Binaural echo error detection gives self-supervised learning to improve landmark classification

Binaural processing is typically associated with source localization using interaural time and level differences. This paper describes its role for improving landmark classification from an echo sequence. A brain-inspired system explores the blind human echolocation problem of differentiating two foliage targets with different sized leaves using audible echoes. A biomimetic sonar views each target by producing left and right-ear monaural echo waveforms whose target-specific frequency power spectra are classified using template matching. Binaural processing implements error detection by ignoring views that give contradictory monaural classifications. Binaural classification using consistent monaural classifications also produces self-labeled data that update monaural template estimates to continually improve target classification accuracy through additional target encounters. Monaural classification error probability starts at 0.225 after supervised learning with 10 encounters with each known target and reaches 0.134 after an additional 90 self-supervised template updates. The binaural classification error probability starts at 0.032 with an average of 1.49 binaural views required to achieve consistent monaural classifications and reaches 0.019 with an average of 1.32 views.