GDescent

 The site is : playground.tensorflow.org

One can play with the parameters to a neural network problem and run

the algorithm. At this point, just using the two splotches of dots from the left

as a sample program.

Starting with random weights to a linear equation, one builds up knowledge 

through iteration. Gradient descent (Cauchy) identify the direction in which

the weight measurements should move.

So one begins with a perceptron - whih only does linear relationsships, as seen in

Supervised Learning - and one propagtes error backward for a better fit.

There are refinements to consider, testing in smaller subsets to avoid being misled

by non-generalizable data. One also adjusts for overfitting by running partial

versions of the model.Ultimately, one works with a library for 'robust' results.

                                                            *     *     *

Training data for recognizing handwritten numers...