Final Year Project

Project Title

Compression Based Perceiver

Description

Biological systems perceive the world by processing high-dimensional inputs from modalities as diverse as vision, audio, touch, etc. However, the perception models in Deep Learning are designed for individual modalities, often relying on domain-specific assumptions. These assumptions introduce helpful biases, but also lock models to individual modalities. The Perceiver, developed by Google DeepMind,is a multi-modal architecture that has low biases, allowing it to scale to multi-modal inputs. However, the architecture is extremely expensive in computational resources, meaning that it cannot be used in most practical applications. The Compression-Based Perceiver is a novel architecture for Computer Vision problems and building up from Google and DeepMind's Perceiver. Our research question evaluates and investigates this architecture and our hypothesis is that if a compression block can compute efficient lower-dimensional data representations, the architecture will be more resource optimized and will remain independent of the input modality, and still maintain comparable accuracy metrics.