Week 3 – Practicum: Natural signals properties and CNNs

Course website: http://bit.ly/DLSP20-web
Playlist: http://bit.ly/pDL-YouTube
Speaker: Alfredo Canziani
Week 3: http://bit.ly/DLSP20-03

0:00:00 – Week 3 – Practicum

PRACTICUM: http://bit.ly/DLSP20-03-3
Properties of signals that are most relevant to CNNs are discussed, namely:- Locality, Stationarity, and Compositionality. How a kernel exploits these features by using Sparsity, Weight sharing and Stacking of layers is explored next, along with the concepts of padding and pooling. A performance comparison between FCN and CNN for different data modalities was also made.
0:00:26 – Properties of natural signals
0:17:54 – Exploiting Properties of Natural Signals to Build Standard Spatial CNN
0:39:36 – Pooling and Covnet - Jupyter Notebook

#Deep Learning #Yann LeCun #NYU #PyTorch #convolutional neural networks #CNN #ConvNet #locality #stationarity #compositionality #sparsity #parameter sharing #hierarchical #receptive field #kernels #padding #pooling