ABSTRACT
DNNs have revolutionized the approach of generating high-accurate predictions for various tasks. However, they are computationally and memory intensive. Hence, their deployment on resource-constrained devices is challenging. Moreover, the current trends in the ML community demonstrated that, despite having high learning capabilities, advanced DL architectures pose even more stringent constraints due to their high complexity. In addition, DNNs suffer from various vulnerability threats that undermine their integrity and question their practical deployments in safety-critical applications. This book tackles these challenges by exploiting the potential of energy reductions and security improvements of advanced DL systems. To enable this, novel techniques are proposed at both the software and hardware levels. Multi-objective techniques are employed to achieve cross-layer optimizations for energy efficiency and robustness. The high complexity of advanced DL models like CapsNets and SNNs requires dedicated designs and optimizations for energy efficiency while offering unique possibilities for enhancing their robustness.
