Sem2 Week 4

Constructing a decision boundary using data points where you know the inputs and the true output. Has two parts: a way to represent a decision boundary (hypothesis space) and a way to adjust it to fit data (learning).

Learning from one set of data (training data) and being able to apply the same decision boundary to make decisions about new, never-before-seen data points.

Learning a decision boundary that overly matches the training data but does not generalize well to new data.

The inputs to a supervised learning model. Often, the practitioner must decide what these are.

A type of neural network that can represent a non-linear decision boundary, using hidden layers between input and output.

Hidden layer nodes represent values that are based on, but different from, the inputs. They allow the network to build more complex, non-linear decision boundaries.

A method for doing gradient descent in a neural network: repeatedly computing the derivative of the error function with respect to the weights, starting at the output layer and working backwards through the network.

An optimization algorithm that calculates the gradient (∇F) of a function and takes a step in the direction -∇F to minimize the function. Step size is controlled by the learning rate η: Step = η × -∇F.

A hyperparameter controlling the step size in gradient descent. Each weight update step = η × -∇F.

The step function is not differentiable, so you cannot compute the gradient needed for gradient descent.

A smooth, S-shaped activation function used instead of the step function. It is differentiable everywhere, making gradient descent possible.

The space of all possible weights — not the inputs. Gradient descent adjusts weights to minimize this error. The error surface is a landscape over weight space — every point in that landscape is a different combination of weight values, and the height at that point tells you how wrong the network is with those weights.

Using all training data at every weight update step. Most accurate gradient estimate but computationally expensive with large datasets.

Using a random subset (batch) of data at each update step. Less accurate per step but much faster to compute.

One complete pass through all of the training data during training.

Starting at a random point in the search space, looking at neighbouring points, and moving downhill. Often works with enough random restarts.

They are difficult to train: every weight influences not just its own node but all downstream nodes, so any single weight change ripples through the entire network.

No. ANNs are complex nonlinear functions; ‘learning’ is just setting weights. There are many differences between ANNs and biological neurons despite the inspiration.

A convolutional neural network (CNN) architecture from 2012 with Around 60 million Parameters, split across two GPUs.

Supervised learning (e.g. classifying MRI scans as tumour/no tumour, predicting consumer purchases); Unsupervised learning (e.g. clustering customers by behaviour without predefined labels); Reinforcement learning (e.g. training a game-playing agent by rewarding good moves).

Classification learns a decision boundary that divides data into distinct classes (e.g. spam/not spam). Regression learns a continuous function to predict a numerical output (e.g. predicting house prices).

A decision tree is a non-neural supervised learning method that represents complex logical functions by splitting data on feature values. It is learned using algorithms like ID3. Unlike neural networks, it does not use weights or gradient descent.

1) Identifying venomous snakes from photos. 2) Recognising handwritten or printed letters. 3) Predicting what products a website visitor will buy. 4) Detecting tumours in MRI scans.