Table of contents: 

Project summary:

In this project I will explain how machine learning (ML) works through pieces of code, and then I will describe how AI/ML can be used, and debate its advantages and disadvantages.

Just as we learn through examples and by processing various inputs, machines can do the same. 

The human brain

Cerebrum. Largest part of the brain. Divided into 2 hemispheres which each have 4 lobes…

• Frontal - reasoning, problem-solving, emotional control

• Parietal - manages sensory inputs like touch and temperature

• Temporal - manages auditory information

• Occipital - processes visual information

Cerebellum

• Beneath cerebrum

• 10% of cranial cavity

• More than 50% of the brain's neurons

• Controls balance and complex movements --> requires abundance of wiring

Brainstem

• The oldest part of the brain

• Controls sleeping, breathing and keeping the heart going. 

Also present in the brain are a variety of smaller structures, e.g. the hippocampus and hypothalamus. 

What is AGI

AGI: artificial general intelligence - does anything a human can do

AI: artificial intelligence tuned to a specific purpose - e.g. a smart speaker, cars, web search, AI in industries like healthcare, farming and factories.

Machine learning

Let us imagine a scenario where we are predicting the price of a house, based on information such as the area, number of bedrooms and location. Given data, a machine can build a model to make such predictions. 

Here's what we will do...

• Import packages

• Find some data (kaggle is great for this)

• Visualise the data

• Process the data (e.g. converting 'yes' and 'no' to 1s and 0s for the model to work)

• Fit the data

• Train the model

• Evaluate the model's accuracy, e.g. root mean squared error (RMSE)

Machine learning - how it works:

Linear regression essentially forms a 'line of best fit' - this is used in the code below. (Note that this is not the only form of machine learning, there are others such as decision trees, random forests and classification models such as logistic regression.) 

2 of the above pictures are from machine learning specialisation course by Andrew Ng (Coursera). Here is code that implements the Linear Regression model. 

Neural networks - deep learning:

Neural networks can be programmed using a software package called keras, which is used to make layers containing nodes. Each node has a certain 'activation function' (such as sigmoid or ReLu which is used in the code below) and takes an input, performs a function, and provides an output. 'Forward propagation' consists of reducing numbers of nodes per layer, and the first layer has the same number of nodes as there are 'features' in a dataset. The neural network learns the relationship between the factors (we can do this by visualising the data using matplotlib - see below) and narrows down to 1 output node. In the code below this predicts the price of a house. 

The human brain is a vastly complicated structure. Because we have almost no idea of how the brain actually works, we cannot precisely mimic it. In deep learning, neural networks use ‘nodes’ which are very different to human neurons and synapses, being much simpler.

The difference between human neurones and artificial ones is shown below.

Neural network model - house price prediction

House price prediction - summary, conclusion and some further explanation

We were able to to different models on the same dataset to make predictions for house prices. For this task and quite often neural networks can work better because they are a more complex model and are able to incorporate non-lineartities. However, it is recommended to start from simpler models to avoid overfitting. (See below.)

The optimiser Adam is used in the house prediction code above. The Adam algorithm improve gradient descent, illustrated below...

Activation functions are shown below.