EEG Event Classification w/ Dilated CNN’s and Multi-Task Learning; A Comparison (Theory + Results + Code)

Dilated Convolutions (DL's) are pretty cool. If you haven't heard of them before, I'd recommend http://www.inference.vc/dilated-convolutions-and-kronecker-factorisation/ as a good starting point to the benefits of it. One of the first applications of it in ML seems to be in Multi-Scale Context Aggregation By Dilated Convolutions, which primarily discusses its application in semantic segmentation (in which it performs …

Conditionally Thresholded CNN’s for Weakly-Supervised Image Segmentation (Results)

The first post can be found here. I realized a few things in the past two weeks. 1. MS COCO is difficult. Like really difficult. Difficult to the point where certain images have the object-of-interest being around 40~ pixels in total size. In fact, it was just last year that weakly-supervised segmentation was first attempted …

Conditionally Thresholded CNN’s for Weakly-Supervised Image Segmentation (Theory)

Consider the average Convolutional Neural Network. For the uninitiated, there are a class of NN's that consider the immediate neighborhood of a parameters when optimizing it's weights connections. In the case of images, these 'parameters' are a single pixel; which means that instead of the network considering the importance of singular pixels versus other, singular …