Statistical Analysis and Modeling of Robot Data for Category Formation

Supervisor: Dr. Katrien Beuls (katrien@ai.vub.ac.be)

Part 1: Statistical Analysis of Robot Data - Identify Statistical Regularities

* Load all data and compute different statistics, explore and visualize

dependencies

* Compute statistical measures: moments

* Dimension reduction: PCA, MDS, ISOMAP - extract channels and channel

dependencies for maximizing variance and dispersion

* Apply Vector Quantization algorithms: SOM, GNG, sparse coding

 

Part 2: Discrimination games (Which channels matter for distinguishing

objects?)

* Learn classifiers that are most distinctive for sets of 2, 4, 8, 16, … objects

* Test different algorithms for maximizing distance between objects: vector

quantization, Prototypes, weighted likelihood classifiers, PCA/LDA, decision

trees

 

Part 3: One-Word Naming Games (first link to language)

* 1) Learn classifiers in discrimination games, learn to attach names

* 2) Learn classifiers and names together

* Use different strategies and methods: cross-situational statistics, single-channel

biases, shaping

 

Part 4: Multi-Word Naming Games

* Learn classifiers in discrimination game, learn names (but allow multi-word)

* Learn classifiers and names together

* Use different strategies and methods: cross-situational statistics, single-channel

biases, shaping

 

Outcome: paper in IROS or similar

Literature: wikipedia (PCA, MDS, ISOMAP, SOM); Spranger (2012); Wellens,

Loetzsch, Steels (2010)

Software: numpy, scikit-learn, scipy, bokeh or similar