This repository contains the crux of my research project with regards to my MSc in neuroimaging at King's College London (2019) .
Here is the abstract from my submitted dissertation:
Context: Current deep convolution neural networks treat vision as a static sequential readout
of pixel information. Active inference defines an agent’s selection of policies (set of actions)
as belief based. In sampling the environment, these beliefs are generatively update in order
to minimize surprise.
Objectives: We hypothesize human vision to be based upon the active inference framework
as we actively seek out salient visual features. Therefore, we infer human vision to be a topdown
process based on our current beliefs about the environment.
Methods: A visual eye-tracking study was performed on healthy participants to analyse gaze
scan paths in trying to identify a scene. The scenes contained partially obstructed handwritten
digits from the MNIST database. To further test our hypothesis, we are developing a Markov
Decision Process (MDP) computational model based on the active inference framework. The
MNIST database was selected as this platform is a benchmark for computer vision algorithm
testing. Behavioural data analysis was performed using a novel method we developed, taking
inspiration from fMRI data analysis, whereby scan paths were converted to volumetric NIfTI
format for SPM analysis.
Results: In all tasks, participants showed better performance in trial duration and less scene
exploration when prior beliefs created an expected scene identity.
Conclusions: Our results support our hypothesis that human vision is a predictive top-down
process, based on our prior beliefs, as per active inference. Results from our MDP model show
how an agent learns to look for salient features for scene identification but is still in
development.
Here, I share my entire work on this project:
- The experimental paradigm I built for data collection (github.com/jonathanmonney/MSC_ActInf/tree/master/01-Paradigm)
- The data analysis methods (github.com/jonathanmonney/MSC_ActInf/tree/master/02-Data_Analysis)
- The submitted thesis (github.com/jonathanmonney/MSC_ActInf/blob/master/JM_thesis_N.pdf)
For any comments or questions, please contact me at jonathanmonney.jm@gmail.com.