Understanding neural networks: what are they and what can we use them for?
By: Xareni Galindo, PhD student; Thesis topic: Deep learning for 3D high- and super-resolution microscopy. Thesis supervisor: Jean-Baptiste Sibarita / Team Quantitative imaging of the cell IINS. https://www.linkedin.com/in/xareni-galindo-b18019b3/
If we have access to tv, cinema, or even science fiction books, we have heard of artificial intelligence and most of the time we may have believed that it was something that was just fictional, only reachable by very advanced civilizations or indeed by humans but in a long time. However, artificial intelligence is here, living among us, but maybe we have been too busy to look for it, and contrary to what we may think, its presence goes beyond humanoid robots or sci-fi movies and books.
Colloquially, we can define AI (Artificial Intelligence) as the ability of machines to imitate (the role of) a human brain, like “learning”, “thinking”, “solving problems” and “recognize objects”. Like humans, machines use neural networks to perform these tasks, we call these Artificial neural networks (ANNs), usually simply called neural networks (NNs), which are inspired in biological neural networks.
ANNs are built using a structure that is loosely modelled on a biological human brain, they use artificial neuron that loosely model the neuron in a biological brain, like synapses in a biological brain, a connection between the artificial neurons, is in charge of transmitting information to the other ones, whereas biological networks use electrical impulses, artificial neural networks use numbers.
Neural networks need an external initial stimulus to begin processing the information contained in them, this initial stimulus can be images, videos, audios, or documents. After the information contained in the stimuli is processed, the task for which the neural network was designed (such as recognizing an object in an image) is completed, we can have our final result. An example of this can be a photo of a dog and a cat as initial stimulus and the same image with a rectangle framing the dog as a result.
Artificial neural networks have evolved into a broad family of techniques across multiple domains, one of the most widely used branches used nowadays is Deep Learning. This technique has been applied to a broad number of subjects, such as speech recognition, medical image analysis, audio recognition, social network filtering, drug design, medical image analysis, self-driving cars, face recognition, finance, and board game programs, to mention a few examples. The use of deep learning, neural networks in these fields has produced results comparable to human expert performance.
A daily-life example of the usage of deep learning is Google Traductor© that uses deep learning algorithms to translate between more than 100 languages. Google Maps© uses deep learning pattern recognition algorithms to show you how similar a restaurant is to your taste or to help us to find a parking lot. Deep learning is also the backbone of websites advertising algorithms, content suggestion algorithms in social networks like Facebook©, Instagram©, and TikTok©. Face-swaps applications and filters are also possible thanks to deep learning.
Nowadays, deep learning is widely used in the medical and biomedical field where its application has begun to show substantial results in developing new treatments, medical diagnosis, and clinical decision-making. The application of computational algorithms to health care and biological studies can bring a new era in the treatment of diseases and drug development. Yet, there still are challenges to deal with, such as the lack of accessible data, the lack of flexibility and multitasking, and the necessity of high-processing machines.
Now, can you say how many times deep learning has helped you in your everyday life? Neither can I, but I am happy that scientists have been able to take this huge step in technology because as a Computer Science Ph.D. student, I know the limits of its applications have not been reached yet.