Week 3

$$\gdef \sam #1 {\mathrm{softargmax}(#1)}$$ $$\gdef \vect #1 {\boldsymbol{#1}} $$ $$\gdef \matr #1 {\boldsymbol{#1}} $$ $$\gdef \E {\mathbb{E}} $$ $$\gdef \V {\mathbb{V}} $$ $$\gdef \R {\mathbb{R}} $$ $$\gdef \N {\mathbb{N}} $$ $$\gdef \relu #1 {\texttt{ReLU}(#1)} $$ $$\gdef \D {\,\mathrm{d}} $$ $$\gdef \deriv #1 #2 {\frac{\D #1}{\D #2}}$$ $$\gdef \pd #1 #2 {\frac{\partial #1}{\partial #2}}$$ $$\gdef \set #1 {\left\lbrace #1 \right\rbrace} $$
🎙️ Yann LeCun

Lecture part A

We first see a visualization of a 6-layer neural network. Next we begin with the topic of Convolutions and Convolution Neural Networks (CNN). We review several types of parameter transformations in the context of CNNs and introduce the idea of a kernel, which is used to learn features in a hierarchical manner. Thereby allowing us to classify our input data which is the basic idea motivating the use of CNNs.

Lecture part B

We give an introduction on how CNNs have evolved over time. We discuss in detail different CNN architectures, including a modern implementation of LeNet5 to exemplify the task of digit recognition on the MNIST dataset. Based on its design principles, we expand on the advantages of CNNs which allows us to exploit the compositionality, stationarity, and locality features of natural images.

Practicum

Properties of natural signals that are most relevant to CNNs are discussed in more detail, namely: Locality, Stationarity, and Compositionality. We explore precisely how a kernel exploits these features through sparsity, weight sharing and the stacking of layers, as well as motivate the concepts of padding and pooling. Finally, a performance comparison between FCN and CNN was done for different data modalities.


📝