The human brain, with its massive computational capability and power efficiency in small form factor, continues to inspire the ultimate goal of building machines that can perform tasks without being explicitly programmed. In an effort to mimic the natural information processing paradigms observed in the brain, several neural network generations have been proposed over the years. Among the neural networks inspired by biology, second-generation Artificial or Deep Neural Networks (ANNs/DNNs) use memoryless neuron models and have shown unprecedented success surpassing humans in a wide variety of tasks. Unlike ANNs, third-generation Spiking Neural Networks (SNNs) closely mimic biological neurons by operating on discrete and sparse events in time called spikes, which are obtained by the time integration of previous inputs.
Implementation of data-intensive neural network models on computers based on the von Neumann architecture is mainly limited by the continuous data transfer between the physically separated memory and processing units. Hence, non-von Neumann architectural solutions are essential for processing these memory-intensive bio-inspired neural networks in an energy-efficient manner. Among the non-von Neumann architectures, implementations employing non-volatile memory (NVM) devices are most promising due to their compact size and low operating power. However, it is non-trivial to integrate these nanoscale devices on conventional computational substrates due to their non-idealities, such as limited dynamic range, finite bit resolution, programming variability, etc. This dissertation demonstrates the architectural and algorithmic optimizations of implementing bio-inspired neural networks using emerging nanoscale devices.
The first half of the dissertation focuses on the hardware acceleration of DNN implementations. A 4-layer stochastic DNN in a crossbar architecture with memristive devices at the cross point is analyzed for accelerating DNN training. This network is then used as a baseline to explore the impact of experimental memristive device behavior on network performance. Programming variability is found to have a critical role in determining network performance compared to other non-ideal characteristics of the devices. In addition, noise-resilient inference engines are demonstrated using stochastic memristive DNNs with 100 bits for stochastic encoding during inference and 10 bits for the expensive training.
The second half of the dissertation focuses on a novel probabilistic framework for SNNs using the Generalized Linear Model (GLM) neurons for capturing neuronal behavior. This work demonstrates that probabilistic SNNs have comparable perform-ance against equivalent ANNs on two popular benchmarks - handwritten-digit classification and human activity recognition. Considering the potential of SNNs in energy-efficient implementations, a hardware accelerator for inference is proposed, termed as Spintronic Accelerator for Probabilistic SNNs (SpinAPS). The learning algorithm is optimized for a hardware friendly implementation and uses first-to-spike decoding scheme for low latency inference. With binary spintronic synapses and digital CMOS logic neurons for computations, SpinAPS achieves a performance improvement of 4x in terms of GSOPS/W/mm$^2$ when compared to a conventional SRAM-based design.
Collectively, this work demonstrates the potential of emerging memory technologies in building energy-efficient hardware architectures for deep and spiking neural networks. The design strategies adopted in this work can be extended to other spike and non-spike based systems for building embedded solutions having power/energy constraints.
|
![[NJIT logo]](http://archives.njit.edu/vhlib/images/njit_logo.gif){kind=logo}
![[Van Houten Library logo]](images/vhl-top-image.jpg)
