Thanks to the advancement of computing, artificial intelligence and machine learning are here to stay. What are the current advantages and applications? What is commonly known as “deep learning” or “deep learningmachine learning” refer to a set of techniques, a family of algorithms that teach computers to do a certain task. “It’s a different kind of learning than human beings do.: If I tell a child ‘that’s yellow’ three or four times, he already understands and learns. Instead, the computer will probably need a few hundred thousand examples, it’s a statistical learning”, he explains to the UNQ science news agency, Emmanuel Iarussicomputer engineer, Conicet Researcher at the Artificial Intelligence Laboratory at the Torcuato Di Tella University.
It turns out that if there is a large database, the computer learns to see patterns or trends. There are differences from what is known as classical programming. The scientist explains that when programming, it is necessary to methodically clarify each step so that it ends up performing the necessary task. “In machine learning, this is reversed: the algorithm is given different labels and asked to learn it,” he adds. An example would be tagging photos of cats and dogs: after scanning millions of images, the computer can recognize these tags in other photos it has never seen before. Iarussi clarifies that computationally it is very expensive because you have to repeat it many times.
Machine learning uses some techniques that have been known for a long time but have exploded in the present for two fundamental reasons. On the one hand, there is a large amount of data; The Internet and computers have notably increased this availability, something unthinkable in the 1990s. On the other hand, there is now the hardware to handle all this information, as machine learning algorithms are very demanding.
Algorithms Everywhere: Dumb and Powerful
Currently, machine learning algorithms are already are part of everyday life. An example of this is the recommendation systems of entertainment platforms: the algorithm learns from thousands and thousands of users with certain characteristics until it finally knows which movies or series they prefer and can recommend them to other consumers. This is something relatively simple, more complex tasks can be performed, for example what is known as “false deep”. In this type of task, you predict what a video of a person saying something will look like. Iarussi clarifies, between laughs, that the problem is that “it works very well!”. There are millions of videos of people talking to a camera: if you separate the audio and train the algorithm to say a certain sentence, the result is a very real video of a person saying something that didn’t actually happen. The implications of this are unimaginable, you can put phrases in people’s mouths they’ve never said and that’s highly believable.
Machine learning algorithms are included in what is known as artificial intelligence. For Iarussi, “artificial intelligence is an aspiration, it sounds like an almost human entity and the reality is that we are very far from it. We see amazing things, but they don’t have the ability to do things that we humans do.” And he gives an example: “A neural network that I train for a certain task is usually not good for another. people have the ability to do various jobs.” The Conicet researcher also explains that, sometimes, Algorithms are often “dumb”, because when the data are not presented in a similar way to what was learned, they cannot perform the requested task. “A true artificial intelligence must do many things at the same time with the same structure”, he concludes.
All these apps come in “black boxes” and sometimes get excessive power. Algorithms are sometimes used to rule out job searches. “If two thousand resumes arrive, to make the selection faster, several are discarded through an algorithm that is trained with the data of employees who have already done well within the company. It saves work, but the rejection that occurs is pretty grim. There are characteristics beyond the data that an algorithm cannot make a census and this whole dimension is left out”, he details.
“I like to think of these techniques as a tool that enhances capabilities, not a replacement. These methodologies enablegenerate a revolution similar to the one that photography would have generated at the time”, he reflects.
The biomedical revolution
Molecules generally interact with each other and react thanks to their structure; it is a kind of interconnected game at the submicroscopic level. It is known that the behavior of proteins is determined by their shape and that is why the company “DeepMInd”, owned by Google, developed “Alphafold”. This program is a set of algorithms that makes predictions of the spatial geometry of proteins thanks to machine learning. “It’s a huge protein database that can predict structures very accurately. You can’t write infinitely complex algorithms, but thanks to databases you can learn automated tasks,” says Iarussi.
The scientist gets emotional when sharing some of his research: “We were contacted by experts who needed to quickly count and differentiate between dead and live cells in breast cancer cultures.”. For this, they managed to train an algorithm to make it capable of differentiating microscopy images: “We helped solve this task and managed to give some tools to the specialists”.
Iarussi’s team is also studying the application of geometry generation algorithms. An example of this is the bone CT scan, which reconstructs the geometry of the bones, but with very low resolution. “It’s not enough to generate a 3D model of the bone,” he explains. What is done is training an algorithm using bones scanned at very high resolution. This cannot be done on people as it is a technique that requires a lot of energy and would cause great damage. However, bone scans can be done ex vivo to train the algorithm and compare with low resolution tomography. “Although we are still far from direct application, tools are improving by leaps and bounds”, comments.