From the Basics of Machine Learning to Understanding TensorFlow Internals with Raphael Gontijo Lopes..
Machine Learning
Data has high-level features
Some of those features are called labels
You can extract those features and labels from data using classifiers
Machine Learning is how computers learn to do this (train classifiers)
Supervised Learning is ML where you learn from labeled data
Neural Networks
Data is composed of attributes
Labels can be extracted from those attributes
Some attributes are more relevant for determining a label (so we weight it higher)
Perceptrons uses those weights to compute higher-level features
Layers of perceptrons can approximate any kind of function (including extracting labels from data) and even higher level features that themselves depend on high level features
We find optimal weights using Gradient Descent
(modifying all weights using derivative of a loss function)
Neural Network Abstractions
Whole layers of Neural Networks can be expressed with a matrix multiplication
GPUs are very fast for multiplying matrices
Whole Neural Networks can be represented by a computational graph, which perform operations over those matrices (or tensors)
TensorFlow
Define the computational graph first, then run data through it to train weights
Think about what’s going on under the hood if you run into issues
Use TF’s summary features to gain insight into your training
Practical Advice
Not learning well? Be careful with learning rates!
Divide your dataset into train and test set to see how well your NN generalizes
Get more data (!)
If you can’t get more data, fine tune
