The laws of physics in the service of machine learning

Object detection, handwriting recognition, machine translation, and data annotation…Machine learning is revolutionizing modern technology, and is today the cornerstone of artificial intelligence. There are some tools in physics that can help understand these intelligent machines, and it’s a safe bet that both areas of research will grow from studying these connections.

The artificial intelligence revolution

Over the past two decades, artificial intelligence has made tremendoaus progress and these successes are transforming our relationship with computing. Today, we are not surprised by the automatic focus of our camera on the faces, nor the relevance of the automatic responses offered by our e-mail server. What is behind it is a collection of computer programs that allow computers to perform these very complex, so-called intelligent tasks, such as recognizing a face in a video or having a conversation.

Machine learning is about creating algorithms that can learn to perform a task from a set of examples. Say you want to sort pictures of cats and dogs. First, the algorithm is provided with a set of images for which the desired task has been performed beforehand: the result (cat or dog) for each image is known. On these examples, the algorithm trains, to then know how to distinguish the two categories on any new image. What is crucial in this method is that we only give the algorithm the training photos, no additional information on what differentiates a cat from a dog, and that it ultimately succeeds to generalize on new photos. Hence the idea of machine learning.

Let’s go into more detail, how in practice can we train an algorithm? The first step is to define a function with a lot of parameters, whose starting space is the set of cat and dog photos and whose end space is the two-element set {“cat”, ” dog “}. The training then consists in adjusting the parameters of this function such that it associates a cat image with the value “cat”, and respectively a dog image with the value “dog”.

Without knowing it, you may have already used one of the simplest models of machine learning: linear regression. From a set of training points (x, y) ∈R2, you have fitted the parameters (a and b) of an affine function y = f (x) = ax + b. So you were able to predict for a new value of x, the corresponding value of y.

The two examples above, the classification of cats and dogs and linear regression, are illustrations of supervised learning. This type of learning involves learning the relationship between an input and an output. In our examples, one entry corresponds to a photo of an animal or a value of x and an output corresponds to a “cat” or “dog” value, or a value of y. We talk about supervision because we indicate to the algorithm for a given input the correct answer, that is to say the corresponding output value.

Unsupervised learning is a bit more complex. It’s about automatically discovering structures in a database, without having to classify or predict an output value.