NeurIPS 2024 Papers

Can we modify the training data distribution to encourage the underlying optimization method toward finding solutions with superior generalization performance on in-distribution data? In this work, we approach this question for the first time by comparing the inductive bias of gradient descent (GD) with that of sharpness-aware minimization (SAM). By studying a two-layer CNN, we rigorously prove that SAM learns different features more uniformly, particularly in early epochs. That is, SAM is less susceptible to simplicity bias compared to GD. We also show that examples constraining features that are learned early are separable from the rest based on the model’s output. Based on this observation, we propose a method that (i) clusters examples based on the network output early in training, (ii) identifies a cluster of examples with similar network output, and (iii) upsamples the rest of examples only once to alleviate the simplicity bias. We show empirically that USEFUL effectively improves the generalization performance on the original data distribution when training with various gradient methods, including (S)GD and SAM. Notably, we demonstrate that our method can be combined with SAM variants and existing data augmentation strategies to achieve, to the best of our knowledge, state-of-the-art performance for training ResNet18 on CIFAR10, STL10, CINIC10, Tiny-ImageNet; ResNet34 on CIFAR100; and VGG19 and DenseNet121 on CIFAR10.

Atmospheric turbulence, caused by random fluctuations in the atmosphere's refractive index, introduces complex spatio-temporal distortions in imagery captured at long range. Video Atmospheric Turbulence Mitigation (ATM) aims to restore videos affected by these distortions. However, existing video ATM methods, both supervised and self-supervised, struggle to maintain temporally consistent mitigation across frames, leading to visually incoherent results. This limitation arises from the stochastic nature of atmospheric turbulence, which varies across space and time. Inspired by the observation that atmospheric turbulence induces high-frequency temporal variations, we propose ConVRT, a novel framework for consistent video restoration through turbulence. ConVRT introduces a neural video representation that explicitly decouples spatial and temporal information into a spatial content field and a temporal deformation field, enabling targeted regularization of the network's temporal representation capability. By leveraging the low-pass filtering properties of the regularized temporal representations, ConVRT effectively mitigates turbulence-induced temporal frequency variations and promotes temporal consistency. Furthermore, our training framework seamlessly integrates supervised pre-training on synthetic turbulence data with self-supervised learning on real-world videos, significantly improving the temporally consistent mitigation of ATM methods on diverse real-world data. More information can be found on our project page: https://convrt-2024.github.io/

Adam has become one of the most favored optimizers in deep learning problems. Despite its success in practice, numerous mysteries persist regarding its theoretical understanding. In this paper, we study the implicit bias of Adam in linear logistic regression. Specifically, we show that when the training data are linearly separable, the iterates of Adam converge towards a linear classifier that achieves the maximum $\ell_\infty$-margin in direction. Notably, for a general class of diminishing learning rates, this convergence occurs within polynomial time. Our result shed light on the difference between Adam and (stochastic) gradient descent from a theoretical perspective.

We examine the continuous-time counterpart of mirror descent, namely mirror flow, on classification problems which are linearly separable. Such problems are minimised ‘at infinity’ and have many possible solutions; we study which solution is preferred by the algorithm depending on the mirror potential. For exponential tailed losses and under mild assumptions on the potential, we show that the iterates converge in direction towards a $\phi_\infty$-maximum margin classifier. The function $\phi_\infty$ is the horizon function of the mirror potential and characterises its shape ‘at infinity’. When the potential is separable, a simple formula allows to compute this function. We analyse several examples of potentials and provide numerical experiments highlighting our results.

Understanding the training dynamics of transformers is important to explain the impressive capabilities behind large language models. In this work, we study the dynamics of training a shallow transformer on a task of recognizing co-occurrence of two designated words. In the literature of studying training dynamics of transformers, several simplifications are commonly adopted such as weight reparameterization, attention linearization, special initialization, and lazy regime. In contrast, we analyze the gradient flow dynamics of simultaneously training three attention matrices and a linear MLP layer from random initialization, and provide a framework of analyzing such dynamics via a coupled dynamical system. We establish near minimum loss and characterize the attention model after training. We discover that gradient flow serves as an inherent mechanism that naturally divide the training process into two phases. In Phase 1, the linear MLP quickly aligns with the two target signals for correct classification, whereas the softmax attention remains almost unchanged. In Phase 2, the attention matrices and the MLP evolve jointly to enlarge the classification margin and reduce the loss to a near minimum value. Technically, we prove a novel property of the gradient flow, termed \textit{automatic balancing of gradients}, which enables the loss values of different samples to decrease almost at the same rate and further facilitates the proof of near minimum training loss. We also conduct experiments to verify our theoretical results.

Large Language Models (LLMs) have demonstrated remarkable capabilities, surpassing human experts in various benchmark tests and playing a vital role in various industry sectors. Despite their effectiveness, a notable drawback of LLMs is their inconsistent moral behavior, which raises ethical concerns. This work delves into symmetric moral consistency in large language models and demonstrates that modern LLMs lack sufficient consistency ability in moral scenarios. Our extensive investigation of twelve popular LLMs reveals that their assessed consistency scores are influenced by position bias and selection bias rather than their intrinsic abilities. We propose a new framework tSMC, which gauges the effects of these biases and effectively mitigates the bias impact based on the Kullback–Leibler divergence to pinpoint LLMs' mitigated Symmetric Moral Consistency. We find that the ability of LLMs to maintain consistency varies across different moral scenarios. Specifically, LLMs show more consistency in scenarios with clear moral answers compared to those where no choice is morally perfect. The average consistency score of 12 LLMs ranges from $60.7\%$ in high-ambiguity moral scenarios to $84.8\%$ in low-ambiguity moral scenarios.

Large language models (LLMs) have been observed to exhibit bias towards certain cultures due to the predominance of training data obtained from English corpora. Considering that multilingual cultural data is often expensive to procure, existing methodologies address this challenge through prompt engineering or culture-specific pre-training. However, these strategies may neglect the knowledge deficiency of low-resource cultures and necessitate substantial computing resources. In this paper, we propose CultureLLM, a cost-effective solution to integrate cultural differences into LLMs. CultureLLM employs the World Value Survey (WVS) as seed data and generates semantically equivalent training data through the proposed semantic data augmentation. Utilizing only $50$ seed samples from WVS with augmented data, we fine-tune culture-specific LLMs as well as a unified model (CultureLLM-One) for $9$ cultures, encompassing both rich and low-resource languages. Extensive experiments conducted on $60$ culture-related datasets reveal that CultureLLM significantly surpasses various counterparts such as GPT-3.5 (by $8.1$\%) and Gemini Pro (by $9.5$\%), demonstrating performance comparable to or exceeding that of GPT-4. Our human study indicates that the generated samples maintain semantic equivalence to the original samples, offering an effective solution for LLMs augmentation. Code is released at https://github.com/Scarelette/CultureLLM.

In contextual dynamic pricing, a seller sequentially prices goods based on contextual information. Buyers will purchase products only if the prices are below their valuations. The goal of the seller is to design a pricing strategy that collects as much revenue as possible. We focus on two different valuation models. The first assumes that valuations linearly depend on the context and are further distorted by noise. Under minor regularity assumptions, our algorithm achieves an optimal regret bound of $\tilde{\mathcal{O}}(T^{2/3})$, improving the existing results. The second model removes the linearity assumption, requiring only that the expected buyer valuation is $\beta$-H\"older in the context. For this model, our algorithm obtains a regret $\tilde{\mathcal{O}}(T^{d+2\beta/d+3\beta})$, where $d$ is the dimension of the context space.

Two competing paradigms exist for self-supervised learning of data representations. Joint Embedding Predictive Architectures (JEPAs) is a class of architectures in which semantically similar inputs are encoded into representations that are predictive of each other. A recent successful approach that falls under the JEPA framework is self-distillation, where an online encoder is trained to predict the output of the target encoder, sometimes with a lightweight predictor network. This is contrasted with the Masked Auto Encoder (MAE) paradigm, where an encoder and decoder are trained to reconstruct missing parts of the input in ambient space rather than its latent representation. A common motivation for using the JEPA approach over MAE is that the JEPA objective prioritizes abstract features over fine-grained pixel information (which can be unpredictable and uninformative). In this work, we seek to understand the mechanism behind this empirical observation by analyzing deep linear models. We uncover a surprising mechanism: in a simplified linear setting where both approaches learn similar representations, JEPAs are biased to learn high influence features, or features characterized by having high regression coefficients. Our results point to a distinct implicit bias of predicting in latent space that may shed light on its success in practice.

In recent years, a large number of algorithms for label integration and noise correction have been proposed to infer the unknown true labels of instances in crowdsourcing. They have made great advances in improving the label quality of crowdsourced datasets. However, due to the presence of intractable instances, these algorithms are usually not as significant in improving the model quality as they are in improving the label quality. To improve the model quality, this paper proposes an instance weighting-based bias-variance trade-off (IWBVT) approach. IWBVT at first proposes a novel instance weighting method based on the complementary set and entropy, which mitigates the impact of intractable instances and thus makes the bias and variance of trained models closer to the unknown true results. Then, IWBVT performs probabilistic loss regressions based on the bias-variance decomposition, which achieves the bias-variance trade-off and thus reduces the generalization error of trained models. Experimental results indicate that IWBVT can serve as a universal post-processing approach to significantly improving the model quality of existing state-of-the-art label integration algorithms and noise correction algorithms.

With the arrival of the Noisy Intermediate-Scale Quantum (NISQ) era, Variational Quantum Algorithms (VQAs) have emerged to obtain possible quantum advantage. In particular, how to effectively incorporate hard constraints in VQAs remains a critical and open question. In this paper, we manage to combine the Hamming Weight Preserving ansatz with a topological-aware parity check on physical qubits to enforce error mitigation and further hard constraints. We demonstrate the combination significantly outperforms peer VQA methods on both quantum chemistry problems and constrained combinatorial optimization problems e.g. Quadratic Assignment Problem. Intensive experimental results on both simulators and superconducting quantum processors are provided to verify that the combination of HWP ansatz with parity check is among the most promising candidates to demonstrate quantum advantages in the NISQ era to solve more realistic problems.

Disentangled representation learning strives to extract the intrinsic factors within the observed data. Factoring these representations in an unsupervised manner is notably challenging and usually requires tailored loss functions or specific structural designs. In this paper, we introduce a new perspective and framework, demonstrating that diffusion models with cross-attention itself can serve as a powerful inductive bias to facilitate the learning of disentangled representations. We propose to encode an image into a set of concept tokens and treat them as the condition of the latent diffusion model for image reconstruction, where cross attention over the concept tokens is used to bridge the encoder and the U-Net of the diffusion model. We analyze that the diffusion process inherently possesses the time-varying information bottlenecks. Such information bottlenecks and cross attention act as strong inductive biases for promoting disentanglement. Without any regularization term in the loss function, this framework achieves superior disentanglement performance on the benchmark datasets, surpassing all previous methods with intricate designs. We have conducted comprehensive ablation studies and visualization analyses, shedding a light on the functioning of this model. We anticipate that our findings will inspire more investigation on exploring diffusion model for disentangled representation learning towards more sophisticated data analysis and understanding.

We study stochastic approximation algorithms with Markovian noise and constant step-size $\alpha$. We develop a method based on infinitesimal generator comparisons to study the bias of the algorithm, which is the expected difference between $\theta_n$ ---the value at iteration $n$--- and $\theta^*$ ---the unique equilibrium of the corresponding ODE. We show that, under some smoothness conditions, this bias is of order $O(\alpha)$. Furthermore, we show that the time-averaged bias is equal to $\alpha V + O(\alpha^2)$, where $V$ is a constant characterized by a Lyapunov equation, showing that $E[\bar{\theta}_n] \approx \theta^*+V\alpha + O(\alpha^2)$, where $\bar{\theta}_n$ is the Polyak-Ruppert average. We also show that $\bar{\theta}_n$ converges with high probability around $\theta^*+\alpha V$. We illustrate how to combine this with Richardson-Romberg extrapolation to derive an iterative scheme with a bias of order $O(\alpha^2)$.

Machine learning systems often acquire biases by leveraging undesired features in the data, impacting accuracy variably across different sub-populations of the data. However, our current understanding of bias formation mostly focuses on the initial and final stages of learning, leaving a gap in knowledge regarding the transient dynamics. To address this gap, this paper explores the evolution of bias in a teacher-student setup that models different data sub-populations with a Gaussian-mixture model. We provide an analytical description of the stochastic gradient descent dynamics of a linear classifier in this setup, which we prove to be exact in high dimension. Notably, our analysis identifies different properties of the sub-populations that drive bias at different timescales and hence shows a shifting preference of our classifier during training. By applying our general solution to fairness and robustness, we delineate how and when heterogeneous data and spurious features can generate and amplify bias. We empirically validate our results in more complex scenarios by training deeper networks on synthetic and real data, i.e. using CIFAR10, MNIST, and CelebA datasets.

Implicit bias describes the phenomenon where optimization-based training algorithms, without explicit regularization, show a preference for simple estimators even when more complex estimators have equal objective values. Multiple works have developed the theory of implicit bias for binary classification under the assumption that the loss satisfies an *exponential tail property*. However, there is a noticeable gap in analysis for multiclass classification, with only a handful of results which themselves are restricted to the cross-entropy loss. In this work, we employ the framework of Permutation Equivariant and Relative Margin-based (PERM) losses [Wang and Scott, 2024] to introduce a multiclass extension of the exponential tail property. This class of losses includes not only cross-entropy but also other losses. Using this framework, we extend the implicit bias result of Soudry et al. [2018] to multiclass classification. Furthermore, our proof techniques closely mirror those of the binary case, thus illustrating the power of the PERM framework for bridging the binary-multiclass gap.

Recent works in Variational Inference have examined alternative criteria to the commonly used exclusive Kullback-Leibler divergence. Encouraging empirical results have been obtained with the family of alpha-divergences, but few works have focused on the asymptotic properties of the proposed algorithms, especially as the number of iterations goes to infinity. In this paper, we study a procedure that ensures a monotonic decrease in the alpha-divergence. We provide sufficient conditions to guarantee its convergence to a local minimizer of the alpha-divergence at a geometric rate when the variational family belongs to the class of exponential models. The sample-based version of this ideal procedure involves biased gradient estimators, thus hindering any theoretical study. We propose an alternative unbiased algorithm, we prove its almost sure convergence to a local minimizer of the alpha-divergence, and a law of the iterated logarithm. Our results are exemplified with toy and real-data experiments.

The remarkable capability of over-parameterised neural networks to generalise effectively has been explained by invoking a ``simplicity bias'': neural networks prevent overfitting by initially learning simple classifiers before progressing to more complex, non-linear functions. While simplicity biases have been described theoretically and experimentally in feed-forward networks for supervised learning, the extent to which they also explain the remarkable success of transformers trained with self-supervised techniques remains unclear. In our study, we demonstrate that transformers, trained on natural language data, also display a simplicity bias. Specifically, they sequentially learn many-body interactions among input tokens, reaching a saturation point in the prediction error for low-degree interactions while continuing to learn high-degree interactions. To conduct this analysis, we develop a procedure to generate \textit{clones} of a given natural language data set, which rigorously capture the interactions between tokens up to a specified order. This approach opens up the possibilities of studying how interactions of different orders in the data affect learning, in natural language processing and beyond.

Consider a hiring process with candidates coming from different universities. It is easy to order candidates with the same background, yet it can be challenging to compare them otherwise. The latter case requires additional costly assessments, leading to a potentially high total cost for the hiring organization. Given an assigned budget, what would be an optimal strategy to select the most qualified candidate? We model the above problem as a multicolor secretary problem, allowing comparisons between candidates from distinct groups at a fixed cost. Our study explores how the allocated budget enhances the success probability of online selection algorithms.

Large language models (LLMs) are increasingly essential in processing natural languages, yet their application is frequently compromised by biases and inaccuracies originating in their training data. In this study, we introduce \textbf{Cross-Care}, the first benchmark framework dedicated to assessing biases and real world knowledge in LLMs, specifically focusing on the representation of disease prevalence across diverse demographic groups. We systematically evaluate how demographic biases embedded in pre-training corpora like $ThePile$ influence the outputs of LLMs. We expose and quantify discrepancies by juxtaposing these biases against actual disease prevalences in various U.S. demographic groups. Our results highlight substantial misalignment between LLM representation of disease prevalence and real disease prevalence rates across demographic subgroups, indicating a pronounced risk of bias propagation and a lack of real-world grounding for medical applications of LLMs. Furthermore, we observe that various alignment methods minimally resolve inconsistencies in the models' representation of disease prevalence across different languages. For further exploration and analysis, we make all data and a data visualization tool available at: \url{www.crosscare.net}.