• Question: Beyond reinforcement, what other kinds of machine learning are common?

    Asked by Luka on 31 Dec 2019.
    • Photo: Lewis Griffin

      Lewis Griffin answered on 31 Dec 2019: last edited 31 Dec 2019 3:56 pm

      Reinforcement learning says (roughly) do more of what works, and less of what doesn’t. Many neural networks are instead trained using back-propagation. Consider, for example, a network that attempts to decide whether or not an image contains a dog; when that network gives the wrong answer for a training image (i.e. it says that there is a dog present when it is not, or vice versa), back-propagation works out how to slightly adjust the parameters of the network so that it gives a slightly better answer for that training image. When this procedure is repeated for many many training images, many times, eventually the network will give the right answer for all of its training images, and (hopefully!) will also give the right answer for images that it has never seen before.

    • Photo: Giuseppe Cotugno

      Giuseppe Cotugno answered on 2 Jan 2020:

      I like to think of machine learning as a box of recipes which tell you how you can cook your data in order to highlight a patter or some feature. As such, as there isn’t a single common recipe, there isn’t a single common machine learning approach in industry, however there are favourites.

      Those people who work with images (computer vision) like to use deep neural networks and all the possible variants (recursive, convolutional, encorders/decoders etc.) because it is proven that those techniques work very well. Once branch of reinforcement learning (deep reinforcement learning) is a mix between neural networks and reinforcement.

      A classic technique is logistic regression, which basically is fitting curves into points. This is an evergreen and it is still used for a number of tasks, especially those which require to guess how data is evolving over time (e.g. prices of share, prices of houses, future population, etc.).

      Deep neural networks are probably very popular now as they generally give good results, but there are several other techniques (support vector machines, boosting to name a few) that occasionally pop-up on research papers.

    • Photo: Maja Popovic

      Maja Popovic answered on 6 Jan 2020:

      Deep neural networks have become widely used for language processing/computational linguistics in the recent years. (important note: they are called “deep” only because they have many hidden layers between the input and the output layer, not because they are capable of deep understanding of any language).

      Other machine learning methods are Support Vector Machines, Naive Bayes, Random Forest, Maximum Entropy, Conditional Random Fields.