Biology-inspired methods for pitch recognition

New publication on biology-inspired methods for pitch recognition

Members of the Department of Theoretical Electrical Engineering, in cooperation with Fraunhofer IDMT, have published an article demonstrating new insights in the field of bio-inspired information processing. A computer model based on biologically plausible principles simulates audio signal processing in a neuromorphic hardware description layer. Octopus cells of the inferior colliculi are modeled that transform incoming signals into rememberable sensory input.

Dendrites of octopus cells selectively link in their local receptive fields via synapses with signal-conducting fibers of the auditory nerve. Selectively triggered signaling flows terminate at the soma of the octopus cell. We postulate for the first time in the article a new backpropagation learning rule. It states that all involved preactivated synapses are incremented when they have triggered a backpropagation signal at the soma.

The result shows that learned weights can be used to learn a pathway selection that results in a reduction of interval times to the main input tone.

Results of the simulation with the created model of an octopus network. The network consists of 11 octopus neurons overlapping with 9 dendrites each connecting to the auditory nerve fibers. a) The output in the form of inverted spike intervals showing clustering around the central interval of 3.83 ms associated with tone C4. b) Adjustment of synapse weights in the learning procedure. c) Statistical distribution of the inter-spike intervals and representation of the median (orange line).

F. Feldhoff, H. Toepfer, T. Harczos, F. Klefenz
Periodicity Pitch Perception Part III: Sensibility and Pachinko Volatility.
Frontiers in Neuroscience, 16


Ansprechpartner: M.Sc. Frank Feldhoff