Machine Learning

We may have realized it's easier to build a brain than to understand one

if you know who said this let me know please

The process of making computers learn to solve problems themselves, based on examples of the problem.

Suppose we arrange for some automatic means of testing the effectiveness of any current weight assignment in terms of actual performance and provide a mechanism for altering the weight assignment so as to maximize the performance. We need not go into the details of such a procedure to see that it could be made entirely automatic and to see that a machine so programmed would "learn" from its experience.

Arther Samuel, Artificial Intelligence: A Frontier of Automation

Important concepts

• "weight assignment"
  • weights are variables, assignments are the values
  • weights are sometimes also referred to as parameters
• weight assignment has "actual performance"
  • the model's ability achieve its goal
  • this is distinct from the "results", meaning the output e.g. results are moves taken vs actual performance is winning the game
• "automatic means" of testing performance
• "mechanism" for improving performance by changing weight assignments
  • can compare two versions performance and move weights closer to the better performing model
• results != performances
  • results are the outputs
    • results are also called predictions
  • performance is the quality of outputs
    • measure of performance is called loss
• classification model is one that predicts a discrete category
• regression model is one that predicts >=1 numeric qualities
  • do not use regression to refer to linear regression models
• "overfitting" can occur when a model is trained for too long, and it starts to memorize the input data too precisely, rather than determining generalised rules
• validation set is the subset of all labelled data that is seperated from the training set, used for measuring the accuracy of the model
• model architecture is the functional form of a model
• metric is a function that measures the quality of a models predictions using the validation set
  • a simple example is the error rate of a classifier
• metric != loss
  • metric is designed for human consumption and review of a model
    • ideally this is as close to the end task as possible
  • loss is a measure that allows a training system to update its model weights automatically
  • they may be the same, they may not be
• "epochs", which are the number of times the training system runs on each label-value pair in the dataset
• transfer learning
  • when working with a pretrained model and using it for a different task to what it was originally trained for
• fine tuning is a transfer learning technique of additionally training a pretrained model with further data

Definitions from Practical Deep Learning Glossary, ch. 1:

The training loop:

Interesting Links

• The Ghosts in the Data
• Knowing Machines