Early exits enable the network's forward pass to terminate early by attaching trainable internal classifiers to the backbone network. Existing early-exit methods typically adopt either a joint training approach, where the backbone and exit heads are trained simultaneously, or a disjoint approach, where the heads are trained separately. However, the implications of this choice are often overlooked, with studies typically adopting one approach without adequate justification. This choice influences training dynamics and its impact remains largely unexplored. In this paper, we introduce a set of metrics to analyze early-exit training dynamics and guide the choice of training strategy. We demonstrate that conventionally used joint and disjoint regimes yield suboptimal performance. To address these limitations, we propose a mixed training strategy: the backbone is trained first, followed by the training of the entire multi-exit network. Through comprehensive evaluations of training strategies across various architectures, datasets, and early-exit methods we present strengths and weaknesses of the early exit training strategies. In particular, we show consistent improvements in performance and efficiency using the proposed mixed strategy.

Foundation models for computer vision built on Vision Transformer (ViT) architectures have become increasingly widespread. However, their fine-tuning process is resource-intensive, slowing their adoption in edge or low-energy applications. We introduce ALaST ( Adaptive Layer Selection for ViT Fine-Tuning ), a novel approach that dynamically optimizes the fine-tuning process to significantly reduce computational cost, memory consumption, and training time. Our method is founded on the critical observation that during fine-tuning, the importance of individual layers and tokens varies substantially across training iterations and depends on the specific mini-batch being processed. ALaST leverages this insight by adaptively estimating layer importance at each fine-tuning step and allocating computational resources—or “compute budgets”—proportionally. Layers assigned lower budgets are either trained with a reduced token set or temporarily frozen. Through comprehensive empirical evaluation on standard benchmarks, we demonstrate that ALaST achieves substantial efficiency gains: up to 1.3 × reduction in training time, 1.5 × reduction in FLOPs, and 2 × decrease in memory requirements, all while maintaining model performance within 0.5 % of full fine-tuning. Notably, our approach provides an automatic schedule for distributing computational resources across layers and can be combined with existing parameter-efficient fine-tuning techniques, offering an orthogonal dimension of optimization for Vision Transformers.

Early-exit neural networks (EENNs) endow a standard deep neural network (DNN) with early-exit classifiers (EECs) to provide predictions at intermediate points of the processing when enough confidence in classification is achieved. This leads to many benefits in terms of effectiveness and efficiency. Currently, the design of EENNs is carried out manually by experts, a complex and time-consuming task that requires accounting for many aspects, including the correct placement, the thresholding, and the computational overhead of the EECs. For this reason, the research is exploring the use of neural architecture search (NAS) to automate the design of EENNs. Currently, few comprehensive NAS solutions for EENNs have been proposed in the literature, and a fully automated, joint design strategy taking into consideration both the backbone and the EECs remains an open problem. To this end, this work presents neural architecture search for hardware-constrained early exit neural networks (NACHOS), the first NAS framework for the design of optimal EENNs satisfying constraints on the accuracy and the number of multiply and accumulate (MAC) operations performed by the EENNs at inference time. In particular, this provides the joint design of backbone and EECs to select a set of admissible (i.e., respecting the constraints) Pareto optimal solutions in terms of the best trade-off between the accuracy and the number of MACs. The results show that the models designed by NACHOS are competitive with the state-of-the-art EENNs. Additionally, this work investigates the effectiveness of two novel regularization terms designed for the optimization of the auxiliary classifiers of the EENN.

Recent research has found that neural networks are vulnerable to several types of adversarial attacks, where the input samples are modified in such a way that the model produces a wrong prediction that misclassifies the adversarial sample. In this paper we focus on black-box adversarial attacks, that can be performed without knowing the inner structure of the attacked model, nor the training procedure, and we propose a novel attack that is capable of correctly attacking a high percentage of samples by rearranging a small number of pixels within the attacked image. We demonstrate that our attack works on a large number of datasets and models, that it requires a small number of iterations, and that the distance between the original sample and the adversarial one is negligible to the human eye.

In this paper, we propose a novel ensembling technique for deep neural networks, which is able to drastically reduce the required memory compared to alternative approaches. In particular, we propose to extract multiple sub-networks from a single, untrained neural network by solving an end-to-end optimization task combining differentiable scaling over the original architecture, with multiple regularization terms favouring the diversity of the ensemble. Since our proposal aims to detect and extract sub-structures, we call it Structured Ensemble. On a large experimental evaluation, we show that our method can achieve higher or comparable accuracy to competing methods while requiring significantly less storage. In addition, we evaluate our ensembles in terms of predictive calibration and uncertainty, showing they compare favourably with the state-of-the-art. Finally, we draw a link with the continual learning literature, and we propose a modification of our framework to handle continuous streams of tasks with a sub-linear memory cost. We compare with a number of alternative strategies to mitigate catastrophic forgetting, highlighting advantages in terms of average accuracy and memory.