[ExI] i'll bee seeing you... was: RE: old software fun (spike)

Anders Sandberg anders at aleph.se
Tue Jul 14 10:35:24 UTC 2015


Just an update: I saw a poster here at the IJCNN conference that is relevant:

Effective Insect Recognition Using a Stacked Autoencoder with Maximum Correntropy Criterion
Yu Qi, Goktug Cinar, Vinicius Souza, Gustavo Batista, Yueming Wang and Jose Principe
http://www.cnel.ufl.edu/files/1425197934.pdf

Basically, they look at input data detected using a laser (but one of the told me he thought a microphone would work). Then they train an autoencoder (a feedforward network mapping a N neuron input via a smaller bottleneck layer to an identical N neuron output) on the basic sound patterns. The bottleneck layer represents the data more efficiently, and once it has been trained those values are used as inputs to another network that is trained to classify the sounds as insects. According to the poster the correntropy method for training is slightly better than standard methods of training. 

This might be a decent bee detector. 



Anders Sandberg, Future of Humanity Institute Philosophy Faculty of Oxford University

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20150714/c9711e88/attachment.html>


More information about the extropy-chat mailing list