Liquid neural networks are ready to catapult artificial intelligence. Word from MIT
This new type of artificial neural network is more flexible, efficient and has greater expressiveness than other neural networks.
Liquid neural networks. It seems like an exotic idea, and yes, in a way it is. MIT (Massachusetts Institute of Technology) has been working on them together with other highly reputable universities, such as the Austrian Institute of Science and Technology, at least since the beginning of this decade . And it is obtaining very promising results . Understanding what a neural network is is not difficult if we avoid the most complex details. After all, it is a computer program inspired by the structure and functioning of the human brain with the purpose of processing information and inferring new knowledge.
An artificial neural network is made up of several layers of logical objects known as nodes or artificial neurons. Each node has its own entity and is capable of processing information to obtain a result and deliver it to one or more nodes in the next layer of the neural network. There are several different types of neural networks , but they all pursue the same goal: processing information to generate new knowledge. Convolutional neural networks, for example, are used to identify patterns, recognize images, interpret voice, or implement computer vision algorithms.
However, there is a very important component of this technology that we have not yet delved into: training . This procedure simply requires delivering information to the neural network so that it can be processed, but not with the purpose of returning one or more results, but rather so that it learns to work with this type of information and is able to make predictions or classifications. when we later provide you with the data we need to analyze. This description is a bit crude, but it is useful for us to understand what we are talking about without complicating this article too much.
Liquid neural networks are the best allies of the coming artificial intelligence
Unlike other types of neural networks, which, as we have just seen, have been designed to learn during the training phase from a predetermined input, liquid neural networks also learn during the analysis of the information given to them. for the purpose of inferring new knowledge. This simply means that they are capable of continuously adapting to the new input data they receive in order to learn dynamically and uninterruptedly. As you can see, it sounds very good.
The researchers who have designed these artificial neural networks have chosen the adjective ‘liquid’ precisely to suggest their flexibility and ability to constantly adapt. However, their proposal has two more qualities that are worth not overlooking: presumably they are more efficient than other artificial neural networks and have greater expressiveness .
According to Ramin Hasani , who is one of the MIT researchers specialized in liquid neural networks, expressiveness describes the ease with which engineers can act on the performance of the neural network by modifying the representation of the artificial neurons. This characteristic, again according to Hasani, allows these networks to address levels of complexity that are not manageable with other information processing structures.
We still need to explore something very important about these neural networks. The most important thing, actually: its fields of application. MIT engineers have already successfully used them to implement an autonomous navigation algorithm for drones that is capable of adapting in real time to the most complex navigation spaces. However, this application is nothing more than the prelude to what liquid neural networks can theoretically do.
And their designers hope that in the short term they can be used to catapult autonomous driving of cars, develop much more precise medical diagnosis systems, process video in real time or analyze large volumes of financial data, among other applications. If they finally prove to be as flexible and capable as their creators claim, we could be on the verge of a great leap in the field of artificial intelligence .