Machine Learning (ML)

This is a library of deep learning and predictive analytics written from scratch in Elixir.

graph LR;
  AI[Artificial Intelligence]-->ML[Machine Learning];
  ML-->DL[Deep Learning];
  ML-->PA[Predictive Analytics];

Loading

Installation

This educational project is not published on Hex.

Clone the repository and run mix test from its root.

FizzBuzz classifier

ML.Classifiers.FizzBuzz.classify/1 trains a four-class, single-layer neural network with stochastic gradient descent, then predicts an integer's label. It accepts positive and negative integers, including zero, and returns one of "None", "Fizz", "Buzz", or "FizzBuzz".

ML.Classifiers.FizzBuzz.classify(7)   #=> "None"
ML.Classifiers.FizzBuzz.classify(9)   #=> "Fizz"
ML.Classifiers.FizzBuzz.classify(10)  #=> "Buzz"
ML.Classifiers.FizzBuzz.classify(15)  #=> "FizzBuzz"

The three inputs are a bias and two divisibility flags. Training uses the representative numbers 1, 3, 5, and 15. Each call starts from zero weights and updates them over 500 passes through those examples. This keeps the example deterministic and makes training observable through train/0; it also means every classification pays the cost of training.

For each example, the training step computes four scores with dot products, applies softmax to get probabilities, and subtracts a learning-rate-scaled gradient from each weight. The gradient for a weight is (predicted_probability - one_hot_target) * input_feature. ML.Classifiers.FizzBuzz.loss/2 returns the cross-entropy loss for a probability vector and correct class index. Training uses its gradient directly without calculating the scalar loss.

weights = ML.Classifiers.FizzBuzz.train()
probabilities = ML.Classifiers.FizzBuzz.predict_probabilities(3, weights)
ML.Classifiers.FizzBuzz.loss(probabilities, 1) # 1 is the "Fizz" class index

The probability order is "None", "Fizz", "Buzz", "FizzBuzz" (indices 0 through 3). A lower loss means more probability was assigned to the correct class. ML.fizz_buzz/1 is a shortcut for classify/1.

Guides

See the Livebook guides, starting with the FizzBuzz notebook.