A simple implementation of a forward propopagation network without any external libraries.
Just a quick think about perceptrons
Train a network to regocnize tetris-shapes from scratch.
Why? To get a deeper understanding of neural networks.
How: Python numpy for the matris store
Todo: Identify the network model. Do some test of activiation functions.
The 5 tetromino shapes
Straight: vertical and horizontal reflection symmetry, and two-fold rotational symmetry Square: vertical and horizontal reflection symmetry, and four-fold rotational symmetry T: vertical reflection symmetry only L: no symmetry S: two-fold rotational symmetry only
Other notes. Thinking aloud. Cosine similarity loss function. Nope. categorical cross-entropy is better for this kind of classification.