Simultaneous growth in the volume of available data along with rapid advancements in computing and hardware technology have paved the way for unprecedented breakthroughs in the field of Artificial Intelligence (AI). In particular, a modern class of AI algorithms, dubbed Deep Learning (DL), has shown great promise by achieving or even surpassing human-level capabilities in many tasks. The rise of DL has brought forth a new industrial revolution by taking over the modern landscape of smart applications, e.g., self-driving cars, virtual assistants, drug discovery, and manufacturing. Nevertheless, to date, there exist quite a few challenges for wide-scale adoption of DL in real-life scenarios.Firstly, confidence characterization and ensuring robustness of DL-enabled services is imperative, particularly in safety-critical autonomous systems. Secondly, concerns over the scalability and efficiency of DL hinder its training and deployment on diverse hardware platforms. This dissertation addresses the above-mentioned challenges via a holistic customization of DL algorithm and system from the standpoint of task-based metrics (e.g., accuracy), physical constraints (e.g., memory and power budget), as well as new design metrics that facilitate DL integration in safety-sensitive tasks. The presented research in this dissertation interlinks theoretical fundamentals, domain-specific architecture design, and automated tools that enable co-optimization of the DL algorithm with the underlying platform while satisfying various constraints. The key contributions of this dissertation are as follows:
1) Devising CuRTAIL, the first end-to-end and automated framework that simultaneously enables efficient and safe execution of DL models in face of adversarial attacks. CuRTAIL formalizes the goal of thwarting adversarial attacks as an optimization problem and trains parallel defense modules to minimize vulnerability. The proposed framework leverages hardware/algorithm co-design and customized acceleration to enable just-in-time execution in resource-constrained settings.
2) Designing a novel framework, dubbed ACCHASHTAG, which identifies any faults occurring during DL inference in real time. I propose to summarize the ground-truth DL model as a unique hash signature, which is used to verify the model’s integrity on the fly. Notably, ACCHASHTAG, for the first time, provides guaranteed lower bounds on the detection rate using a formal statistical analysis of hash collision.
3) Proposing CLEANN, the first end-to-end framework that enables online mitigation of backdoor, a.k.a. Trojan, attacks on DL. CLEANN uses sparse recovery and statistical analysis to identify incoming Trojan samples and remove their effect on the victim model’s prediction. I design the algorithmic solutions as well as customized hardware-accelerated engines to enable real-time DL model decision verification via CLEANN.
4) Innovating an approach for restructuring inter-layer connections in DL models, leading to faster convergence to a desired accuracy during training. This is achieved by transforming the DL model into a small-world network using principles from graph theory. The obtained DL model, dubbed SWANN, is a highly-connected, small-world topology with enhanced signal propagation characteristics and faster learning speed.
5) Developing LTS, the first training-free, hardware-aware neural architecture search for autoregressive Transformers. The proposed method delivers high-performance specialized architectures for inference on a target hardware. The core of LTS is an ultra-low-cost proxy that can estimate the performance of candidate architectures without any need for training. Using this novel proxy, the search can be performed entirely on the target hardware, allowing us to incorporate hardware measurements, e.g., peak memory utilization and latency, within the architecture search loop.
6) Automating DL model customization for various target hardware by formulating it as a constrained optimization. The optimization goal is to compress a large model to satisfy given accuracy and hardware performance constraints. I propose a highly-scalable blackbox optimizer, dubbed AdaNS, to solve the aforesaid optimization problem. AdaNS leverages adaptive non-uniform sampling with carefully crafted probabilistic distributions to locate and reconstruct the optimization objective function around its maximizers.