Learning Hierarchical Spectral Representations of Human Speech with the Information Dynamics of Thinking
As a first implementation of Information Dynamics of Thinking theory, this thesis sought to investigate how the system responds to human speech input data. After formalizing the processes of segmentation, categorization, and abstraction central to the operation of the theory, the implemented system found evidence of semantic categories at higher levels of abstraction consistent with similar human speech syllables – the first indication of the viability of this cognitive architecture.
The Maximum Entropy Bayesian Actor Critic algorithm alters standard actor critic algorithms with the hope of ameliorating the problems of sample inefficiency and convergence brittleness that hamstring Deep RL methods. By combining Bayesian actor critic for its sample efficiency and maximum entropy actor critic for its robust convergence, MEBAC hopes to retain the best of both of these methods and avoid the problems faced by Deep RL altogether.
Maximum Entropy Bayesian Actor Critic
As a master student, Steve Homer formalized and implemented the first version of the Information Dynamics of Thinking. Now as a PhD student under Geraint Wiggins, his primary interests lie in computation cognitive science, where he looks to combine concepts from information theory, Bayesian probability theory, and analytical geometry to better understand the human mind.