ML BuzzWords Guide

A single layer neural network is capable of representing any function

Network Types

Most basic is perception, feed forward neural network

GANS(Generative Adversarial Networks)

Generator creates images and Discriminator decides whether the images are real or fake.

CCN(Convolutional Neural Net)

Use a convolutional matrix/filter on images instead of using raw pixels as input

RNN

Recurrent Neural Networks are good for time series data that has a concept of memory, introduces loop back from neurons to make training more efficient

LSM

Long term short term memory, a neuron has a forget and memory type thing

Auto-encoders

Like Principal component analysis (PCA) in reduce dimensions of data, but neural net trying to learn identity function in small amount of space and for general objects

Training Methods

Grid search, monte carlo, gradient descent, adadelta(change gradient according to sample)

Regularization penalizes complexity

Batch Gradient Descent

Batch of 1(stochastic gradient) doesn't use vectorization well and so can be slower than larger batches

Batch of m is too slow for large m. If less than say 2k, just do batch.

Else your minibatch should fit in Cpu/gpu mem and often good as a power of 2 like 128, 256, or 512.

Activation Function

Adds non linearity which allows neural networks to get any math function

Softmax- for multiple categories

Rlu- max(0, z) less complexity and helps neural nets train faster

Leaky Rlu - prevents dead neurons by having less than 0 be less than 0