Natural Signals Properties and the Convolution

Explore the fundamentals of convolutional neural networks in this comprehensive lecture by Alfredo Canziani. Delve into natural signal properties, including stationarity and locality in both 1D and 2D contexts. Examine the concept of compositionality in 2D signals and review fully connected networks. Understand how locality leads to sparsity and stationarity results in parameter sharing. Learn about 1D kernels, padding techniques, and the application of ConvNets to image processing. Discover pooling methods and compare fully connected networks with convolutional networks using a Jupyter Notebook. Investigate the effects of deterministic pixel shuffling on signal properties before concluding with a final comparison and summary.

Syllabus

– Happy birthday to the TAs!
– Today topic: convolutional neural nets
– Input layer, points, and signals
– Natural signal properties
– 1D stationarity
– 1D locality
– 2D stationarity
– 2D locality
– 2D compositionality
– Fully connected recap
– Locality ⇒ sparsity
– Stationarity ⇒ parameter sharing
– 1D kernels
– 1D padding
– ConvNet for images and tensor reshaping
– Pooling
– Jupyter Notebook: fully connected vs. convnet
– Deterministic pixel shuffling: breaking signal properties
– Final comparison
– Goodbye