CGSC170 Lecture Notes - Lecture 8: Terry Sejnowski, The Algorithm, Pattern Recognition
20 Dec 2019
School
Department
Course
Professor
Document Summary
The algorithm needs to find a way of calculating error in hidden units that do not have a target activation. It does this but calculating for each hidden unit its degree of responsibility for error at the output units. This error value is used to adjust the weights of hidden units. Hidden units do not have targeted output. Distinguish between the sonar echo produced by a rock and the sonar echo produced by a mine. Each input unit is dedicated to a different frequency. The input unit"s activation level is based on the level of energy in the sonar echo at the frequency. Each input unit is customized to a frequency band. Use patterns of echos to categorize rock or mine. Vector of activation values defined over the input units. Reflects the unique fingerprint of each sonar echo. Weights change every time the network comes up with a wrong output.
