Neural Architecture Transfer 2: A Paradigm for Improving Efficiency in Multi-Objective Neural Architecture Search

Deep learning is increasingly impacting various aspects of contemporary society. Artificial neural networks have emerged as the dominant models for solving an expanding range of tasks. The introduction of Neural Architecture Search (NAS) techniques, which enable the automatic design of task-optimal networks, has led to remarkable advances. However, the NAS process is typically associated with long execution times and significant computational resource requirements. Once-For-All (OFA) and its successor, Once-For-All-2 (OFAv2), have been developed to mitigate these challenges. While maintaining exceptional performance and eliminating the need for retraining, they aim to build a single super-network model capable of directly extracting sub-networks satisfying different constraints. Neural Architecture Transfer (NAT) was developed to maximise the effectiveness of extracting sub-networks from a super-network. In this paper, we present NATv2, an extension of NAT that improves multi-objective search algorithms applied to dynamic super-network architectures. NATv2 achieves qualitative improvements in the extractable sub-networks by exploiting the improved super-networks generated by OFAv2 and incorporating new policies for initialisation, pre-processing and updating its networks archive. In addition, a post-processing pipeline based on fine-tuning is introduced. Experimental results show that NATv2 successfully improves NAT and is highly recommended for investigating high-performance architectures with a minimal number of parameters.

Enhancing Once-For-All: A Study on Parallel Blocks, Skip Connections and Early Exits

The use of Neural Architecture Search (NAS) techniques to automate the design of neural networks has become increasingly popular in recent years. The proliferation of devices with different hardware characteristics using such neural networks, as well as the need to reduce the power consumption for their search, has led to the realisation of Once-For-All (OFA), an eco-friendly algorithm characterised by the ability to generate easily adaptable models through a single learning process. In order to improve this paradigm and develop high-performance yet eco-friendly NAS techniques, this paper presents OFAv2, the extension of OFA aimed at improving its performance while maintaining the same ecological advantage. The algorithm is improved from an architectural point of view by including early exits, parallel blocks and dense skip connections. The training process is extended by two new phases called Elastic Level and Elastic Height. A new Knowledge Distillation technique is presented to handle multi-output networks, and finally a new strategy for dynamic teacher network selection is proposed. These modifications allow OFAv2 to improve its accuracy performance on the Tiny ImageNet dataset by up to 12.07% compared to the original version of OFA, while maintaining the algorithm flexibility and advantages.

Anticipate, Ensemble and Prune: Improving Convolutional Neural Networks via Aggregated Early Exits

Today, artificial neural networks are the state of the art for solving a variety of complex tasks, especially in image classification. Such architectures consist of a sequence of stacked layers with the aim of extracting useful information and having it processed by a classifier to make accurate predictions. However, intermediate information within such models is often left unused. In other cases, such as in edge computing contexts, these architectures are divided into multiple partitions that are made functional by including early exits, i.e. intermediate classifiers, with the goal of reducing the computational and temporal load without extremely compromising the accuracy of the classifications. In this paper, we present Anticipate, Ensemble and Prune (AEP), a new training technique based on weighted ensembles of early exits, which aims at exploiting the information in the structure of networks to maximise their performance. Through a comprehensive set of experiments, we show how the use of this approach can yield average accuracy improvements of up to 15% over traditional training. In its hybrid-weighted configuration, AEP's internal pruning operation also allows reducing the number of parameters by up to 41%, lowering the number of multiplications and additions by 18% and the latency time to make inference by 16%. By using AEP, it is also possible to learn weights that allow early exits to achieve better accuracy values than those obtained from single-output reference models.