Progressive intelligence for low-power devices
Progressive intelligence for low-power devices
As our world has become increasingly advanced so too has the pervasiveness of technology throughout it. This has led to a desire to have these technological capabilities embedded in the devices we interact with on a daily basis, this spans from entire industrial sectors down to individual consumer devices. Artificial intelligence, in particular machine learning, is one of these advancements which is progressing at an unprecedented rate. However, it is also becoming incredibly resource intensive to operate. Progressive intelligence is a solution to this issue, which approaches the machine learning inference process incrementally, first making low-cost--low-confidence predictions and improving on these incrementally according to application requirements.
This work explores progressive intelligence, and discusses where it fits in the machine learning field. Machine learning approaches can be split into data-based and model-based approaches, it is shown progressive intelligence can be implemented in both. In the data-based paradigm an upper bound on error as the number of samples increases is derived. In the model-based paradigm branched neural networks are selected as a good system on which to implement progressive intelligence ideas. A comprehensive study of progressive intelligence in branched neural networks is undertaken, from their training to their inference processes. Following this, the novel concept of an early-exit reject is introduced, where three scenarios are studied in which samples can be rejected at inference time to save on computational resources.
University of Southampton
Dymond, Jack
3309494f-9bac-4891-a6a7-06a900eebf87
2024
Dymond, Jack
3309494f-9bac-4891-a6a7-06a900eebf87
Gunn, Steve
306af9b3-a7fa-4381-baf9-5d6a6ec89868
Stein, Sebastian
cb2325e7-5e63-475e-8a69-9db2dfbdb00b
Dymond, Jack
(2024)
Progressive intelligence for low-power devices.
University of Southampton, Doctoral Thesis, 182pp.
Record type:
Thesis
(Doctoral)
Abstract
As our world has become increasingly advanced so too has the pervasiveness of technology throughout it. This has led to a desire to have these technological capabilities embedded in the devices we interact with on a daily basis, this spans from entire industrial sectors down to individual consumer devices. Artificial intelligence, in particular machine learning, is one of these advancements which is progressing at an unprecedented rate. However, it is also becoming incredibly resource intensive to operate. Progressive intelligence is a solution to this issue, which approaches the machine learning inference process incrementally, first making low-cost--low-confidence predictions and improving on these incrementally according to application requirements.
This work explores progressive intelligence, and discusses where it fits in the machine learning field. Machine learning approaches can be split into data-based and model-based approaches, it is shown progressive intelligence can be implemented in both. In the data-based paradigm an upper bound on error as the number of samples increases is derived. In the model-based paradigm branched neural networks are selected as a good system on which to implement progressive intelligence ideas. A comprehensive study of progressive intelligence in branched neural networks is undertaken, from their training to their inference processes. Following this, the novel concept of an early-exit reject is introduced, where three scenarios are studied in which samples can be rejected at inference time to save on computational resources.
Text
JackDymond-Final-Thesis
- Version of Record
Text
Final-thesis-submission-Examination-Mr-Jack-Dymond
Restricted to Repository staff only
More information
Published date: 2024
Identifiers
Local EPrints ID: 492900
URI: http://eprints.soton.ac.uk/id/eprint/492900
PURE UUID: 02b532fd-422f-4964-a3bf-6727f0d61b88
Catalogue record
Date deposited: 20 Aug 2024 16:31
Last modified: 21 Aug 2024 01:41
Export record
Contributors
Author:
Jack Dymond
Thesis advisor:
Steve Gunn
Thesis advisor:
Sebastian Stein
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics