The Connectionist Engine (Deep Learning & CNNs)

The Limits of Classic ML

• Last Week: We used an SVM. We had to manually "flatten" the image, destroying all spatial information (e.g., which pixel is next to which).
• The Problem: If you shift a digit one pixel to the left, the flattened 784-number array looks completely different to the SVM!
• The Solution:Deep Learning. We will build a Neural Network that looks at 2D images directly and learns its own features.

1D Flattened Array vs 2D Spatial Grid