Cortically-Inspired Overcomplete Feature Learning for Colour Images

Cowley, Benjamin; Kneller, Adam; Thornton, John

doi:10.1007/978-3-319-13560-1_57

Cited by 2 publications

(2 citation statements)

References 17 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The inclusion of the SP reduced all classification accuracies (although not to the same degree for the Insects set 4 ) suggesting that the SP encoding is degrading the salience of the g/ERB wavelet features, at least the features relevant for a k -NN classifier. This is an unexpected result, given that SP has performed well in other related domains [18] and suggests there is a missing element in our model of audition. One possibility is that the g/ERB wavelet is not a good model of the cochlear signal and so does not capture the salient features required for higher level neocortical processing.…”

Section: Resultsmentioning

confidence: 80%

See 1 more Smart Citation

A Cortically-Inspired Model for Bioacoustics Recognition

Main

Thornton

2015

Neural Information Processing

Self Cite

View full text Add to dashboard Cite

Wavelet transforms have shown superior performance in auditory recognition tasks compared to the more commonly used Mel-Frequency Cepstral Coefficients, and offer the ability to more closely model the frequency response behaviour of the cochlear basilar membrane. In this paper we evaluate a gammatone wavelet as a preprocessor for the Hierarchical Temporal Memory (HTM) model of the neocortex as part of the broader development of a biologically motivated approach to sound recognition. Specifically, we apply for the first time, a gammatone/equivalent rectangular bandwidth wavelet transform in conjunction with the HTM's Spatial Pooler to recognise frog calls, bird songs and insect sounds. Our audio feature detection results show that wavelets perform considerably better than MFCCs on our selected datasets but that combining wavelets with HTM does not produce further improvements. This outcome raises questions concerning the degree of match to the biology required for an effective HTM-based model of audition.

show abstract

Section: Resultsmentioning

confidence: 80%

“…These properties are implemented in the interaction of artificial cortical mini-columns. Research on HTMs has steadily developed over the past ten years, with a focus on image processing [18].…”

Section: Hierarchical Temporal Memorymentioning

confidence: 99%