Modeling Temporal Dynamics in the Classification of Auditory Signals
Abstract
We have pursued an exploratory modeling approach to the response properties of auditory neurons in the zebra finch. Early attempts at modeling auditory neurons in the song system nucleus HVc were not successful, in part because of the relative paucity of biological information at the time of the start of the project and in part because the technology applying neural nets to quantitative neurophysiological data had not been established. Thus, although our ultimate goal is to model HVc auditory neurons, we have started our analysis at the level of the auditory thalamus (ovoidalis), where neurons have simpler response properties. In general, the temporal dynamics of neuronal response is a critical feature to capture in any modeling effort, yet this feature is among the most difficult to model. We have employed a 3-layered connectionist architecture called Time Delay Neural Network (TDNN), which permits a natural representation of time-varying processes and exhibits the essential property of temporal invarience. The supervised learning approach we have taken to training the network (backpropagation of errors) involves applying inputs to the network, and comparing the network output with the desired output. This difference or error signal is then used to alter the values of the weights of the network so as to decrease the error signal in subsequent iterations
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1993
- Accession Number
- ADA267472
Entities
People
- Daniel Margoliash
Organizations
- University of Chicago