Deep Learning in a Browser¶

with ConvNetJS¶

Follow along at https://5x5x5x5.github.io/convNetJStalk/

State of the Art Applications¶

• Speech Recognition
• Image Recognition

Burdened by Terminology¶

These techniques do not capture the complexity of brains, but rather were inspired by what was know about neuron function in the '50s.

These terms might be the best we've got, but don't confuse them for the real thing.¶

Wilson!¶

Neural Networks¶

are concerned with the problem of learning a function from one space to another space, using data.

The Neuron¶

• The building block is the neuron.
• A very simple computational element.
• Computes a very simple function of its inputs
• Activations and weights summed
• Activation function

A Neural Network¶

Learning Algorithm¶

• While not done
  • pick a random training case (x -> y)
  • run neural network on x
  • modify connections to make reduce error
    • derivative of error function points to direction of change
    • step size indicates magnitude of change