Dataset distillation aims to generate a smaller but representative subset from a large dataset, which allows a model to be trained efficiently, meanwhile evaluating on the original testing data distribution to achieve decent performance. Many prior works have aimed to align with diverse aspects of the original datasets, such as matching the training weight trajectories, gradient, feature/BatchNorm distributions, etc. In this work, we show how to distill various large-scale datasets such as full ImageNet-1K/21K under a conventional input resolution of 224$\times$224 to achieve the best accuracy over all previous approaches, including SRe$^2$L, TESLA and MTT. To achieve this, we introduce a simple yet effective ${\bf C}$urriculum ${\bf D}$ata ${\bf A}$ugmentation ($\texttt{CDA}$) during data synthesis that obtains the accuracy on large-scale ImageNet-1K and 21K with 63.2% under IPC (Images Per Class) 50 and 36.1% under IPC 20, respectively. Finally, we show that, by integrating all our enhancements together, the proposed model beats the current state-of-the-art by more than 4% Top-1 accuracy on ImageNet-1K/21K and for the first time, reduces the gap to its full-data training counterpart to less than absolute 15%. Moreover, this work represents the inaugural success in dataset distillation on larger-scale ImageNet-21K under the standard 224$\times$224 resolution. Our code and distilled ImageNet-21K dataset of 20 IPC, 2K recovery budget are available at //github.com/VILA-Lab/SRe2L/tree/main/CDA.
相關內容
A basic question within the emerging field of mechanistic interpretability is the degree to which neural networks learn the same underlying mechanisms. In other words, are neural mechanisms universal across different models? In this work, we study the universality of individual neurons across GPT2 models trained from different initial random seeds, motivated by the hypothesis that universal neurons are likely to be interpretable. In particular, we compute pairwise correlations of neuron activations over 100 million tokens for every neuron pair across five different seeds and find that 1-5\% of neurons are universal, that is, pairs of neurons which consistently activate on the same inputs. We then study these universal neurons in detail, finding that they usually have clear interpretations and taxonomize them into a small number of neuron families. We conclude by studying patterns in neuron weights to establish several universal functional roles of neurons in simple circuits: deactivating attention heads, changing the entropy of the next token distribution, and predicting the next token to (not) be within a particular set.
Strategies synthesized using formal methods can be complex and often require infinite memory, which does not correspond to the expected behavior when trying to model Multi-Agent Systems (MAS). To capture such behaviors, natural strategies are a recently proposed framework striking a balance between the ability of agents to strategize with memory and the model-checking complexity, but until now has been restricted to fully deterministic settings. For the first time, we consider the probabilistic temporal logics PATL and PATL* under natural strategies (NatPATL and NatPATL*, resp.). As main result we show that, in stochastic MAS, NatPATL model-checking is NP-complete when the active coalition is restricted to deterministic strategies. We also give a 2NEXPTIME complexity result for NatPATL* with the same restriction. In the unrestricted case, we give an EXPSPACE complexity for NatPATL and 3EXPSPACE complexity for NatPATL*.
Variational regularization is commonly used to solve linear inverse problems, and involves augmenting a data fidelity by a regularizer. The regularizer is used to promote a priori information and is weighted by a regularization parameter. Selection of an appropriate regularization parameter is critical, with various choices leading to very different reconstructions. Classical strategies used to determine a suitable parameter value include the discrepancy principle and the L-curve criterion, and in recent years a supervised machine learning approach called bilevel learning has been employed. Bilevel learning is a powerful framework to determine optimal parameters and involves solving a nested optimization problem. While previous strategies enjoy various theoretical results, the well-posedness of bilevel learning in this setting is still an open question. In particular, a necessary property is positivity of the determined regularization parameter. In this work, we provide a new condition that better characterizes positivity of optimal regularization parameters than the existing theory. Numerical results verify and explore this new condition for both small and high-dimensional problems.
We propose a new method for estimating subject-specific mean functions from longitudinal data. We aim to do this in a flexible manner (without restrictive assumptions about the shape of the subject-specific mean functions), while exploiting similarities in the mean functions between different subjects. Functional principal components analysis fulfils both requirements, and methods for functional principal components analysis have been developed for longitudinal data. However, we find that these existing methods sometimes give fitted mean functions which are more complex than needed to provide a good fit to the data. We develop a new penalised likelihood approach to flexibly model longitudinal data, with a penalty term to control the balance between fit to the data and smoothness of the subject-specific mean curves. We run simulation studies to demonstrate that the new method substantially improves the quality of inference relative to existing methods across a range of examples, and apply the method to data on changes in body composition in adolescent girls.
Knowledge distillation has emerged as a highly effective method for bridging the representation discrepancy between large-scale models and lightweight models. Prevalent approaches involve leveraging appropriate metrics to minimize the divergence or distance between the knowledge extracted from the teacher model and the knowledge learned by the student model. Centered Kernel Alignment (CKA) is widely used to measure representation similarity and has been applied in several knowledge distillation methods. However, these methods are complex and fail to uncover the essence of CKA, thus not answering the question of how to use CKA to achieve simple and effective distillation properly. This paper first provides a theoretical perspective to illustrate the effectiveness of CKA, which decouples CKA to the upper bound of Maximum Mean Discrepancy~(MMD) and a constant term. Drawing from this, we propose a novel Relation-Centered Kernel Alignment~(RCKA) framework, which practically establishes a connection between CKA and MMD. Furthermore, we dynamically customize the application of CKA based on the characteristics of each task, with less computational source yet comparable performance than the previous methods. The extensive experiments on the CIFAR-100, ImageNet-1k, and MS-COCO demonstrate that our method achieves state-of-the-art performance on almost all teacher-student pairs for image classification and object detection, validating the effectiveness of our approaches.
Distributed antennas must be phase-calibrated (phase-synchronized) for certain operations, such as reciprocity-based joint coherent downlink beamforming, to work. We use rigorous signal processing tools to analyze the accuracy of calibration protocols that are based on over-the-air measurements between antennas, with a focus on scalability aspects for large systems. We show that (i) for some who-measures-on-whom topologies, the errors in the calibration process are unbounded when the network grows; and (ii) despite that conclusion, it is optimal -- irrespective of the topology -- to solve a single calibration problem for the entire system and use the result everywhere to support the beamforming. The analyses are exemplified by investigating specific topologies, including lines, rings, and two-dimensional surfaces.
Machinery for data analysis often requires a numeric representation of the input. Towards that, a common practice is to embed components of structured data into a high-dimensional vector space. We study the embedding of the tuples of a relational database, where existing techniques are often based on optimization tasks over a collection of random walks from the database. The focus of this paper is on the recent FoRWaRD algorithm that is designed for dynamic databases, where walks are sampled by following foreign keys between tuples. Importantly, different walks have different schemas, or "walk schemes", that are derived by listing the relations and attributes along the walk. Also importantly, different walk schemes describe relationships of different natures in the database. We show that by focusing on a few informative walk schemes, we can obtain tuple embedding significantly faster, while retaining the quality. We define the problem of scheme selection for tuple embedding, devise several approaches and strategies for scheme selection, and conduct a thorough empirical study of the performance over a collection of downstream tasks. Our results confirm that with effective strategies for scheme selection, we can obtain high-quality embeddings considerably (e.g., three times) faster, preserve the extensibility to newly inserted tuples, and even achieve an increase in the precision of some tasks.
The adoption of continuous shrinkage priors in high-dimensional linear models has gained momentum, driven by their theoretical and practical advantages. One of these shrinkage priors is the R2D2 prior, which comes with intuitive hyperparameters and well understood theoretical properties. The core idea is to specify a prior on the percentage of explained variance $R^2$ and to conduct a Dirichlet decomposition to distribute the explained variance among all the regression terms of the model. Due to the properties of the Dirichlet distribution, the competition among variance components tends to gravitate towards negative dependence structures, fully determined by the individual components' means. Yet, in reality, specific coefficients or groups may compete differently for the total variability than the Dirichlet would allow for. In this work we address this limitation by proposing a generalization of the R2D2 prior, which we term the Generalized Decomposition R2 (GDR2) prior. Our new prior provides great flexibility in expressing dependency structures as well as enhanced shrinkage properties. Specifically, we explore the capabilities of variance decomposition via logistic normal distributions. Through extensive simulations and real-world case studies, we demonstrate that GDR2 priors yield strongly improved out-of-sample predictive performance and parameter recovery compared to R2D2 priors with similar hyper-parameter choices.
We propose an adjusted Wasserstein distributionally robust estimator -- based on a nonlinear transformation of the Wasserstein distributionally robust (WDRO) estimator in statistical learning. The classic WDRO estimator is asymptotically biased, while our adjusted WDRO estimator is asymptotically unbiased, resulting in a smaller asymptotic mean squared error. Meanwhile, the proposed adjusted WDRO has an out-of-sample performance guarantee. Further, under certain conditions, our proposed adjustment technique provides a general principle to de-bias asymptotically biased estimators. Specifically, we will investigate how the adjusted WDRO estimator is developed in the generalized linear model, including logistic regression, linear regression, and Poisson regression. Numerical experiments demonstrate the favorable practical performance of the adjusted estimator over the classic one.
Humans perceive the world by concurrently processing and fusing high-dimensional inputs from multiple modalities such as vision and audio. Machine perception models, in stark contrast, are typically modality-specific and optimised for unimodal benchmarks, and hence late-stage fusion of final representations or predictions from each modality (`late-fusion') is still a dominant paradigm for multimodal video classification. Instead, we introduce a novel transformer based architecture that uses `fusion bottlenecks' for modality fusion at multiple layers. Compared to traditional pairwise self-attention, our model forces information between different modalities to pass through a small number of bottleneck latents, requiring the model to collate and condense the most relevant information in each modality and only share what is necessary. We find that such a strategy improves fusion performance, at the same time reducing computational cost. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple audio-visual classification benchmarks including Audioset, Epic-Kitchens and VGGSound. All code and models will be released.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191