« Back to home

Introduction to Machine Learning

Week One:

  • Online quiz on learn.gold
  • Wekinator Assignment 1 on learn.gold
  • Supervised Learning Ideas (forum link)

Reflection:

The aim of machine learning (ML) is to process an input in order to reach an output, without actually programming the decision-making stage (or the model in the pipeline) explicitly. My initial thoughts about this generalised ML schematic were that it seems extremely broad. Indeed, the input can be virtually any type of data; from single values to video streams. Furthermore, the many algorithms used to produce an output seem sparsely related; some (e.g. kNN) enable classification of an input from a set of known categories, while others (e.g. neural networks) enable regression using the relationship of inputs and outputs of the system. Possibilities continue with temporal analysis used for speech and gesture recognition, and interoperation of algorithms to fit different goals (e.g. classification using neural networks). The concept of supervised learning brought my slightly overwhelmed understanding together. An extension to the [input -> model -> output] schematic, supervised learning provides a method to build the model by feeding training data (a set of corresponding input and output pairs) to the learning algorithm. After the model training stage, the running model can then analyse new input data in real-time.

With supervised machine learning in mind, a group including myself discussed some potential applications. One targets audio engineering - setting lots of effects that don’t always match the music being played (a high-pass filter probably wouldn’t suit bass-heavy hip-hop, for example). A model could be taught which settings pair with various notes in order to automate the process in real-time. This would benefit from models so complex that they couldn’t be feasibly programmed, due to the complex nature of musical appreciation. However I’d expect it to require a huge training set to be remotely useful. If it were not useful to replace the hard work of an audio engineer, I think the expressivity could be useful to musical artists in other creative ways; the output could be used in accompanying visuals which would change with the sound, or the output could be applied to another instrument to give it a strange new profile.

For Data and Machine Learning for Creative Practice (IS53055A)