Fundamental Issues

• Top-down approach: pre-processed to do things
• Bottom-up approach: simulate brains, learn new behaviours

Travelling Salesman Problem

Combinatorial Explosion Problem: One of the major unsolved theoretical problems in computer science

• Has to visit a number of cities, must start and end at the same one.
• Need to find the minimum distance/cost solution

Towers of Hanoi - Move 64 gold rings to form a tower, need to move the tower

Combinatorial Explosion

- Where the number of problem solutions grows exponentially with its size. An exhaustive search is not an option [Insert image]

The Turing Test

• Work out which candidate is the machine or human, only by asking them questions.
• If the machine can fool the interrogator 30% of the time, the machine is considered intelligent.
• E.g Alexa, Google AI Cloud contact.

Chinese Room Experiment

• Rule processor happens to be intelligent but has no understanding of the rules

Machine Learning: ANN

Machine Learning

• Computer programs that automatically improve their performance through experience.
• Study of computer algorithms that can improve automatically through experience and by the use of data.

Top-down/Classic: Model all different functions and wire all these 'agents' together. Bottom-up/Deep Learning: Give the system a lot of data, so it can discover by itself

Neural Networks

• Signals move between neurons
• Sum of inputs >= threshold, the neuron 'fires'. If theirs a long-term firing pattern, then the AI is learning

Weighted Sum: $Y=x_1\times w_1+x_2\times w_2+x_3\times w_3$ Neurons connected by directed, weighted paths. Positive weight: excitatory, otherwise inhibitory. $x_1$ and $x_2$ encourage the neuron to fire. $x_3$ prevents the neuron from firing

Each neuron has a fixed threshold $t$. Weighted sum input to the neuron $Y >= t$. Then the neuron fires.

Training

Weights are normally randomly assigned

• Epoch: Entire training set feed into the neural network. The AND function: an epoch consists of four sets of inputs(patterns) feed into the network
• Training Value, T: Value that we require the network to produce
• Error, Err: The amount the output by the network O differs from the training value T
• X_i: Inputs to Neuron
• w_i: Weight from input X_i, to the output
• LR (Learning Rate) How quickly the network converges. It is set by the experimentation, typically 0.1. This is the adjustment amount

While epoch produces an error
    Check next inputs(pattern) from epoch
    Err = T - O
    If Err <> 0 then
        w_i = w_i + LR * X_i * Err
    End If
End While

• Mean Squared Error: [(T-O)^2]/n. Measures the different between T and O during training.
• Linearly Separable: Functions which can be separated. Only these can be represented by a single layer NN

• T is the expected output the learning algorithm it trying to learn.
• O is the output of the ANN with the current set of weights during the training.

If you can separate your inputs and outputs with a straight line then you have found the weights.

Plot your inputs on a graph If you can draw a line such that your inputs that lead to an expected output are on one side of the line and all other inputs are on the other side, then congrats

Machine Learning: Intro

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.

Three pillars of machine learning: 1) Models and algorithms. Doesn't require knowledge engineers 2) Powerful and cheaper computation 3) Massive data warehouse

Process

Partition the total datasets into subsets:

• Learning the parameters of the model
• How to generalise the independent data (overfitting)
• Overfitting: When a statistical model fits exactly against its training data. When this happens the algorithm cannot perform accurately against unseen data.
• When doing evaluation do 70% training, 30% testing. Multiple runs using different partitions(K-Fold)

Data Mining

• Data mining: Exploration and analysis of large quantities of data to discover valid, novel, useful and understandable patterns in data.
• Machine learning predicts with models whereas data mining explains patterns

Machine Learning: Techniques

• Supervised Learning: From labelled training data, with examples of inputs and desired outputs (Classifications, regression)
• Unsupervised Leaning: A function to describe hidden structure from unlabelled data. (Clustering, association rules)
• Classification: Learn to predict to which set a instance belongs to based on pre-labelled instances.
• Regression: Estimated relationship between variable Y and variable(s) X. Function is based on the given data to minimise its mean square error to fit the data.
• Decision Tree:
  • Internal Nodes: Decision rules on features (variables/inputs).
  • Leaf Nodes: predicted class label (output)
  • +Quick Training time, handle large features, easy to implement
  • -Only simple decision boundaries, problem with lots of missing data and cannot handle complicated relationship
• Neural Network: Set of neurons connected by directed weighted edges
  • +Can lean complicated class boundaries and be more accurate.
  • -Hard to implement, and slow training time. Can overfit the data and hard to interpret

Supervise Learning Applications

Handwriting/Pattern Recognition/Translation

• Clustering
  • Given: Un-labeled data set D and similarity/distance metric
  • Goal: Find 'natural' partitioning, or groups of similar data points
  • Application: Divide a market into distinct subsets of customers
• Association
  • Correlation between any two or more variables
  • Given: Set of records containing items
  • Goal: Produce dependency rules, to predict occurrence of one variable based on other variables
  • Application: Market basket analysis
• Correlation $\ne$ Causation