IMAGE: Ahmed Gad — Pixabay (CC0)

Machine learning: scratching the surface… and imagining a few possibilities

One of the most complex tasks for a teacher of innovation these days is getting their students to imagine a particular technology’s potential and then get them to apply them to the future so as to understand the specific innovations they might bring about. Arguably the most difficult development, conceptually and for its infinite possibilities and applications, is machine learning.

Curiously enough, one of the most difficult tasks for any manager in relation to machine learning projects is defining objectives, because in all likelihood you will not have much idea of how to use it. With this in mind, I thought I’d outline some recent practical applications of machine learning, either as an inspiration, or to better understand some of the changes to our lives we can expect over the next few years.

What’s needed for a machine learning project? Fundamentally, data. I could write chapter and verse about the characteristics of such data, their structure and preparation, which usually takes up the most time, but we’ll will leave that for a more detailed entry at some point in the future. Today I’m going to talk about practical examples:

  • Eliminating material on social networks that misinforms and spreads hate speech: an ongoing Facebook priority project is trying to locate posts inciting violence towards Myanmar’s Rohingya community. This is a very difficult task given the versatility of human language and the infinite ways violence can be incited, but made easier by the thousands of examples it already has. The company has hired 60 people tasking them with labelling entries. It intends to hire 100 by the end of the year, all of whom are experts in some of the 100 or so languages ​​spoken in Myanmar in addition to the official language, which is spoken by two thirds of the population. The information will then be fed into an algorithm. Mark Zuckerberg says the algorithm will take between five to 10 years to deliver fully reliable results, but when it does, it will be very useful for the social network, whose role in inciting violence in Myanmar against the Muslim minority there has been questioned. Facebook estimates the algorithm is able to detect 52% of posts potentially inciting violence before users report them.
  • Again on Facebook, another application intrinsic to its activity: replacing closed eyes with open ones in photographs. This is a common problem with group photographs, and if anybody is going to have plenty of examples, it’s going to be Facebook. Its algorithm will locate more photographs of the individual in question, taking into account the angle, lighting and other characteristics and insert their eyes in the original photograph. This is a task for a Generative Adversarial Network (GAN), two neural networks that compete with each other, one generating images, and another judging their level of realism, using the photographs of the person with their eyes open.
  • Another Facebook application, although not related to its social network activities: cutting the time it takes to carry out a magnetic resonance to a few minutes, thus reducing stress and discomfort for patients and increasing the number of tests. Facebook intends to release this project as open source, hoping to use it to improve algorithm training. The key here is acquiring the right data. The company has teamed up NYU, whose clinic has a file of 10,000 anonymous magnetic resonances, so that through deep learning it can train an algorithm to recognize bones, muscles, ligaments and other components of the human body. Once this model is built, the machine will be able to spot anomalies in magnetic resonances.
  • Google and data center cooling: up to a few years ago, cooling was one of the main costs of data centers. In the most modern, this has been reduced to 10%, but even so, there are still significant potential savings to be made. The challenge for data centers is getting the temperature and humidity right, either through the right combination of air or water, ventilation, aeration or air conditioning. To do this, it has used reinforcement learning, an algorithm that allows it to see what a software agent should choose in a given environment in order to maximize a certain function, in this case, spending on electricity. There is a level of human supervision that allows it to see if the algorithm’s choices are excessively risky, but at this moment, the algorithm is already obtaining savings of around 40%.
  • Google again: this time using a medical imaging application similar to Facebook’s that is able to examine images of the eye obtained by optical coherence tomography (OCT). The case is not similar to the previous one about MRI: OCTs are not complicated and relatively fast to perform, but the problem is that, until now, they had to be analyzed by a doctor, which took time and was subject to human error. The company has set up a team with Moorfields Eye Hospital that analyzes around 1,000 images of this type every day in a two-stage process: segmentation through a neural network that converts the OCT into a map of three-dimensional tissues with clearly defined color segments that can detect up to 50 possible diseases, trained with 877 images segmented manually by ophthalmologists, followed by a second classification network that analyzes the three-dimensional tissue map and makes diagnostic decisions about the type of disease and how urgent treatment is. This second network was trained through 14,884 tissue maps produced by the segmentation network that were reviewed by ophthalmologists and optometrists. The resulting process is already able to improve the efficiency of medical diagnosis teams.
  • Detecting cancer: medical imaging already able to recognize skin cancers (95% success rate against 87% by a team of 28 dermatologists), prostate, head and neck, colon and rectal and breast tumors.
  • Preventing natural disasters: Microsoft has developed a tool capable of using high definition aerial images and then algorithmically recognizing canals, rivers, trees, fields, roads and buildings, which combined with the readings of meteorological sensors can predict, plan and prepare for possible floods. As the algorithms are fed with more and more such maps, they learn to recognize more objects, such as different species of trees buildings, etc. — and its application is more reliable.
  • My personal favorite, for obvious reasons: an algorithm capable of grading essay-type exams, which works in Chinese and English and is able to develop a knowledge base to interpret the general logic and meaning of an essay, and that in tests in 60,000 academic institutions with a total of 120 million students, agreed with the assessments of teachers 92% of the time. What’s more, it improves as it grades more exams. Imagine what I could do with my time if I did not have to spend it grading exams, without doubt the task that takes up most of a teacher’s time and the one we hate the most.

These are just a few ideas that featured in news stories I collected since May. A small sample, but one that covers a range of areas and gives us food for thought about the infinite possibilities of a technology we have barely scratched the surface of and that will change the world as we know it; in fact, it’s changing it already.


(En español, aquí)