A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E.

The process of learning itself is not the task. (Learning is our means of attaining the ability to perform the task.)

types of experience

supervised (data is labelled, or a target is specified),
unsupervised (features are learned from unlabelled data)
reinforced (no data is given, learning happens via interaction with an environment),

types of tasks (not complete):

classification (supervised)
- sort input into a predetermined set of categories
- e.g. object recognition, face recognition
- image classification demo: https://www.clarifai.com/demo
clustering (unsupervised)
- sort input into arbitrary categories
regression
- predict an output given some input
- e.g. predicting prices, drawing a new image based on an original
transcription
- unstructured data —> structured sequence
- e.g. optical character recognition (image to text), speech recognition (audio to text)
translation
- structured sequence —> structured sequence
structured output
- outputs multiple values that are related to each other in some way
- e.g. parsing a sentence into parts of speech, providing a caption for an image
anomaly detection
- picks out unusual examples from an input set
- e.g. fraud detection
synthesis
- similar to “structured output”, but variation in output is valued:
  - generating texture
  - synthesizing speech
suggesting or generating missing items of a set (unsupervised)
- interpellation
- compression
denoising
- predicting a “clean” example from a “corrupted” example

performance measure

(how training happens)

also referred to as the cost function.

back propagation: weights are adjusted moving from output layer back towards input layer
gradient descent: a performance measure for the performance measure, or, which direction does each change in weight push the overall network?
beware of overfitting - the challenge of fitting the training data differs from the challenge of finding patterns that generalize to new data

a few types of artificial neural networks

“feedforward networks”

perceptron

multilayer perceptron

Each connection between neurons carries a trainable Weight and offset Bias. By adjusting these values through a training process, the strength of individual connections are made stronger or weaker.

In a single layer type, Input Neurons are connected directly to Output Neurons.

In a multilayer type, Hidden Layer Neurons each have their own activation function.

“Recurrent Neural Networks” (RNN)

RNN

RNN (Recurrent Neural Network)

LSTM

LSTM (Long / Short Term Memory)

GRU

GRU (Gated Recurrent Units)

used when sequence is important (what came before determines what comes next)

can be generalized to include text, audio, video (anything time-based), attention (scanning a scene, etc.)

RNNgenerate RNNread