Thesis
Visual feature extraction through brain inspired algorithms : towards efficiency, accuracy and continual learning
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2025
- Thesis identifier
- T17248
- Person Identifier (Local)
- 202067948
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- Machine learning and Artificial intelligence have already revolutionised the worldwe live in. Nevertheless, these technologies are expected to progress even further and advance living standards far beyond today’s reality. To achieve thisrevolution, several limitations of current AI need to be addressed. Demands oncomputing resources and energy supply are a major obstacle, which ultimatelylimit the capabilities of the AI we can deploy at the edge. Furthermore, currentsystems are ill-suited for continual learning (CL) on real world data, as they require to train with all available data as an independent and identically distributed(i.i.d.) set. This thesis focuses on these problems by working on visual featureextraction, and contributing to more efficient and accurate algorithms, with thecapacity for CL.With the objective of developing energy efficient feature extraction, a majorpart of the thesis is focused on Spiking Neural Networks (SNNs). SNNs have become an interesting alternative to conventional artificial neural networks (ANN)thanks to their temporal processing capabilities and energy efficient implementations in neuromorphic hardware. However, the challenges involved in trainingthem have limited their performance in terms of accuracy and thus their applications. Improving learning algorithms and neural architectures for a more accuratefeature extraction is therefore a priority. Contributing towards this aim, this workpresents a study on the key components of modern spiking architectures, an indepth study on the possible implementations of spiking residual connections, anda novel spiking version of the successful residual network architecture. Additionally, the effect of different state of the art techniques are empirically compared inimage classification tasks to provide a state of the art guide to SNN design. Finally, the proposed network outperforms previous SNN architectures in multipledatasets, while using less parameters.In order to exploit SNNs for more efficient AI, it is also of interest to understand the full scope of their exploitable properties. These networks are characterised by their unique temporal dynamics, but the properties and advantages of such computations are still not fully understood. In order to provide answers, inthis work it is demonstrated how spiking neurons can enable temporal feature extraction in feed-forward neural networks without the need for recurrent synapses,and how recurrent SNNs can achieve comparable results to LSTM with a smallernumber of parameters. This shows how their bio-inspired computing principlescan be successfully exploited beyond energy efficiency gains, and evidences theirdifferences with respect to conventional artificial neural networks. These resultsare obtained through a new task, DVS-Gesture-Chain (DVS-GC), which allows,for the first time, to evaluate the perception of temporal dependencies in a realevent-based action recognition dataset. Furthermore, this setup allows to revealthe role of the leakage rate in spiking neurons for temporal processing tasks anddemonstrated the benefits of ”hard reset” mechanisms.Finally, the focus is switched to the capacity for training feature extractorsin continual learning scenarios, a major milestone for the development of trulyautonomous systems and artificial general intelligence. The challenge in thissetup is to avoid catastrophic forgetting, where artificial systems forget previousknowledge if they are trained in new data without revisiting the old. Often, themethods used to alleviate forgetting make use of either rehearsal buffers, pretrained backbones or external indication of the task to solve. However, these requirements result in severe limitations regarding scalability, privacy preservationand efficient deployment. This work explores how to eliminate the need for suchrequirements and proposes a new method, Low Interference Feature ExtractionSub-networks (LIFES). Additionally, the study breaks down the CatastrophicForgetting (CF) problem into 4 causes, allowing to better understand the effectof CL methods. The proposed LIFES algorithm achieves competitive results instandard incremental learning scenarios, providing an alternative to approacheswith more restrictive requirements. Moreover, it provides solutions for specificcauses of the CF problem, making it complementary to other methods
- Advisor / supervisor
- Di Caterina, Gaetano
- Resource Type
- DOI
- Funder
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
PDF of thesis T17248 | 2025-05-07 | Public | Download |