Intelligent Robotics

Master’s Thesis Topic on Discrimative Filtering/Feature Extraction for Classification of Tactile Signals

Supervisor: Professor Ville Kyrki (ville.kyrki@aalto.fi)

Advisor: Samuli Hynninen (samuli.hynninen@aalto.fi)

The Intelligent Robotics Group is providing a Master’s thesis topic for a student interested in robotic perception and signal processing in robotics. The main goal of the thesis project is to explore and develop novel filtering and/or feature extraction methods to enhance the classifiability of force-torque signals. The student is provided with a dataset of real-world force-torque measurements, collected from a robot interacting in a standardized way with different materials. However, ideally the developed methods work for diverse sets of data, making a general contribution to the field of discriminative filtering and feature extraction in machine learning.

Discriminative filtering is an umbrella term for a collection of diverse signal processing techniques used to enhance the classifiability of the signals for downstream classification tasks. The established methods include, for example, the Common Spatial Patterns method [1], which aims to maximize the ratio of the variances of data from two different classes, or Linear Discrimnant Analysis, which aims to minimize the intraclass variation while maximizing the interclass variation. Also, a whole variety of methods to design discriminative filter banks exist, often working in frequency domain or it’s variations such as Mel-Frequency Cepstral Coefficient space.

Feature extraction/selection, in turn, can be considered an even more general umbrella term, including discriminative filtering but also methods such as Minimum Redundancy Maximum Relevance Algorithm [2], ReliefF algorithm [3], and Laplacian Score [4].

The student is provided with the freedom to use their own discretion when it comes to the selection of proper family of methodologies, but one interesting approach could be to use traditional infinite impulse response filters, such as Butterworth or Chebyshev filters, so that an optimality criterion on their discriminative properties is set.

Deliverables:

– An algorithm for discriminative filtering or feature extraction

– A thesis providing a theoretical background for the developed methods

Pre-requisites:

– Good understanding of engineering mathematics

– Basic programming skills

– Previous knowledge on the basics of machine learning and signal processing

References:
[1] Koles et al. (1990) Spatial patterns underlying population differences in the background EEG

[2] Ding et al. (2005) Minimum redundancy feature selection from microarray gene expression data

[3] Robnik-Sikonja and Kononenko (2003) Theoretical and Empirical Analysis of ReliefF and RreliefF

[4] He et al (2005) Laplacian Score for Feature Selection