Machine Learning

Machine Learning is what mad scientists like to call "Artificial Intelligence" these days. We call it that because we promised to make AI's really quickly a long time ago and failed. We don't say we are working on AI anymore, we just talk about Machine Learning. We've adjusted our goals. But ML is the first step towards AI.

Methods

Linear Regression

In school, you should have learned the standard linear equation:

Y = mX + b

You put in values of X, positive or negative, picked from a line we can show going left to right, and you get back values on the Y axis, which we show vertically:

The vertical position is computed from the horizontal position by multiplying it by some number we call "m" and then adding another number, we call "b".

If we have a bunch of points, taken from real world data, and we write a program that finds values for m and b which calculate a Y from each points X, which is pretty close to the actual points Y, then that program has learned to predict where new points will probably appear on that graph.

This is called Linear Regression and it's the first step in understanding Machine Learning.

• http://techref.massmind.org/techref/method/ai/LinearRegresion.htm
• http://www.code-in-javascript.com/linear-regression-in-javascript/ cached
• https://github.com/Tom-Alexander/regression-js

Neural Nets

Neural Networks use Regression to train a slightly more complex formula which also makes predictions based on it's inputs. Instead of one X, there are many X's and each has it's own m. But let's change the letter from "m" to "w" for "Weight". Like:

Y = w₁X₁ + w₂X₂ ... + b

And we connect many of these formulas, each called a node, together to form a network:

• http://playground.tensorflow.org Watch a neural net train and design one that can learn faster / better.
• http://techref.massmind.org/techref/method/ai/NeuralNets.htm How NN work and how to write them.

Demos

• https://storage.googleapis.com/tfjs-examples/webcam-transfer-learning/dist/index.html
  Needs a webcam. Take pictures of yourself looking up, down, left, right by clicking the associated direction. Take at least 20 pictures of each pose, moving slightly between each one. Once you have all the pictures, click "Training" and wait for the loss to drop to a very low number. Then press play. Look in the direction you wish to move.

• https://magenta.tensorflow.org/demos/performance_rnn/index.html#3|2,0,1,0,1,1,0,1,0,1,0,1|1,1,1,1,1,1,1,1,1,1,1,1|1,1,1,1,1,1,1,1,1,1,1,1|false
  Like a piano player "noodling" around. Not playing a specific song, but playing things that could be parts of songs. The computer is making up sounds that are "musical" in nature.