Machine learning, data science and artificial intelligence

We will compare machine learning to a classic example from statistics: computing a regression line to identify a trend in a scatter plot.

To illustrate, we take some data about marketing budgets and sales figures in the corresponding period:

"Regular statistics" enables, among other things:

→ we see that sales are correlated with marketing spendings. It is likely that more marketing spending causes more sales.

→ by tracing the line further (using the formal model), we can predict the effect of more marketing spending

"Regular statistics" is advanced by scientists who:

→ a key approach is inference: by defining a sample of the data of just the correct size, we can reach conclusions which are valid for the entire dataset.

→ they neglect how hard it can be to run their models on computers, in terms of taking advantage of new types of hardward (ex: GPUs, see below).

→ since they focus on sampling the data, they are not concerned with handling entire datasets with related IT issues.

Machine learning does similar things to statistics, but in a slightly different way:

Machine learning is advanced by scientists who are typically:

(footnote: so machine learning, in my opinion, shares the spirit of "getting things done" as was operations research in the early days)

The pursuit of improved models in traditional statistics is not immune to the notion of computational efficiency - it does count as a desirable property - but in machine learning this is largely a pre-requisite.

A key illustration of the difference between statistics and machine learning can be provided with the use of graphic cards.

Graphic cards (also called GPUs: graphics processing units) are these electronic boards full of chips found inside a computer, which are used for the display of images and videos on computer screens:

In the 1990s, video gaming developed a lot from arcades to desktop computers. Game developers created computer games showing more and more complex scenes and animations. (see an evolution of graphics, and advanced graphics games in 2017).

These video games need powerful video cards (aka GPUs) to render complex scenes in full details - with calculations on light effects and animations made in real time.

This pushed for the development of ever more powerful GPUs. Their characteristics is that they can compute simple operations to change pixel colors, for each of the millions of pixels of the screen in parallel, so that the next frame of the picture can be rendered in milliseconds.

Millions of simple operations run in parallel for the price of a GPU (a couple of hundreds of dollars), not the price of dozens of computers running in parallel (can be dozens of thousands of dollars)? This is interesting for computations on big data!

If a statistical problem for prediction can be broken down into simple operations which can be run on a GPU, then a large dataset can be analyzed in seconds or minutes on a laptop, instead of cluster of computers.

To illustrate the difference in speed between a mathematical operation run without / with a GPU:

The issue is: to use a GPU for calculations, you need to conceptualize the problem at hand as one that can be:

→ then, the calculations will be able to be done on a GPU, which can accelerate the treatment by 10x, 100x or more.

Machine learning typically pays attention to these dimensions of the problem right from the design phase of models and techniques, where classic statistics would typically not consider the issue, or only downstream: not at the design phase but at the implementation phase, which is too late.

Now that we have seen how statistics and machine learning differ in their approach, we still need to understand how does machine learning get good results, if it does not rely on modelling / sampling the data like statistics does?

Machine learning can be categorized in 3 families of tricks:

Unsupervised learning designates the methods which take a fresh dataset and find interesting patterns in it, without inferring from previous, similar datasets.

How does supervised learning work? Let’s take an example. In a wedding reception, how to sit people with similar interests at the same tables?

The set up:

A possible solution using an unsupervised approach:

And repeat for all tables, many times, until no exchange of sits improves the similarity. When this stage is achieved, we say the model has "converged" to one of the best possible solutions.

This approach makes it possible to identify groups of people who have common points. It is obviously very useful to organize the world around us in business, from a segmentation of customers or prospects, to a classification of products in categories for evaluation or portfolio management purposes. There is a very large field of scientific research devoted to designing better clustering techniques suiting a variety of situations. One of the most popular of these techniques remains the"k-means", and was invented in the 1950s:

Supervised learning is the approach consisting in calibrating a model based on the history of past experiences, in order to predict a new occurrence of the same experience with great accuracy. Let’s take an exmple.

Imagine we collected from Instagram 50,000 images of cats and dogs.

etc…​.

The challenge for the computer / software is this one: if we get a new image of a cat without a caption, will it be able to guess the label "cat"?

Supervised learning proceeds this way:

Now, when you get new pictures of cats and dogs (the fresh set), applying the trained model should output a correct prediction (label "cat" or label "dog").

Supervised learning is currently the most popular family of machine learning and obtains excellent results especially in image recognition, even though some cases remain hard to crack:

(source)

It is called supervised learning because the learning is very much constrained / supervised by the intensive training performed:

→ there is limited or no "unsupervised discovery" of novelty.

Important take aways on the supervised approach:

To understand reinforcement learning in an intuitive sense, we can think of how animals can learn quickly by ignoring undesirable behavior and rewarding desirable behavior.

It’s easy and only takes a few seconds. The following video shows B. F. Skinner, a central figure in behavioral psychology in the 1950s-1970s, who teaches a pigeon how to turn around. For this, Skinner proceeds simply by rewarding the pigeon with seeds, as soon as the pigeon makes rotational movements. At the end, the pigeon finally made a complete turn on himself, because he learned that it would give him a reward.

Footnote: how does this apply to learning in humans? On the topic of learning and decision making, I warmly recommend Foundations of Neuroeconomic Analysis by Paul Glimcher, professor of neuroscience, psychology and economics at NYU:

Besides pigeons, reinforcement learning can be applied to any kind of "expert agents".

Take the case of a video game like Super Mario Bros:

Structure of the game / the task:

Reinforcement learning works by:

After looping from 1. to 4. thousands of times, Mario memorizes all the actions to do to complete the game, without any human player:

Reinforcement learning is perceived as corresponding to an important side of human learning / human intelligence (goal oriented, "trial and error").

Now, let’s imagine that we create a game in which two learning machines are competing: one that controls Mario Bros, the other who controls an enemy character in the game, and tries to defeat Mario Bros. By having them battle thousands of rounds and adapt their behavior by learning from their mistakes, the two will learn much faster and improve a lot more than by reinforcement learning alone. This type of artificial intelligence is called "generative antagonistic networks" and many observers see a bright future for it in AI.

Using machine learning can be a waste of resource, when well known statistics could be easily applied.

Hints that "classic" statistical modeling (maybe as simple as a linear regression) should be enough:

Finally, there is a situation in which machine learning is absolutely not the right solution. If the question is that of the role of this or that factor in the determination of a result, machine learning remains silent on this subject. Let’s take the example of pictures of cats and dogs:

Weak AI designates computer programs capable of surpassing humans in complex tasks on a narrow and precisely delimited domain (such as playing chess). Weak AI works through expert systems or with machine learning techniques as seen above. The AI applications we see all around us are weak AI: driving aids and autonomous vehicles, chatbots, computers capable of beating humans at go or Mario Bros, …​

Strong AI is an intelligence that would be able to solve problems of general scope, able to set its own goals, to be aware of itself, or to solve problems of varied and original natures. Today, nothing approaches this and the consensus says that current machine learning techniques are not adapted to create this type of intelligence.

This means that talks on current applications of AI mean we actually discuss weak AI, which is largely synonymous with machine learning.

This presentation is by Ian Hogarth and Nathan Benaich.