Skip to content

Machine Learning

Machine Learning aims to use patterns in data to make estimates. There are four main components to the learning process:

  • Data: specifically features (inputs) and labels (expected output)
  • Model: makes estimates from features.
  • Objective: what the model is trying to achieve
  • Optimiser: used to improve the performance of the model

Models consist of logic and parameters that are used to transform the input into an estimate. Models are chosen based on their logic rather than the parameter values. Parameters are discovered during the training process. Training is achieved using a cost (objective) function and an optimiser function. The cost function is used to calculate how well a model performed. The optimiser function is used to change the models parameters to improve the cost performance. A gradient-descent optimiser tracks the decrease in cost as parameter values are changed and identifying when the cost has been brought to it's lowest value.

Crucial to the optimiser, is the step-size: how much a parameter value is adjusted for the next iteration. If the step-size is too small, the optimiser can settle for a local minimal cost. If the step-size is too large, the minimal cost can be missed entirely. Finding the optimal step-size requires experimentation.

Machine Learning training can be either supervised or unsupervised. Supervised learning assesses a model's performance by comparing its estimates to the correct answer (labels). Unsupervised training only uses features. Because unsupervised training does not have access to expected output, it is often used in circumstances where no single correct answer exists.

Once a Model has been trained, it can be saved and deployed.

Model training is dependant on the data used in training. Ideally, this should be representative, free from errors and missing data points. A representative dataset contains sufficient values (samples) to represent the whole population. Errors in measurement values can also skew training results. Where data points are absent, models can be chosen that handle missing data, or the missing data can be substituted with reasonable values, or the records with missing data can be removed.