Amazon’s Chilean explores the “black boxes” of artificial intelligence algorithms

Chilean scientist René Vidal, current academic at Johns Hopkins University (JHU), professor at the University of Pennsylvania and researcher at the multinational Amazon, will be the keynote speaker at the conference organized by the National Center for Artificial Intelligence (Cenia) to commemorate one year since his creation.

The event will be held this Thursday, January 12, at 9:00, in the Aula Magna of the Santiago Campus of the Federico Santa María Technical University (Vicuña Mackenna 3939, Metro Camino Agrícola), and will address the progress and challenges of the discipline in the country.

Originally from the municipality of Lautaro, in the Araucanía region, Vidal has worked in the United States for more than 20 years. One of his latest projects is to explore ways to explain the decisions made by artificial intelligence algorithms.

This with the aim of understanding why these sets of rules or sequences of steps, typically related to mathematics or computer science, make their decisions. According to the Chilean specialist who lives in the North American country, it is a fundamental understanding for future developments in this field.

These include autonomous vehicles or the diagnosis of diseases through medical images. Furthermore, in recent years, the researcher has led development teams that try to detect diseases with a single drop of blood, using artificial intelligence tools.


“Over the past decade, the use of artificial intelligence techniques has been a real revolution in science and engineering, but also in many of the things that impact us as human beings,” said Vidal, who is was recently appointed director of the IDEAS initiative at the University of Pennsylvania.

“Artificial intelligence is being used to develop self-driving cars, with TESLA, or will be used more and more on a daily basis to count how many people are on the subway, to shop at the supermarket without going through the checkout or to ask more and more complex questions to voice assistants such as Siri or Alexa”, added the main speaker of the Cenia seminar on this discipline.

The National Center for Artificial Intelligence, an entity funded by the National Agency for Research and Development (ANID) under its Core Centers Program, aims to make Chile the pillar of AI development in Latin America, promoting the technological harmony of development with the human being and its environment in order to improve the quality of life of our societies and individuals.

For Chile, added Vidal, it is essential to strengthen R&D investments to exploit the potential of these technologies, as well as promote initiatives that connect different research groups with the aim of obtaining greater impacts.

“The AI ​​area has great application potential, but collaboration is needed. Otherwise, we will develop a series of super cool algorithms, without any impact. Being able to create techniques for many practical applications requires a lot of collaboration. These projects are the way to transform AI developments into concrete applications. What the State of Chile is doing today with CENIA is something very important and hopefully will multiply”.

black boxes

Vidal’s latest research aims to understand why the algorithms that make Artificial Intelligence solutions work make decisions. As they are increasingly used to solve more complex tasks, it is crucial to determine if they work well, how accurately, if they respect privacy, if they are robust or if it is possible to know why you did one thing and not another.

“This can become very important if you have a self-driving car and it has crashed. Who is responsible? If the car went wrong or if it is forced to crash into a person or a building, which one you choose at that moment. If you have to sue someone, who pays and why? Being able to explain why an algorithm takes one decision and not another will be a very important issue in the future”, reflects the scientist.

However, the fundamental problem is that all current algorithms are like a real black box or a big mystery for the community of scientists behind AI platforms. Basically, explains the expert, there is an enormous amount of data that allows you to train the algorithms, which after this process learn to make decisions.

“But if they’re wrong, we don’t know why that happened. We don’t even know why he made the right decision. All the research we’re doing (the Chilean leads a lab of 20 scientists at JHU) has to do with how to make the algorithms responsible for making predictions also capable of explaining why they made that decision.”

New method

The method to explain the decisions of an artificial intelligence algorithm developed by Vidal and his collaborators is based on making the algorithm answer a series of questions understandable to human beings. The sequence of questions and answers constitutes the explanation, which depends on the type of prediction desired and the data available.

The logic, says the Amazon specialist, is similar to that of the traditional children’s game of 20 questions, typical of the Anglo-Saxon world. The child is told: “I have a number in my head between one and ten.” The most logical answer is five, to split the search in half.

René Vidal presents two examples: one, that of a group of birds with the task of identifying which type of bird an image corresponds to. Today, he points out, there are databases of images of birds tagged with more than 300 attributes (such as breast color, wing shape), which can be used to learn how to predict the attributes present in an image.

This way, the algorithm can establish a path of questions to predict bird type faster from very few attributes.

Something similar happens with medical diagnoses: the specialist asks a series of questions to reach a conclusion. By getting that answer, you also define a logical sequence of questions that were asked to achieve this goal. In the case of the bird, because the color of the legs or the elongation of the head is one, or in the case of health care, because the symptoms are one and not the other.

“Normally you don’t need to ask a thousand questions, you want simple explanations. The question is how do you develop an algorithm that can predict someone’s illness or what happened this time by asking a sequence of questions and theoretically asking the minimum number of questions?

This understanding is essential to explore new potentials of technology.

The Chilean researcher describes two of the challenges that leading companies in this field have to face: in Amazon, for example, one of the trends is that its voice assistant, Alexa, asks increasingly complex and sophisticated questions, with which its privacy will simultaneously be more demanding. For TESLA, meanwhile, the ability of algorithms to explain why they decide to speed up or turn right is much more critical.

Data in a drop of blood

For the past 20 years, René Vidal has been a professor in the Department of Biomedical Engineering and Director of the Mathematical Institute for Data Science (MINDS) at Johns Hopkins University, one of the world’s leading centers for the use of technologies applied to health.

Starting January 1, the Chilean scientist, also an Amazon scholar, will begin a new role as the leader of the Innovation in Data Engineering and Science (IDEAS) initiative at the University of Pennsylvania.

The developments in which the Chilean scientist collaborated had various applications. One of the most particular is the development of a new blood test that he created with the Belgian biotechnology company miDiagnostics, the Imec research center, specialized in nanoelectronic instruments, and the R&D area of ​​Johns Hopkins, which then allowed the design of a rapid test for covid-19 that was used at Brussels airport.

The initiative involves the use of a silicon chip that requires only a small volume of samples, which are analyzed through the information provided by the patient’s own fluids. Results with laboratory-quality sensitivity and specificity are displayed on a connected device in an easy-to-use user interface.

JHU specialist – whose work in healthcare focuses on the development of diagnostic imaging techniques – led the team that, before the pandemic, worked on the design of this platform with the challenge of using blood drops for diagnosis of dozens of diseases in minutes.

“The test we were developing is based on images. It’s a technique called holography, the idea is that now you have a video with the blood moving, and the question is if you can detect each of the cells and produce the count of how many are red blood cells, white blood cells and inside the cells blood cells white can be classified into monocytes, lymphocytes and granulocytes, and then produce the test result. All at the same speed as a covid test ”.

The solution, based on microfluidic imaging, had a similar aim to that of the controversial Theranos company, led by Elizabeth Holmes, a biotech startup dedicated to simplifying laboratory tests, which promised to detect hundreds of diseases, including cancer, with a drop of blood through state of the art machines.

Could a Chilean scientist solve the problem that bankrupted Holmes, who before the age of 30 amassed a fortune of more than $4 billion and raised funds from the most reputable investors in the United States?

“From the point of view of the goal and what was expected to achieve, it is similar, but not from the practical application and approach. In our case, through artificial intelligence algorithms. And also, of course, in the ethical part behind each project,” he replies.

To learn more about what is happening in the world of science and culture, join our Cultívate community, El Mostrador newsletter on these topics. sign up for free HERE

Follow us on

The Google News office

Leave a Comment

Your email address will not be published. Required fields are marked *