Event Details
Deep Neural Networks Implemented in Fixed-Point Arithmetic
Presenter: Lucy Blecha
Supervisor:
Date: Wed, October 7, 2026
Time: 11:00:00 - 00:00:00
Place: Zoom - see below.
ABSTRACT
Zoom meeting link:
Join Zoom Meeting
https://uvic.zoom.us/j/85271931486?pwd=hwpXiuLtzZ0FEj2lq6GDPAICWmyXCx.1
Meeting ID: 852 7193 1486
Password: 463097
Abstract: The classification performance of a deep neural network implemented in fixed-point arithmetic is investigated. A synthetic classification problem is defined, which consists of 5-dimensional concentric hyper-spheres corresponding to five non-linearly separable classes. Compared to a baseline implementation in 64-bit floating-point arithmetic, our fixed-point deep neural network achieves nearly the same classification accuracy with a word length of less than 16 bits. This is a promising result, indicating that the implementation complexity of deep neural networks can be significantly reduced, with beneficial effects in latency and power consumption.