Efficient Multi-Objective NeuroEvolution in Computer Vision and Applications for Threat Identification.

Abstract

Concealed threat detection is at the heart of critical security systems designed to en- sure public safety. Currently, methods for threat identification and detection are primarily manual, but there is a recent vision to automate the process. Problematically, developing computer vision models capable of operating in a wide range of settings, such as the ones arising in threat detection, is a challenging task involving multiple (and often conflicting) objectives. Automated machine learning (AutoML) is a flourishing field which endeavours to dis- cover and optimise models and hyperparameters autonomously, providing an alternative to classic, effort-intensive hyperparameter search. However, existing approaches typ- ically show significant downsides, like their (1) high computational cost/greediness in resources, (2) limited (or absent) scalability to custom datasets, (3) inability to provide competitive alternatives to expert-designed and heuristic approaches and (4) common consideration of a single objective. Moreover, most existing studies focus on standard classification tasks and thus cannot address a plethora of problems in threat detection and, more broadly, in a wide variety of compelling computer vision scenarios. This thesis leverages state-of-the-art convolutional autoencoders and semantic seg- mentation (Chapter 2) to develop effective multi-objective AutoML strategies for neural architecture search. These strategies are designed for threat detection and provide in- sights into some quintessential computer vision problems. To this end, the thesis first introduces two new models, a practical Multi-Objective Neuroevolutionary approach for Convolutional Autoencoders (MONCAE, Chapter 3) and a Resource-Aware model for Multi-Objective Semantic Segmentation (RAMOSS, Chapter 4). Interestingly, these ap- proaches reached state-of-the-art results using a fraction of computational resources re- quired by competing systems (0.33 GPU days compared to 3150), yet allowing for mul- tiple objectives (e.g., performance and number of parameters) to be simultaneously op- timised. This drastic speed-up was possible through the coalescence of neuroevolution algorithms with a new heuristic technique termed Progressive Stratified Sampling. The presented methods are evaluated on a range of benchmark datasets and then applied to several threat detection problems, outperforming previous attempts in balancing multiple objectives. The final chapter of the thesis focuses on thread detection, exploiting these two mod- els and novel components. It presents first a new modification of specialised proxy scores to be embedded in RAMOSS, enabling us to further accelerate the AutoML process even more drastically while maintaining avant-garde performance (above 85% precision for SIXray). This approach rendered a new automatic evolutionary Multi-objEctive method for cOncealed Weapon detection (MEOW), which outperforms state-of-the-art models for threat detection in key datasets: a gold standard benchmark (SixRay) and a security- critical, proprietary dataset. Finally, the thesis shifts the focus from neural architecture search to identifying the most representative data samples. Specifically, the Multi-objectIve Core-set Discovery through evolutionAry algorithMs in computEr vision approach (MIRA-ME) showcases how the new neural architecture search techniques developed in previous chapters can be adapted to operate on data space. MIRA-ME offers supervised and unsupervised ways to select maximally informative, compact sets of images via dataset compression. This operation can offset the computational cost further (above 90% compression), with a minimal sacrifice in performance (less than 5% for MNIST and less than 13% for SIXray). Overall, this thesis proposes novel model- and data-centred approaches towards a more widespread use of AutoML as an optimal tool for architecture and coreset discov- ery. With the presented and future developments, the work suggests that AutoML can effectively operate in real-time and performance-critical settings such as in threat de- tection, even fostering interpretability by uncovering more parsimonious optimal models. More widely, these approaches have the potential to provide effective solutions to chal- lenging computer vision problems that nowadays are typically considered unfeasible for AutoML settings.