SciData's activity for the Halloween Week 2019 at FDI UCM

• P5 editor project: https://editor.p5js.org/david_ucm/sketches/mHT4YDhnN
• my own artifical intelligence github repo: https://github.com/dlgeraghty/artificial_intelligence
• octopudna: https://dlgeraghty.github.io/artificial_intelligence/drawNeuralNetwork/draw_nn.html
• magenta sketch rnn game: https://magenta.tensorflow.org/assets/sketch_rnn_demo/index.html
• google quick draw game: https://quickdraw.withgoogle.com/#
• tutorial reference for this example: https://www.youtube.com/watch?v=GMaKcRkiNoM
• wikipedia RNN: https://en.wikipedia.org/wiki/Recurrent_neural_network
• Ml5 docs: https://ml5js.org/
• p5 docs: https://p5js.org/
• quick draw dataset github (categories): https://github.com/googlecreativelab/quickdraw-dataset/blob/master/categories.txt
• RDP algorithm: https://www.youtube.com/watch?v=nSYw9GrakjY

The conexion of the nodes that form the neural network form a graph along a tempopral sequence.
This helps the model behave according to a given time instance.
The key that makes them good for speech and hand written digits recognition is the fact that they use a special kind of their own memory.

There are two different types:

• Finite impulse -> can be easily converted to feedforward
• Inifinite impulse -> Can not be converted to feedforward as they form a cyclick graph.

The most simple form of it, is organised as a feedforward would be , but the connections are represented as in a graph. Each neuron can be activated or deactivated.

• Ml5 library! This makes it super easy to do this otherwise complex task and allows to represent it in the browser which is a good thing.
• P5 js! This allows for easy drawing and manipulation of drawings in the canvas, it is also super easy to use!
• In this example, we want to do something similar to the magenta sketch rnn , but as we will see, there will be some key differences.
• Key functions and values:
  1. Preload function: value in the ml5.sketchRNN('value') is the object that we will be moldelling.
  2. sketchRNNstart: sketchRNN.generate(prevPath, newPath) allows us to give it a certain path and then it continues from there.
  3. All the drawing stuff is handled in the draw loop (this is all p5 js and some magic).