Introduction to Neural Networks

Neural networks have become a nearly ubiquitous tool in visual perception systems for robots.

This is due largely to two key properties that they hold: compositionality and differentiability.

Together, these allow us to compose the atomic building blocks (neurons) and learn parameters by propagating gradients.

We briefly cover the simplest neural network architecture, the multi-layer perceptron, and discuss how it can learn through a process called “stochastic gradient descent” (SGD).