Open the Pod Bay Doors, HAL or Perhaps Alexa can do it?
Traditional computer programs consist of a list of instructions, put together by a programmer. It is the programmer's job to find the most concise, logical set of instructions to perform a task.
The AI system in Final Fantasy XII shows a traditional set of instructions.
A Machine Learning project can learn from itself or from data that you provide. Instead of programming logic yourself, you will train a model and then let that model make decisions. These models can often find patterns in complex data that even the smartest programmer could not have discovered.
- Artificial Intelligence: you program a character to drink a potion once his health gets too low.
- Machine Learning: you train a model with previous combat data, and then let it decide what is the best moment to drink a health potion.
In the age of Big Data, it's becoming harder to find patterns and solutions using traditional programming. Machine Learning can analyse huge amounts of data.
- Medical diagnosis
- Financial analysis
- Chatbots
- Language Translation - Japanese translation was not possible before Machine Learning
- Recommendations in shops
- Self-driving cars
- Live video analysis
- Pose recognition (kinect)
- iPhone face ID
- Playing GO
- Email spam detection
- Anne Frank project
This workshop is about supervised learning. We supply the computer with labeled data, and then we build a model that allows us to label new unknown data.
A model is built using an algorithm. When starting a Machine Learning project, you have to look carefully at the data you have available, and the kind of result you need, to determine which algorithm fits that need.
Using an algorithm to train a model to recognise cat drawings. This step is needed only once.
Once you have the model, you can ask to what degree a new drawing resembles a cat. Or you could even let it imagine new cat drawings!
Replace the cat images with credit card transactions, including data that mentions if the transaction was fraudulent. After training a model with this data, you can test new transactions for being fraudulent.
model
Train a model with date, amount, user, location, shop as numbers in an array. The data is labeled as valid or invalid.
model.train(
[12122018, 33, 2282, 3, 55], valid
[10212017, 24, 4343, 6, 45], valid
[12122018, 112, 1231, 6, 12], invalid
)
test a new transaction
var valid = model.test([10212017, 24, 4343, 6, 45])
There are many different algorithms and methods to create a self-learning model, for example:
- Nearest Neighbour
- Neural Networks
This algorithm uses the distance between points to classify a new point. In the above image, the green circle will be classified as red, because there are more red triangles in its vicinity.
The above example will classify the green circle as RED, because there are more red triangles in its vicinity than blue squares. This classification is based on x,y coordinates of a point. But surprisingly, this calculation works just as well when those two numbers do not represent a physical coordinate, but for example, the number of legs and eyes of a creature.
You can imagine that this also works in three coordinates: x,y,z, so we can use the algorithm for datasets with three numbers. The real beauty of this algorithm is that it also works with even more than 3 coordinates.
In other words, we could build the credit card example using Nearest Neighbour, by supplying our credit card data as an array.
This algorithm is much faster than a neural network, so it may be a better fit for simpler projects!
A Neural Network is inspired by the human brain. Input will flow through neurons in the network. Each neuron has weights that decide to which next neuron the data should flow.
By repeating the training (iterations), the weights are adjusted until input results in the desired output.
- Weight The network starts with random weights for each input.
- Output The combined inputs and weights result in one of the two outputs
- Iteration Adjust the weights until the output matches the desired output
Links
- Creating a single neuron in Javascript
- Neural Network basics explained
- But what is a neural network? - youtube
- Neural networks demystified
Instead of writing algorithms yourself, you will often choose an existing library, based on your data and your expected output. Tools exist for:
- Image Recognition
- Speech Recognition
- Language Generation
- Body Pose Recognition
- And many more, see the reading list