Successful adoption of deep learning (DL) in the wild requires models to be: (1) compact, (2) accurate, and (3) robust to distributional shifts. Unfortunately, efforts towards simultaneously meeting these requirements have mostly been unsuccessful. This raises an important question: "Is the inability to create Compact, Accurate, and Robust Deep neural networks (CARDs) fundamental?" To answer this question, we perform a large-scale analysis of popular model compression techniques which uncovers several intriguing patterns. Notably, in contrast to traditional pruning approaches (e.g., fine tuning and gradual magnitude pruning), we find that "lottery ticket-style" approaches can surprisingly be used to produce CARDs, including binary-weight CARDs. Specifically, we are able to create extremely compact CARDs that, compared to their larger counterparts, have similar test accuracy and matching (or better) robustness--simply by pruning and (optionally) quantizing. Leveraging the compactness of CARDs, we develop a simple domain-adaptive test-time ensembling approach (CARD-Deck) that uses a gating module to dynamically select appropriate CARDsfrom the CARD-Deckbased on their spectral-similarity with test samples. The proposed approach builds a "winning hand" of CARDsthat establishes a new state-of-the-art [8] on CIFAR-10-C accuracies (i.e., 96.8% standard and 92.75% robust) and CIFAR-100-C accuracies (i.e., 80.6% standard and 71.3% robust) with better memory usage than non-compressed baselines (pretrained CARDs available at [8]). Finally, we provide theoretical support for our empirical findings.

Continual lifelong learning is essential to many applications.

Neural Information Processing Systems http://nips.cc/

Deep model compression has been extensively studied, and state-of-the-art methods can now achieve high compression ratios with minimal accuracy loss.

Vision Transformers (ViTs) are widely used in a variety of applications, while they usually have a fixed architecture that may not match the varying computational resources of different deployment environments. Thus, it is necessary to adapt ViT architectures to devices with diverse computational overheads to achieve an accuracy-efficient trade-off. This concept is consistent with the motivation behind Learngene. To achieve this, inspired by polynomial decomposition in calculus, where a function can be approximated by linearly combining several basic components, we propose to linearly decompose the ViT model into a set of components called learngenes during element-wise training. These learngenes can then be recomposed into differently scaled, pre-initialized models to satisfy different computational resource constraints. Such a decomposition-recomposition strategy provides an economical and flexible approach to generating different scales of ViT models for different deployment scenarios. Compared to model compression or training from scratch, which require to repeatedly train on large datasets for diverse-scale models, such strategy reduces computational costs since it only requires to train on large datasets once. Extensive experiments are used to validate the effectiveness of our method: ViTs can be decomposed and the decomposed learngenes can be recomposed into diverse-scale ViTs, which can achieve comparable or better performance compared to traditional model compression and pre-training methods. The code for our experiments is available in the supplemental material.

Knowledge Distillation (KD) has emerged as a promising technique for model compression but faces critical limitations: (1) sensitivity to hyperparameters requiring extensive manual tuning, (2) capacity gap when distilling from very large teachers to small students, (3) suboptimal coordination in multi-teacher scenarios, and (4) inefficient use of computational resources. We present \textbf{HPM-KD}, a framework that integrates six synergistic components: (i) Adaptive Configuration Manager via meta-learning that eliminates manual hyperparameter tuning, (ii) Progressive Distillation Chain with automatically determined intermediate models, (iii) Attention-Weighted Multi-Teacher Ensemble that learns dynamic per-sample weights, (iv) Meta-Learned Temperature Scheduler that adapts temperature throughout training, (v) Parallel Processing Pipeline with intelligent load balancing, and (vi) Shared Optimization Memory for cross-experiment reuse. Experiments on CIFAR-10, CIFAR-100, and tabular datasets demonstrate that HPM-KD: achieves 10x-15x compression while maintaining 85% accuracy retention, eliminates the need for manual tuning, and reduces training time by 30-40% via parallelization. Ablation studies confirm independent contribution of each component (0.10-0.98 pp). HPM-KD is available as part of the open-source DeepBridge library.

Neural Information Processing Systems http://nips.cc/