Artificial intelligence (AI)

Any human-made system mimicking animal intelligence. This is a large and very diverse field.

Machine learning (ML)

A subfield of AI that can be defined as computer programs whose performance at a task improves with experience.

Deep learning (DL)

A subfield of ML using artificial neural networks with two or more hidden layers.

ML allows to achieve previously impossible tasks

In typical computing, a programmer writes code that gives a computer detailed instructions of what to do.

Coding all the possible ways—pixel by pixel—that an image can represent, say, a dog is an impossibly large task: there are many breeds of dogs, the image can be a picture, a blurred picture, a drawing, a cartoon, the dog can be in all sorts of positions, wearing clothes, etc.

There just aren’t enough resources to make the traditional programming approach able to create a computer program that can identify a dog in images.

By feeding a very large number of dog images to a neural network however, we can train that network to recognize dogs in images that it has never seen (without explicitly programming how it does this!).

Old concept … new computing power

The concept is everything but new: Arthur Samuel came up with it in 1949 and built a self-learning Checkers-playing program in 1959.

Supervised learning

• Regression is a form of supervised learning with continuous outputs
• Classification is supervised learning with discrete outputs

Supervised learning uses training data in the form of example input/output pairs.

Goal

Find the relationship between inputs and outputs.

Unsupervised learning

Clustering, social network analysis, market segmentation, PCA … are all forms of unsupervised learning.

Unsupervised learning uses unlabelled data.

Goal

Find structure within the data.

Reinforcement learning

The algorithm explores by performing random actions and these actions are rewarded or punished (bonus points or penalties).

This is how algorithms learn to play games.

Decide on an architecture

The architecture won’t change during training.

The type of architecture you choose (e.g. CNN, Transformer) depends on the type of data you have (e.g. vision, textual). The depth and breadth of your network depend on the amount of data and computing resource you have.

Set some initial parameters

You can initialize them randomly or get much better ones through transfer learning.

While the parameters are also part of the model, those will change during training.

Get some labelled data

When we say that we need a lot of data for machine learning, we mean “lots of labelled data” as this is what gets used for training models.

Make sure to keep some data for testing

Those data won’t be used for training the model. Often people keep around 20% of their data for testing.

Pass data and parameters through the architecture

The train data are the inputs and the process of calculating the outputs is the forward pass.

The outputs of the model are predictions

Compare those predictions to the train labels

Since our data was labelled, we know what the true outputs are.

Calculate train loss

The deviation of our predictions from the true outputs gives us a measure of training loss.

Adjust parameters

The parameters get automatically adjusted to reduce the training loss through the mechanism of backpropagation. This is the actual training part.

This process is repeated many times. Training models is pretty much a giant for loop.

From model to program

Remember that the model architecture is fixed, but that the parameters change at each iteration of the training process.

While the labelled data are key to training, what we are really interested in is the combination of architecture + final parameters.

When the training is over, the parameters become fixed. Which means that our model now behaves like a classic program.

Evaluate the model

We can now use the testing set (which was never used to train the model) to evaluate our model: if we pass the test inputs through our program, we get some predictions that we can compare to the test labels (which are the true outputs).

This gives us the test loss: a measure of how well our model performs.

Use the model

Now that we have a program, we can use it on unlabelled inputs to get what people ultimately want: unknown outputs.

This is when we put our model to actual use to solve some problem.

While biological neurons are connected in extremely intricate patterns, artificial ones follow a layered structure. Another difference in complexity is in the number of units: the human brain has 65–90 billion neurons. ANN have much fewer units.

The information in biological neurons is an all-or-nothing electrochemical pulse or action potential. Greater stimuli don’t produce stronger signals but increase firing frequency. In contrast, artificial neurons pass the computation of their inputs through an activation function and the output can take any of the values possible with that function.

Which activation function to use depends on the type of problem and the available computing budget. Some early functions have fallen out of use while new ones have emerged (e.g. sigmoid got replaced by ReLU which is easier to train).

Central to both systems is the concept of learning.

Fully connected neural networks

Convolutional neural networks

Convolutional neural networks (CNN) are used for spatially structured data (e.g. images).

Images have huge input sizes and would require a very large number of neurons in a fully connected neural net. In convolutional layers, neurons receive input from a subarea (called local receptive field) of the previous layer. This greatly reduces the number of parameters. Optionally, pooling (combining the outputs of neurons in a subarea) reduces the data dimensions.

Recurrent neural networks

Recurrent neural networks (RNN) such as Long Short-Term Memory (LSTM) are used for chain structured data (e.g. text).

They are not feedforward networks (i.e. networks for which the information moves only in the forward direction without any loop).

Transformers

A combination of two RNNs (the encoder and the decoder) is used in sequence to sequence models for translation or picture captioning.

In 2014 the concept of attention (giving added weight to important words) was developed, greatly improving the ability of such models to process a lot of data.

The problem with recurrence is that it is not easily to parallelize (and thus to run fast on GPUs).

In 2017, a new model—the transformer—was proposed: by using only attention mechanisms and no recurrence, the transformer achieves better results in an easily parallelizable fashion.

With the addition of transfer learning, powerful transformers emerged in the field of NLP (e.g. Bidirectional Encoder Representations from Transformers (BERT) from Google and Generative Pre-trained Transformer-3 (GPT-3) from OpenAI).