lecture 4 + 5 Flashcards
What is supervised learning?
The agent learns a function from inputs, this is done using an algorithm which receives a set of sample inputs and their associated desired outputs.
What is unsupervised learning?
The learning algorithm receives a set of training data and has to work out what regularities it contains.
What is reinforcement learning?
The agent receives data and generates actions in response, also receiving a reinforcement signal, either negative or positive depending on its actions.
What is induction?
Assume a true unknown function which always returns the correct output for a given input. Indirect evidence is a pair, the learning procedure will take a set of examples and try and get it as close to expected as possible, ideally it will generalise well to new examples.
What is a consistent hypothesis? Which should we prefer?
One which agrees with all the training examples. We should always prefer the simplest example.
What does it mean when a model is overtrained?
This is when a consistent model does not perform well on test data, it essentially overfits the data.
Why do we typically split up training and testing data?
To ensure that the model can not overfit the data.
What is a simple model of a neuron?
The McCulloch-Pitts unit, it has a bias weight and input links (each of which has a weight) going into an input function, if (the input values*their weight)- the bias is greater than 0 the neuron fires.
What does changing the weight vector and bias do to a 2-D problem using neurons?
Bias shifts the line along one direction while the weights change the rotation.
What do neuron networks do to problems with M-D input?
Define a hyperplane, splitting the problem space into two half spaces.
What is a binary classification function?
One which returns a class label from the set of two labels, 0 and 1.
What is the perceptron learning rule?
- perceptron stimulated by environment
- perceptron parameters undergo changes as a result of stimulation. Weight changes base on error*desired output. Bias changes based on -error.
- perceptron responds in new way
What are the two ways of using a learning rule?
Online training: change weights and bias after each training example.
Batch training: Remember weight changes for each output for full set of training examples and apply average weight change.
What is a classification function?
Returns a class label from a set of possible labels.
