Neural Networks From Scratch, Lecture 2: The Multilayer Perceptron

Neural networks are the dominant class of machine learning algorithms for computer vision and natural language processing today, and their rise in the past 10 years feels like a revolution. But the history of neural networks is not one of revolution, but evolution. Modern neural networks represent decades of fine-tuning of old ideas. In the previous lecture, we discussed linear regression, logistic regression, and gradient descent. Those notes are available on Medium. Now we can build on these concepts to build the state-of-the-art neural network of the 1980s: The Multilayer Perceptron (MLPs). The MLP is a cascade of linear transformations, each followed by a nonlinear activation. The challenge is h ow to learn weights in this nonlinear model. To perform this task, we introduce the backpropagation algorithm. We then analyze the advantages and shortcomings of the MLP.

Speaker(s): Dr. Miguel Dominguez,

Virtual: https://events.vtools.ieee.org/m/278883