Ayodeji Ijishakin's Projects
The Google foobar is a secret hiring challenge which google sends out to indivdiuals based on their (programming related) google searches. The challenge is comprised of 5 levels of difficulty. It has a pyramid structure for the amount of challenges, the first level has 1 challenge, the second has 2, the third 3, fourth 2 and fifth 1.
A repository which features image processing algorithms, of particular use for medical image analysis. The first is the Euclidean distance transform, which converts pixel values within an image to their distance from the foreground of the image.
Solutions to LeetCode problems
Website for the MANIFOLD lab at UCL.
Bayezian Optimisation
This repository contains projects which use machine learning to investigate the neuroscientific underpinnings of psychedelic compounds. The first project uses a non-linear dimensionality reduction technique called diffusion map embedding to investigate how the dimethyl-tryptamine (DMT) breakdown the cortical hierarchy. In normal cognition higher-order cortical regions such as the default mode and executive networks constrain the flow of information within lower-level regions (e.g somatosensory cortices). As a result, ones perception is dictated heavily by the prior predictions encoded within higher-level cortical regions, as opposed to the sensory information which is gathered. The effect of DMT is to set-off a cascade of hyperactivity within high-level cortical regions, due to their high-affinity to the 5HT2A serotonin receptor, which is particularly ubiquitous within these regions. The hyper-activity shuts down regions such as the default mode network, which allows for information to flow more freely throughout the cortex. This in turn allows for the revision of the priors which previously dominated how sensory information is interpreted. Diffusion map embedding (DME) was used to investigate this, as it was applied to functional connectivity data from individuals who had ingested DMT. DME reduces the dimensionality of the functional connectivity data, into axes of functional connectivity variance. The principal axes spans from higher-level cortical regions (e.g default mode network) to the lower-levels ones (e.g. somatosensory). The analysis revealed that this axes is 'flattened' when compared to controls, as there is more integration of the functional connectivity of regions. This maps directly onto the phenomenological reports of individuals who experience aberrant cognition and perception as a result of the decomposition of the cortical hierarchy.
This repository contains code which can be used to train an MLP, convolutional neural network, residually connected convolutional network and generative adversarial network to improve the resolution of abdominal CT images. It works by downsampling a 192x160 2D CT image, by various downsampling factors (2, 4, 6). The downsampled images act as the input data and the normal resolution image is the output. Once trained, unseen CT images, with dimensions which correspond to the downsampling factor can be fed to the network to be super-resolved. Each network can be trained on the various downsampling factors making for 12 networks in total.