Environmental Bayesian Optimization of Gerneral-Purpose Convolutional Neural Networks

Lead Research Organisation: University of Oxford
Department Name: Engineering Science


With the advent of Big Data and modern Deep Learning technology, areas such as computer vision are entering a renaissance. Performance on standard problems such as visual object recognition is soaring, and very difficult challenges such as automatic image captioning can now be tackled with good results. Yet, despite the advances in such specific tasks, deep networks are still far beyond general-purposes intelligences such as humans. A key missing feature is the ability of excelling in *complex environments*, where thousands of interrelated tasks need to be solved together. One of the outstanding technical challenges in scaling deep learning to general-purpose cognition systems is large scale optimization in such complex environments. In this project, we propose to develop Bayesian Optimization as a potential solution to this challenge.
Bayesian Optimization has proved indispensable in automatically tuning the hyper-parameters of commonly used machine learning algorithms, but has been under-explored in the training and deployment phases as well. Nevertheless, the strengths of Bayesian optimization make it an ideal candidate in large scale distributed optimization as well as optimization on multiple tasks. We will demonstrate this technology on computer vision applications. Whereas current vision algorithms focus on specific tasks (e.g. recognizing 1000 object categories or reading text), we will learn general-purpose and efficient computer vision systems from different and complementary tasks.
The company Mathworks is involved. This project falls within the EPSRC information and communication technologies research area.

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/N509711/1 01/10/2016 30/09/2021
1802210 Studentship EP/N509711/1 01/10/2016 30/09/2020 Sylvestre-Alvise Rebuffi
Description We published several papers in the area of Computer Vision. Two works were about how to reuse the knowledge learned by a neural network and apply it on another domain/topic. In another work, we proposed a novel class discovery method which based on the knowledge of several classes of objects can discover and cluster new data in new classes without human supervision.
Exploitation Route Like several applied AI methods, it can have direct applications for industry: communication, medical, ... Our methods are easy to implement and their code is released.
Sectors Aerospace, Defence and Marine,Digital/Communication/Information Technologies (including Software),Healthcare,Other

Description Spotlight presentation at NeurIPS conference 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Postgraduate students
Results and Impact My paper was selected for a spotlight presentation at the NeurIPS conference. This presentation lasted 5 minutes in front of an audience of researchers in the main hall of the conference.
Year(s) Of Engagement Activity 2018