• Home
  • Eye Science
    • Planning Ahead
    • Popout
    • Background
    • Publications
  • NomiNote
  • Machine Learning
  • About & Details
    • Contact
Jelmer Philip
  • Home
  • Eye Science
    • Planning Ahead
    • Popout
    • Background
    • Publications
  • NomiNote
  • Machine Learning
  • About & Details
    • Contact

Machine learning and AI

AI Experience 

I obtained my Masters from the Utrecht University in Computer Science with a specialisation in Technical Artificial Intelligence. At the time, the focus on neural networks and machine learning was still limited. Therefore, I utilised every opportunity I got to learn more about the topic. For my elective Bachelor project I implemented an adaptation of the echo state network to predict stock exchange rates. In order to be more broadly acquainted with machine learning techniques, for my Master's I decided to implement a multi-layered (or deep as one would call it today) object recognition system utilising both supervised (feed-forward neural networks) and unsupervised (self-organising maps) learning techniques.
Predicting Stock Exchange Rates
My first interaction with neural networks was in 2003 when I built and compared a simple feed-forward network trained with back-propagation to an echo state network incorporating an evolutionary component to update the weights in the recurrent layer of the neural network (in the classic echo state network weights are randomly assigned and this obviously has drawbacks). While certainly the predictions were not perfect comparing the feedforward network to our Echo State approach we found that the echo state network outperformed the standard feed-forward network on our chosen metric, as well as, on most alternative metrics.
Object Recognition
For my Masters I built an object recognition system. The system consisted of several layers and I implemented a number of techniques to extract features from images. Anything from compact radial representations of edge crossings to using filters inspired by the response properties of the neurons in the our own visual cortex.
I utilized both supervised and unsupervised machine learning components. For classification of shapes and objects I used feed-forward neural networks trained with back-propagation. To build representations of combinations of shapes I used self-organising maps (a type of vector quantisation).
Data Science
While I would not refer to myself as data scientist my academic career required extensive data processing, filtering and inferential statistics. Eye movements are abundant and in my experiments, participants would make up to 5 a second. Moving from raw gaze coordinates sampled every millisecond, to movements, to a quantifiable metric naturally involves a lot of data processing. Moreover, testing hypotheses requires inferential statistics. Over time I have performed anything from simple analysis of variance to Bayesian estimation of parameter distributions using MCMC algorithms.
Powered by Create your own unique website with customizable templates.
  • Home
  • Eye Science
    • Planning Ahead
    • Popout
    • Background
    • Publications
  • NomiNote
  • Machine Learning
  • About & Details
    • Contact