Open Lab

OL14 Machine learning on non-uniformly sampled data

Background

Human brain is an amazing organ and one of the greatest computers! There is a lot to be learnt from the way brain handles signal. Most of our sensory signal is preprocessed by an organ called hypothalamus. One of the preprocessing that it applies on signal is scaling it into a log-scale. So that, signal near point of interest is sampled finely and signal far off are sampled coarsely. One of our work was on pre-processing of images with a log-polar scale (the way thalamus does with the signal from eye) and check its implications on machine learning (ML) algorithms.

Aim/goal/research question

Design an ML algorithm with log-scale preprocessing as an embedded part. Determine its relative merits and demerits. Design it to become as much autonomous as possible so that it performs the log-scale pre-processing with minimal effect on the ML algorithm.  (Optional) Work on finding analytical limits/bounds on the loss of performance for the ML algorithms.

Method

In this project, the student would work on the design of pre-processing techniques for signals where the sampling is converted from linear to logarithmic. Then the student would investigate the gain that the ML algorithm gets from this step.  As the next step, the student would try to generalise it to non-uniformly sampled data and check how ML algorithms can deal with such data. A nice reference on this is : https://open.bu.edu/bitstream/handle/2144/1918/03.021.pdf?sequence=1&isAllowed=y.

Recommended past experience/interest

Strong DSP skills. Exposure to ML.

Related Work

TBD

Contact