Deep learning with deep neural networks (DNNs) has become one of the most popular choice for realizing artificial intelligence systems. To apply the advanced technology of deep learning to various fields, it is necessary not only to nurture deep learn...
Deep learning with deep neural networks (DNNs) has become one of the most popular choice for realizing artificial intelligence systems. To apply the advanced technology of deep learning to various fields, it is necessary not only to nurture deep learning experts, but also to enable non-experts to gain DNNs with high-level performance. As an approach based on automatic machine learning, neural architecture search (NAS), which can automatically obtain a DNN architecture yielding the promising performance in the target domain, is garnering considerable interest as a practical alternative to the manual design for DNNs. Precise execution of NAS requires the tremendous search cost, but the weight-sharing makes NAS estimate the performance of DNNs, thereby significantly reducing the search cost. Based on the weight-sharing, diverse NAS methods have been proposed and produced results that exceed the performance of hand-crafted DNNs in various fields. This dissertation contains methods and substantial results on three essential research topics in NAS, including how to reduce the search cost, how to increase search performance by improving search spaces and search algorithms, and how to facilitate NAS for domains that overlook the architectural importance.
The first research is to reduce the search cost of NAS methods without search performance degradation. The significant computational search cost of NAS is a hindrance to researchers who want to employ NAS to their domains. To address this issue, we introduce proxy data, i.e., a representative subset of the target data, for the acceleration of NAS methods. Even though data selection has been used across various fields in deep learning, our evaluation for existing data selection methods on NAS benchmarks reveals that they are not suitable for NAS and a new data selection method is necessary. Based on in-depth analysis on proxy data constructed using these selection methods through data entropy, we propose a novel proxy data selection method tailored for NAS. To empirically demonstrate the effectiveness, we conduct thorough experiments across diverse datasets, search spaces, and NAS algorithms. Consequently, NAS algorithms with the proposed data selection method discover architectures that are competitive with those obtained using the entire dataset. It significantly reduces the search cost: executing DARTS with the proposed data selection method requires only 40 minutes on CIFAR-10 and 7.5 hours on ImageNet with a single GPU. Furthermore, as the inverse approach from the conventional NAS, when the architecture searched on our proxy data of ImageNet is transferred to the smaller datasets, i.e., CIFAR-10, a test error of 2.4% is yielded.
In the second research, we improve the search performance of differentiable NAS working on a cell-based search space by refining the search space. Through continuous relaxation on a super-network cell, differentiable NAS methods can approximately evaluate various cells during the search process. When a cell is derived from an optimized super-network cell, a number of the evaluated cells, which are not included in the conventional cell-based search space but may achieve a high performance, are inevitably discarded by the conventional argmax selection rule. To overcome the limitation of the conventional cell derivation, we extend the search space and propose a top-k survival rule to derive a cell belonging in the extended search space. Meanwhile, we observe that existing NAS methods experience a significant decrease in search performance when coupled with the top-k survival rule. To properly explore the extended search space, we propose BtNAS that individually models the operation strengths as continuous random variables following Beta distributions. BtNAS approximates these true posterior distributions using the variational Bayes method, and the parameters of the variational Beta distributions are trained through gradient-based optimization and the pathwise derivative estimator. BtNAS derives cells that are competitive to those recently reported by differentiable NAS methods and yield networks with fewer parameters. In addition, we adjust operation strengths to make all intermediate node in the super-network cell to be activated, and then achieve a state-of-the-art test error of 2.3% on CIFAR-10.
The last research is to exploit NAS to search for an energy-efficient spiking neural network (SNNs). SNNs, which mimic information transmission in brains with neurons and synapses, have received considerable attention. SNNs can efficiently process spatio-temporal information through event-driven computations with discrete and sparse spikes. Most previous studies focused on training methods to improve the performance and energy-efficiency of SNNs, and the architectural effect regarding SNNs was rarely studied. To facilitate NAS for SNNs, we propose a spike-aware neural architecture search framework, named AutoSNN. We first construct a two-level search space for SNNs by identifying which architectural factors are more suitable for SNNs than for conventional DNNs. Afterwards, to estimate the performance and the spike generation of candidate architectures in the search space, AutoSNN adopts the one-shot super-network training scheme based on weight-sharing and the evolutionary search algorithm. The architecture fitness value used for the architecture evaluation is specifically designed for considering both the accuracy and number of spikes during the search process. An SNN searched by AutoSNN outperforms handcrafted SNNs, in terms of both accuracy and energy-efficiency. We also demonstrate the superiority of AutoSNN on various datasets, including neuromorphic datasets.
Throughout this dissertation, we propose methods for developing an efficient and effective NAS framework and show a strategy to utilize NAS according to the three important research topics. To validate the proposed methods, we provide a significant amount of experimental results, thereby demonstrating that the architecture choice must be considered to maximize the performance of deep learning. Therefore, owing to the approaches introduced in this dissertation, it is expected that NAS will be actively employed and researchers from diverse fields can easily obtain promising results by adopting deep learning with NAS.