AI Hardware Acceleration in ASIC Design Lecture 2: Neural Networks – An Overview

This lecture provides a foundational overview of neural networks, starting with the basic building block: the perceptron (neuron). It details how these neurons function through weighted inputs, summation, and activation, and how connecting multiple neurons creates layers and ultimately, complex networks. The core mathematical operation driving these networks is highlighted as the multiply-accumulate (MAC) operation, which scales rapidly with network size and complexity. The lecture emphasizes that the sheer number of MAC operations is a primary driver of the computational demands – and associated power consumption – of modern AI systems.
The lecture then explores different neural network architectures, including deep neural networks, recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and generative pre-trained transformers (GPTs). It explains how each architecture addresses specific challenges and applications, such as sequential data processing or large language modeling. Finally, the lecture touches upon the training process, emphasizing the massive datasets and computational resources required to optimize the weights and connections within these networks, ultimately leading to a fixed-function system for inference.