NeurIPS 2024 论文列表

We introduce DiffAug, a simple and efficient diffusion-based augmentation technique to train image classifiers for the crucial yet challenging goal of improved classifier robustness. Applying DiffAug to a given example consists of one forward-diffusion step followed by one reverse-diffusion step. Using both ResNet-50 and Vision Transformer architectures, we comprehensively evaluate classifiers trained with DiffAug and demonstrate the surprising effectiveness of single-step reverse diffusion in improving robustness to covariate shifts, certified adversarial accuracy and out of distribution detection. When we combine DiffAug with other augmentations such as AugMix and DeepAugment we demonstrate further improved robustness. Finally, building on this approach, we also improve classifier-guided diffusion wherein we observe improvements in: (i) classifier-generalization, (ii) gradient quality (i.e., improved perceptual alignment) and (iii) image generation performance. We thus introduce a computationally efficient technique for training with improved robustness that does not require any additional data, and effectively complements existing augmentation approaches.

We study a dynamic pricing problem for third-party platform service fees under strategic, far-sighted customers. In each time period, the platform sets a service fee based on historical data, observes the resulting transaction quantities, and collects revenue. The platform also monitors equilibrium prices influenced by both demand and supply. The objective is to maximize total revenue over a time horizon $T$. Our problem incorporates three practical challenges: (a) initially, the platform lacks knowledge of the demand side beforehand, necessitating a balance between exploring (learning the demand curve) and exploiting (maximizing revenue) simultaneously; (b) since only equilibrium prices and quantities are observable, traditional Ordinary Least Squares (OLS) estimators would be biased and inconsistent; (c) buyers are rational and strategic, seeking to maximize their consumer surplus and potentially misrepresenting their preferences. To address these challenges, we propose novel algorithmic solutions. Our approach involves: (i) a carefully designed active randomness injection to balance exploration and exploitation effectively; (ii) using non-i.i.d. actions as instrumental variables (IV) to consistently estimate demand; (iii) a low-switching cost design that promotes nearly truthful buyer behavior. We show an expected regret bound of $\tilde{\mathcal{O}} (\sqrt{T}\wedge\sigma_S^{-2})$ and demonstrate its optimality, up to logarithmic factors, with respect to both the time horizon $T$ and the randomness in supply $\sigma_S$. Despite its simplicity, our model offers valuable insights into the use of actions as estimation instruments, the benefits of low-switching pricing policies in mitigating strategic buyer behavior, and the role of supply randomness in facilitating exploration which leads to a phase transition of policy performance.

In class-incremental learning (class-IL), models must classify all previously seen classes at test time without task-IDs, leading to task confusion. Despite being a key challenge, task confusion lacks a theoretical understanding. We present a novel mathematical framework for class-IL and prove the Infeasibility Theorem, showing optimal class-IL is impossible with discriminative modeling due to task confusion. However, we establish the Feasibility Theorem, demonstrating that generative modeling can achieve optimal class-IL by overcoming task confusion. We then assess popular class-IL strategies, including regularization, bias-correction, replay, and generative classifier, using our framework. Our analysis suggests that adopting generative modeling, either for generative replay or direct classification (generative classifier), is essential for optimal class-IL.

Temporal link prediction, aiming at predicting future interactions among entities based on historical interactions, is crucial for a series of real-world applications. Although previous methods have demonstrated the importance of relative encodings for effective temporal link prediction, computational efficiency remains a major concern in constructing these encodings. Moreover, existing relative encodings are usually constructed based on structural connectivity, where temporal information is seldom considered. To address the aforementioned issues, we first analyze existing relative encodings and unify them as a function of temporal walk matrices. This unification establishes a connection between relative encodings and temporal walk matrices, providing a more principled way for analyzing and designing relative encodings. Based on this analysis, we propose a new temporal graph neural network called TPNet, which introduces a temporal walk matrix that incorporates the time decay effect to simultaneously consider both temporal and structural information. Moreover, TPNet designs a random feature propagation mechanism with theoretical guarantees to implicitly maintain the temporal walk matrices, which improves the computation and storage efficiency. Experimental results on 13 benchmark datasets verify the effectiveness and efficiency of TPNet, where TPNet outperforms other baselines on most datasets and achieves a maximum speedup of $33.3 \times$ compared to the SOTA baseline.

We study the \emph{in-context learning} (ICL) ability of a \emph{Linear Transformer Block} (LTB) that combines a linear attention component and a linear multi-layer perceptron (MLP) component. For ICL of linear regression with a Gaussian prior and a \emph{non-zero mean}, we show that LTB can achieve nearly Bayes optimal ICL risk. In contrast, using only linear attention must incur an irreducible additive approximation error. Furthermore, we establish a correspondence between LTB and one-step gradient descent estimators with learnable initialization ($\mathsf{GD}-\beta$), in the sense that every $\mathsf{GD}-\beta$ estimator can be implemented by an LTB estimator and every optimal LTB estimator that minimizes the in-class ICL risk is effectively a $\mathsf{GD}-\beta$ estimator. Finally, we show that $\mathsf{GD}-\beta$ estimators can be efficiently optimized with gradient flow, despite a non-convex training objective. Our results reveal that LTB achieves ICL by implementing $\mathsf{GD}-\beta$, and they highlight the role of MLP layers in reducing approximation error.

Non-autoregressive Transformers (NATs) are recently applied in direct speech-to-speech translation systems, which convert speech across different languages without intermediate text data. Although NATs generate high-quality outputs and offer faster inference than autoregressive models, they tend to produce incoherent and repetitive results due to complex data distribution (e.g., acoustic and linguistic variations in speech). In this work, we introduce DiffNorm, a diffusion-based normalization strategy that simplifies data distributions for training NAT models. After training with a self-supervised noise estimation objective, DiffNorm constructs normalized target data by denoising synthetically corrupted speech features. Additionally, we propose to regularize NATs with classifier-free guidance, improving model robustness and translation quality by randomly dropping out source information during training. Our strategies result in a notable improvement of about $+7$ ASR-BLEU for English-Spanish (En-Es) translation and $+2$ ASR-BLEU for English-French (En-Fr) on the CVSS benchmark, while attaining over $14\times$ speedup for En-Es and $5 \times$ speedup for En-Fr translations compared to autoregressive baselines.

To protect deep neural networks (DNNs) from adversarial attacks, adversarial training (AT) is developed by incorporating adversarial examples (AEs) into model training. Recent studies show that adversarial attacks disproportionately impact the patterns within the phase of the sample's frequency spectrum---typically containing crucial semantic information---more than those in the amplitude, resulting in the model's erroneous categorization of AEs. We find that, by mixing the amplitude of training samples' frequency spectrum with those of distractor images for AT, the model can be guided to focus on phase patterns unaffected by adversarial perturbations. As a result, the model's robustness can be improved. Unfortunately, it is still challenging to select appropriate distractor images, which should mix the amplitude without affecting the phase patterns. To this end, in this paper, we propose an optimized **Adversarial Amplitude Generator (AAG)** to achieve a better tradeoff between improving the model's robustness and retaining phase patterns. Based on this generator, together with an efficient AE production procedure, we design a new **Dual Adversarial Training (DAT)** strategy. Experiments on various datasets show that our proposed DAT leads to significantly improved robustness against diverse adversarial attacks. The source code is available at https://github.com/Feng-peng-Li/DAT.

We formalize the problem of prompt compression for large language models (LLMs) and present a framework to unify token-level prompt compression methods which create hard prompts for black-box models. We derive the distortion-rate function for this setup as a linear program, and provide an efficient algorithm to compute this fundamental limit via the dual of the linear program. Using the distortion-rate function as the baseline, we study the performance of existing compression schemes on a synthetic dataset consisting of prompts generated from a Markov chain, natural language queries, and their respective answers. Our empirical analysis demonstrates the criticality of query-aware prompt compression, where the compressor has knowledge of the downstream task/query for the black-box LLM. We show that there is a large gap between the performance of current prompt compression methods and the optimal strategy, and propose Adaptive QuerySelect, a query-aware, variable-rate adaptation of a prior work to close the gap. We extend our experiments to a small natural language dataset to further confirm our findings on our synthetic dataset.

Evidential Deep Learning (EDL), grounded in Evidence Theory and Subjective Logic (SL), provides a robust framework to estimate uncertainty for out-of-distribution (OOD) detection alongside traditional classification probabilities.However, the EDL framework is constrained by its focus on evidence that supports only single categories, neglecting the other collective evidences that could corroborate multiple in-distribution categories. This limitation leads to a diminished estimation of uncertainty and a subsequent decline in OOD detection performance.Additionally, EDL encounters the vanishing gradient problem within its fully-connected layers, further degrading classification accuracy.To address these issues, we introduce hyper-domain and propose Hyper-opinion Evidential Deep Learning (HEDL). HEDL extends the evidence modeling paradigm by explicitly integrating sharp evidence, which supports a singular category, with vague evidence that accommodates multiple potential categories.Additionally, we propose a novel opinion projection mechanism that translates hyper-opinion into multinomial-opinion, which is then optimized within the EDL framework to ensure precise classification and refined uncertainty estimation.HEDL integrates evidences across various categories to yield a holistic evidentiary foundation for achieving superior OOD detection. Furthermore, our proposed opinion projection method effectively mitigates the vanishing gradient issue, ensuring classification accuracy without additional model complexity. Extensive experiments over many datasets demonstrate our proposed method outperforms existing OOD detection methods.

We propose a new variant of the Adam optimizer called MicroAdam that specifically minimizes memory overheads, while maintaining theoretical convergence guarantees. We achieve this by compressing the gradient information before it is fed into the optimizer state, thereby reducing its memory footprint significantly. We control the resulting compression error via a novel instance of the classical *error feedback* mechanism from distributed optimization in which *the error correction information is itself compressed* to allow for practical memory gains. We prove that the resulting approach maintains theoretical convergence guarantees competitive to those of AMSGrad, while providing good practical performance. Specifically, we show that MicroAdam can be implemented efficiently on GPUs: on both million-scale (BERT) and billion-scale (LLaMA) models, MicroAdam provides practical convergence competitive to that of the uncompressed Adam baseline, with lower memory usage and similar running time. Our code is available at https://github.com/IST-DASLab/MicroAdam.

The trade-off between cost and performance has been a longstanding and critical issue for deep neural networks. One key factor affecting the computational cost is the width of each layer. However, in practice, the width of layers in a neural network is mostly empirically determined. In this paper, we show that a pattern regarding the variance of weight norm corresponding to different channels can indicate whether the layer is sufficiently wide and may help us better allocate computational resources across the layers. Starting from a simple intuition that channels with larger weights would have larger gradients and the difference in weight norm enlarges between channels with similar weight, we empirically validate that wide and narrow layers show two different patterns with experiments across different data modalities and network architectures. Based on the two different patterns, we identify three stages during training and explain each stage with corresponding evidence. We further propose to adjust the width based on the identified pattern and show that conventional layer width settings for CNNs could be adjusted to reduce the number of parameters while boosting the performance.

The rapid development of Large Language Models (LLMs) has been pivotal in advancing AI, with pre-trained LLMs being adaptable to diverse downstream tasks through fine-tuning. Federated learning (FL) further enhances fine-tuning in a privacy-aware manner by utilizing clients' local data through in-situ computation, eliminating the need for data movement. However, fine-tuning LLMs, given their massive scale of parameters, poses challenges for clients with constrained and heterogeneous resources in FL. Previous methods employed low-rank adaptation (LoRA) for efficient federated fine-tuning but utilized traditional FL aggregation strategies on LoRA adapters. This approach led to mathematically inaccurate aggregation noise, reducing fine-tuning effectiveness and failing to address heterogeneous LoRAs. In this work, we first highlight the mathematical incorrectness of LoRA aggregation in existing federated fine-tuning methods. We introduce a new approach called FLoRA that enables federated fine-tuning on heterogeneous LoRA adapters across clients through a novel stacking-based aggregation method. Our approach is noise-free and seamlessly supports heterogeneous LoRAs. Extensive experiments demonstrate FLoRA's superior performance in both homogeneous and heterogeneous settings, surpassing state-of-the-art methods. We envision this work as a milestone for efficient, privacy-preserving, and accurate federated fine-tuning of LLMs.

In a rapidly evolving job market, skill demand forecasting is crucial as it enables policymakers and businesses to anticipate and adapt to changes, ensuring that workforce skills align with market needs, thereby enhancing productivity and competitiveness. Additionally, by identifying emerging skill requirements, it directs individuals towards relevant training and education opportunities, promoting continuous self-learning and development. However, the absence of comprehensive datasets presents a significant challenge, impeding research and the advancement of this field. To bridge this gap, we present Job-SDF, a dataset designed to train and benchmark job-skill demand forecasting models. Based on millions of public job advertisements collected from online recruitment platforms, this dataset encompasses monthly recruitment demand. Our dataset uniquely enables evaluating skill demand forecasting models at various granularities, including occupation, company, and regional levels. We benchmark a range of models on this dataset, evaluating their performance in standard scenarios, in predictions focused on lower value ranges, and in the presence of structural breaks, providing new insights for further research. Our code and dataset are publicly accessible via the https://github.com/Job-SDF/benchmark.

While numerous Video Violence Detection (VVD) methods have focused on representation learning in Euclidean space, they struggle to learn sufficiently discriminative features, leading to weaknesses in recognizing normal events that are visually similar to violent events (i.e., ambiguous violence). In contrast, hyperbolic representation learning, renowned for its ability to model hierarchical and complex relationships between events, has the potential to amplify the discrimination between visually similar events. Inspired by these, we develop a novel Dual-Space Representation Learning (DSRL) method for weakly supervised VVD to utilize the strength of both Euclidean and hyperbolic geometries, capturing the visual features of events while also exploring the intrinsic relations between events, thereby enhancing the discriminative capacity of the features. DSRL employs a novel information aggregation strategy to progressively learn event context in hyperbolic spaces, which selects aggregation nodes through layer-sensitive hyperbolic association degrees constrained by hyperbolic Dirichlet energy. Furthermore, DSRL attempts to break the cyber-balkanization of different spaces, utilizing cross-space attention to facilitate information interactions between Euclidean and hyperbolic space to capture better discriminative features for final violence detection. Comprehensive experiments demonstrate the effectiveness of our proposed DSRL.

Continual learning has primarily focused on the issue of catastrophic forgetting and the associated stability-plasticity tradeoffs. However, little attention has been paid to the efficacy of continually learned representations, as representations are learned alongside classifiers throughout the learning process. Our primary contribution is empirically demonstrating that existing online continually trained deep networks produce inferior representations compared to a simple pre-defined random transforms. Our approach embeds raw pixels using a fixed random transform, approximating an RBF-Kernel initialized before any data is seen. We then train a simple linear classifier on top without storing any exemplars, processing one sample at a time in an online continual learning setting. This method, called RanDumb, significantly outperforms state-of-the-art continually learned representations across all standard online continual learning benchmarks. Our study reveals the significant limitations of representation learning, particularly in low-exemplar and online continual learning scenarios. Extending our investigation to popular exemplar-free scenarios with pretrained models, we find that training only a linear classifier on top of pretrained representations surpasses most continual fine-tuning and prompt-tuning strategies. Overall, our investigation challenges the prevailing assumptions about effective representation learning in the online continual learning.

Despite the recent observation that large language models (LLMs) can store substantial factual knowledge, there is a limited understanding of the mechanisms of how they acquire factual knowledge through pretraining. This work addresses this gap by studying how LLMs acquire factual knowledge during pretraining. The findings reveal several important insights into the dynamics of factual knowledge acquisition during pretraining. First, counterintuitively, we observe that pretraining on more data shows no significant improvement in the model's capability to acquire and maintain factual knowledge. Next, LLMs undergo forgetting of memorization and generalization of factual knowledge, and LLMs trained with duplicated training data exhibit faster forgetting. Third, training LLMs with larger batch sizes can enhance the models' robustness to forgetting. Overall, our observations suggest that factual knowledge acquisition in LLM pretraining occurs by progressively increasing the probability of factual knowledge presented in the pretraining data at each step. However, this increase is diluted by subsequent forgetting. Based on this interpretation, we demonstrate that we can provide plausible explanations on recently observed behaviors of LLMs, such as the poor performance of LLMs on long-tail knowledge and the benefits of deduplicating the pretraining corpus.

We introduce a novel capacity measure 2sED for statistical models based on the effective dimension. The new quantity provably bounds the generalization error under mild assumptions on the model. Furthermore, simulations on standard data sets and popular model architectures show that 2sED correlates well with the training error. For Markovian models, we show how to efficiently approximate 2sED from below through a layerwise iterative approach, which allows us to tackle deep learning models with a large number of parameters. Simulation results suggest that the approximation is good for different prominent models and data sets.

Multiple types of inference are available for probabilistic graphical models, e.g., marginal, maximum-a-posteriori, and even marginal maximum-a-posteriori. Which one do researchers mean when they talk about ``planning as inference''? There is no consistency in the literature, different types are used, and their ability to do planning is further entangled with specific approximations or additional constraints. In this work we use the variational framework to show that, just like all commonly used types of inference correspond to different weightings of the entropy terms in the variational problem, planning corresponds _exactly_ to a _different_ set of weights. This means that all the tricks of variational inference are readily applicable to planning. We develop an analogue of loopy belief propagation that allows us to perform approximate planning in factored-state Markov decisions processes without incurring intractability due to the exponentially large state space. The variational perspective shows that the previous types of inference for planning are only adequate in environments with low stochasticity, and allows us to characterize each type by its own merits, disentangling the type of inference from the additional approximations that its practical use requires. We validate these results empirically on synthetic MDPs and tasks posed in the International Planning Competition.

Text-to-image generative models have recently attracted considerable interest, enabling the synthesis of high-quality images from textual prompts. However, these models often lack the capability to generate specific subjects from given reference images or to synthesize novel renditions under varying conditions. Methods like DreamBooth and Subject-driven Text-to-Image (SuTI) have made significant progress in this area. Yet, both approaches primarily focus on enhancing similarity to reference images and require expensive setups, often overlooking the need for efficient training and avoiding overfitting to the reference images. In this work, we present the $\lambda$-Harmonic reward function, which provides a reliable reward signal and enables early stopping for faster training and effective regularization. By combining the Bradley-Terry preference model, the $\lambda$-Harmonic reward function also provides preference labels for subject-driven generation tasks. We propose Reward Preference Optimization (RPO), which offers a simpler setup (requiring only 3\% of the negative samples used by DreamBooth) and fewer gradient steps for fine-tuning. Unlike most existing methods, our approach does not require training a text encoder or optimizing text embeddings and achieves text-image alignment by fine-tuning only the U-Net component. Empirically, $\lambda$-Harmonic proves to be a reliable approach for model selection in subject-driven generation tasks. Based on preference labels and early stopping validation from the $\lambda$-Harmonic reward function, our algorithm achieves a state-of-the-art CLIP-I score of 0.833 and a CLIP-T score of 0.314 on DreamBench.