions $f_\star^{(t)} \circ g_\star$ from a function class $\mathcal{F} \times \mathcal{G}$, where $f_\star^{(t)}$ are task specific linear functions and $g_\star$ is a shared non-linear representation. An approximate representation $\hat g$ is estimated using $N$ samples from each of $T$ source tasks, and a fine-tuning function $\hat f^{(0)}$ is fit using $N'$ samples from a target task passed through $\hat g$. Our results show that the excess risk of the estimate $\hat f^{(0)} \circ \hat g$ on the target task decays as $\tilde{\mathcal{O}}\Big(\frac{\mathrm{C}(\mathcal{G})}{N T} + \frac{\text{dim}(\mathcal{F})}{N'}\Big)$, where $\mathrm{C}(\mathcal{G})$ denotes the complexity of $\mathcal{G}$. Notably, our rates match that of the iid setting, while requiring fewer samples per task than prior analysis and admitting *no dependence on the mixing time*. We support our analysis with numerical experiments performing imitation learning over non-linear dynamical systems.

Distributed learning is an effective approach to accelerate model training by using parallel computing power of multiple workers. However, substantial communication delays arise between workers and a parameter server due to the massive costs associated with communicating gradients. SignSGD with majority voting (signSGD-MV) is a simple yet effective optimizer that reduces communication costs through sign quantization, but its convergence rate significantly decreases when adversarial workers arbitrarily manipulate datasets or local gradient updates. In this paper, we consider a distributed learning problem where the workforce comprises a mixture of honest and adversarial workers. In this setting, we show that the convergence rate can remain invariant as long as the number of honest workers providing trustworthy local updates to the parameter server exceeds the number of adversarial workers. The key idea behind this counter-intuitive result is our novel aggregation method, signSGD with federated defense (signSGD-FD). Unlike traditional approaches, signSGD-FD utilizes the gradient information sent by adversarial workers with appropriate weights, obtained through gradient sign decoding. Experimental results demonstrate that signSGD-FD achieves superior convergence rates compared to traditional algorithms in various adversarial attack scenarios.

In 1991, Brenier proved a theorem that generalizes the polar decomposition for square matrices -- factored as PSD $\times$ unitary -- to any vector field $F:\mathbb{R}^d\rightarrow \mathbb{R}^d$. The theorem, known as the polar factorization theorem, states that any field $F$ can be recovered as the composition of the gradient of a convex function $u$ with a measure-preserving map $M$, namely $F=\nabla u \circ M$. We propose a practical implementation of this far-reaching theoretical result, and explore possible uses within machine learning. The theorem is closely related to optimal transport (OT) theory, and we borrow from recent advances in the field of neural optimal transport to parameterize the potential $u$ as an input convex neural network. The map $M$ can be either evaluated pointwise using $u^*$, the convex conjugate of $u$, through the identity $M=\nabla u^* \circ F$, or learned as an auxiliary network. Because $M$ is, in general, not injective, we consider the additional task of estimating the ill-posed inverse map that can approximate the pre-image measure $M^{-1}$ using a stochastic generator. We illustrate possible applications of Brenier's polar factorization to non-convex optimization problems, as well as sampling of densities that are not log-concave.

In multi-agent reinforcement learning (MARL), effective exploration is critical, especially in sparse reward environments. Although introducing global intrinsic rewards can foster exploration in such settings, it often complicates credit assignment among agents. To address this difficulty, we propose Individual Contributions as intrinsic Exploration Scaffolds (ICES), a novel approach to motivate exploration by assessing each agent's contribution from a global view. In particular, ICES constructs exploration scaffolds with Bayesian surprise, leveraging global transition information during centralized training. These scaffolds, used only in training, help to guide individual agents towards actions that significantly impact the global latent state transitions. Additionally, ICES separates exploration policies from exploitation policies, enabling the former to utilize privileged global information during training. Extensive experiments on cooperative benchmark tasks with sparse rewards, including Google Research Football (GRF) and StarCraft Multi-agent Challenge (SMAC), demonstrate that ICES exhibits superior exploration capabilities compared with baselines. The code is publicly available at https://github.com/LXXXXR/ICES.

We study the problem of obtaining optimal and realistic prescriptions when using ReLU networks for data-driven decision-making. In this setting, the network is used to predict a quantity of interest and then is optimized to retrieve the decisions that maximize the quantity (e.g. find the best prices that maximize revenue). However, optimizing over-parameterized models often produces unrealistic prescriptions, far from the data manifold. This phenomenon is known as the Optimizer's Curse. To tackle this problem, we model the requirement for the resulting decisions to align with the data manifold as a tractable optimization constraint. This is achieved by reformulating the highly nonlinear Local Outlier Factor (LOF) metric as a single linear or quadratic constraint. To solve the problem efficiently for large networks, we propose an adaptive sampling algorithm that reduces the initial hard-to-solve optimization problem into a small number of significantly easier-to-solve problems by restricting the decision space to realistic polytopes, i.e. polytopes of the decision space that contain at least one realistic data point. Experiments on publicly available networks demonstrate the efficacy and scalability of our approach.

Dataset distillation has emerged as a strategy to overcome the hurdles associated with large datasets by learning a compact set of synthetic data that retains essential information from the original dataset. While distilled data can be used to train high performing models, little is understood about how the information is stored. In this study, we posit and answer three questions about the behavior, representativeness, and point-wise information content of distilled data. We reveal distilled data cannot serve as a substitute for real data during training outside the standard evaluation setting for dataset distillation. Additionally, the distillation process retains high task performance by compressing information related to the early training dynamics of real models. Finally, we provide an framework for interpreting distilled data and reveal that individual distilled data points contain meaningful semantic information. This investigation sheds light on the intricate nature of distilled data, providing a better understanding on how they can be effectively utilized.

Since its debut in 2016, ResNet has become arguably the most favorable architecture in deep neural network (DNN) design. It effectively addresses the gradient vanishing/exploding issue in DNN training, allowing engineers to fully unleash DNN's potential in tackling challenging problems in various domains. Despite its practical success, an essential theoretical question remains largely open: how well/best can ResNet approximate functions? In this paper, we answer this question for several important function classes, including polynomials and smooth functions. In particular, we show that ResNet with constant width can approximate Lipschitz continuous function with a Lipschitz constant $\mu$ using $\mathcal{O}(c(d)(\varepsilon/\mu)^{-d/2})$ tunable weights, where $c(d)$ is a constant depending on the input dimension $d$ and $\epsilon>0$ is the target approximation error. Further, we extend such a result to Lebesgue-integrable functions with the upper bound characterized by the modulus of continuity. These results indicate a factor of $d$ reduction in the number of tunable weights compared with the classical results for ReLU networks. Our results are also order-optimal in $\varepsilon$, thus achieving optimal approximation rate, as they match a generalized lower bound derived in this paper. This work adds to the theoretical justifications for ResNet's stellar practical performance.

We propose Diffusion Inference-Time T-Optimization (DITTO), a general-purpose framework for controlling pre-trained text-to-music diffusion models at inference-time via optimizing initial noise latents. Our method can be used to optimize through any differentiable feature matching loss to achieve a target (stylized) output and leverages gradient checkpointing for memory efficiency. We demonstrate a surprisingly wide-range of applications for music generation including inpainting, outpainting, and looping as well as intensity, melody, and musical structure control – all without ever fine-tuning the underlying model. When we compare our approach against related training, guidance, and optimization-based methods, we find DITTO achieves state-of-the-art performance on nearly all tasks, including outperforming comparable approaches on controllability, audio quality, and computational efficiency, thus opening the door for high-quality, flexible, training-free control of diffusion models. Sound examples can be found at https://ditto-music.github.io/web/.

We give the first non-trivial decremental dynamic embedding of a weighted, undirected graph $G$ into $\ell_p$ space. Given a weighted graph $G$ undergoing a sequence of edge weight increases, the goal of this problem is to maintain a (randomized) mapping $\phi: (G,d) \to (X,\ell_p)$ from the set of vertices of the graph to the $\ell_p$ space such that for every pair of vertices $u$ and $v$, the expected distance between $\phi(u)$ and $\phi(v)$ in the $\ell_p$ metric is within a small multiplicative factor, referred to as the distortion, of their distance in $G$. Our main result is a dynamic algorithm with expected distortion $O(\log^2 n)$ and total update time $O\left((m^{1+o(1)} \log^2 W + Q)\log(nW) \right)$, where $W$ is the maximum weight of the edges, $Q$ is the total number of updates and $n, m$ denote the number of vertices and edges in $G$ respectively. This is the first result of its kind, extending the seminal result of Bourgain '85 to the expanding field of dynamic algorithms. Moreover, we demonstrate that in the fully dynamic regime, where we tolerate edge insertions as well as deletions, no algorithm can explicitly maintain an embedding into $\ell_p$ space that has a low distortion with high probability.

Optimization objectives in the form of a sum of intractable expectations are rising in importance (*e.g.,*, diffusion models, variational autoencoders, and many more), a setting also known as "finite sum with infinite data." For these problems, a popular strategy is to employ SGD with *doubly stochastic gradients* (doubly SGD): the expectations are estimated using the gradient estimator of each component, while the sum is estimated by subsampling over these estimators. Despite its popularity, little is known about the convergence properties of doubly SGD, except under strong assumptions such as bounded variance. In this work, we establish the convergence of doubly SGD with independent minibatching and random reshuffling under general conditions, which encompasses dependent component gradient estimators. In particular, for dependent estimators, our analysis allows fined-grained analysis of the effect correlations. As a result, under a per-iteration computational budget of $b \times m$, where $b$ is the minibatch size and $m$ is the number of Monte Carlo samples, our analysis suggests where one should invest most of the budget in general. Furthermore, we prove that random reshuffling (RR) improves the complexity dependence on the subsampling noise.

The renowned difference-in-differences (DiD) estimator relies on the assumption of 'parallel trends,' which may not hold in many practical applications. To address this issue, economists are increasingly considering the triple difference estimator as a more credible alternative. Both DiD and triple difference are limited to assessing average effects exclusively. An alternative avenue is offered by the changes-in-changes (CiC) estimator, which provides an estimate of the entire counterfactual distribution by relying on assumptions imposed on the distribution of potential outcomes. In this work, we extend the triple difference estimator to accommodate the CiC framework, presenting the `triple changes estimator' and its identification assumptions, thereby expanding the scope of the CiC paradigm. Subsequently, we empirically evaluate the proposed framework and apply it to a study examining the impact of Medicaid expansion on children's preventive care.

In this work, we analyze both theoretically and empirically the effect of tied input-output embeddings—a popular technique that reduces the model size while often improving training. Interestingly, we found that this technique is connected to Harris (1954)’s distributional hypothesis—often portrayed by the famous Firth (1957)’s quote “a word is characterized by the company it keeps”. Specifically, our findings indicate that words (or, more broadly, symbols) with similar semantics tend to be encoded in similar input embeddings, while words that appear in similar contexts are encoded in similar output embeddings (thus explaining the semantic space arising in input and output embedding of foundational language models). As a consequence of these findings, the tying of the input and output embeddings is encouraged only when the distributional hypothesis holds for the underlying data. These results also provide insight into the embeddings of foundation language models (which are known to be semantically organized). Further, we complement the theoretical findings with several experiments supporting the claims.

Considering generating samples with high rewards, we focus on optimizing deep neural networks parameterized stochastic differential equations (SDEs), the advanced generative models with high expressiveness, with policy gradient, the leading algorithm in reinforcement learning. Nevertheless, when applying policy gradients to SDEs, since the policy gradient is estimated on a finite set of trajectories, it can be ill-defined, and the policy behavior in data-scarce regions may be uncontrolled. This challenge compromises the stability of policy gradients and negatively impacts sample complexity. To address these issues, we propose constraining the SDE to be consistent with its associated perturbation process. Since the perturbation process covers the entire space and is easy to sample, we can mitigate the aforementioned problems. Our framework offers a general approach allowing for a versatile selection of policy gradient methods to effectively and efficiently train SDEs. We evaluate our algorithm on the task of structure-based drug design and optimize the binding affinity of generated ligand molecules. Our method achieves the best Vina score (-9.07) on the CrossDocked2020 dataset.

Second-order information is valuable for many applications but challenging to compute. Several works focus on computing or approximating Hessian diagonals, but even this simplification introduces significant additional costs compared to computing a gradient. In the absence of efficient exact computation schemes for Hessian diagonals, we revisit an early approximation scheme proposed by Becker and LeCun (1989, BL89), which has a cost similar to gradients and appears to have been overlooked by the community. We introduce HesScale, an improvement over BL89, which adds negligible extra computation. On small networks, we find that this improvement is of higher quality than all alternatives, even those with theoretical guarantees, such as unbiasedness, while being much cheaper to compute. We use this insight in reinforcement learning problems where small networks are used and demonstrate HesScale in second-order optimization and scaling the step-size parameter. In our experiments, HesScale optimizes faster than existing methods and improves stability through step-size scaling. These findings are promising for scaling second-order methods in larger models in the future.