FAQs

Deep learning (DL) is a subfield of machine learning (ML), which is in turn a subfield of AI. As in ML, DL algorithms and models also take numerical inputs and either predict a quantity or sort objects into categories. However, they are much more sophisticated, flexible and can detect subtle patterns in their training data, which is why they can be better suited for complex tasks e.g. language translation. That said, they also require a lot more training data to perform well and avoid 'overfitting'.

Words are converted into numbers to make them amenable to machine learning algorithms. One way is to assign each word a unique number based on the order it appears in a dictionary. More sophisticated approaches assign a collection of numbers to each word (somewhat like coordinates) so that similar words will be "near" one another. The process of turning words into numbers is designed to be reversible so that a model's numerical outputs can be turned back into words as necessary. See Principle 4.

ML models function by abstracting data mathematically. The challenge is to translate these abstractions into plain terms and practical insights. The extent to which this can be done will depend on the algorithm used and the insights required (e.g. a model might reveal that it gives 9% more importance to the property postcode than square footage in predicting sale price, but it might not be obvious why the provided data has led to this behaviour in the model).

Alas, it does depend. Does it seem like a regression or classification task, or an NLP use case? Are the expected outputs in a well-defined format like categories (e.g. 'relevant', 'irrelevant')? Instead of an ML algorithm, will a quicker rule-based solution suffice? Can a model afford to make errors on this task? Do we have enough relevant data of good quality? These are some of the questions that should be considered in deciding whether AI or machine learning should be deployed for any task. See also the Data Science module.

Statistical models are trained to produce the most accurate outputs by generalising from their training data. Bias in their predictions therefore often reflect issues and biases in the training data. While some biases are more inherent within the training data and require more sophisticated treatments, certain statistical biases may also arise from technical faults in ensuring data quality. For example, if a model is trained on an unrepresentative data set containing 1 'relevant' document and 99 'irrelevant' ones (relative to the problem statement), it might learn to adopt what seems to be a blanket policy of classifying every new document as 'irrelevant', since this resulted in a high accuracy of 99% on its training data.