So I’ve been wondering what problem to work on for my learning while I work my way through #practicaldeeplearning. I think I’ve found the perfect dataset: https://figshare.com/collections/_/4560497

It’s got everything. The raw 12-lead ECG readings as well as arrhythmia classifications for over 10,000 patients. And best of all it’s easy to convert the ECG readings to images for #deeplearning. Will probably start by trying out #ResNet initially.

Will keep posting my progress here.

Sometimes when creating shortcuts on #ResNet people use a MaxPooling2D or even a convolution.

If you want to pass the identity of your input, you can use a Convolution (with no gradient) initialized with dirac_ (delta function A.K.A identity function for a convolution).

As you see is faster than using a MaxPooling:
#NeuralNetwork #PyTorch

Linking Neural Collapse and L2 Normalization with Improved Out-of-Distribution Detection in Deep ...

Jarrod Haas, William Yolland, Bernhard T Rabus

https://openreview.net/forum?id=fjkN5Ur2d6

#normalization #deep #resnet

To share more than just an introduction post, here's a fun bit of research that I was involved with during the 2019 FDL programme - https://www.climatechange.ai/papers/neurips2019/11 - #ResNet #CloudPhysics #FDL #MODIS #ClimateChange #iResNet #CloudClassification #GCP