The first algorithm for training neural networks was formulated by the Canadian neuropsychologist Donald O. Hebb, who studied neuron interaction and was interested in the principle governing their combination into ensembles. Hebbs ideas were initially criticized, though some years later a group of American scientists were able to model an artificial neural network that could distinguish squares from other geometric shapes.
In 2014, a revolutionary event took place: Deep Artificial Neural Networks were able to recognize objects in pictures with an accuracy of more than 95%, exceeding human capabilities. From this moment, neural networks began their triumphal procession across the globe.
Neural networks of this kind are quite large (composed of several dozen to several hundred layers), and large computational resources are required for training. Every picture, however, consists of approximately the same set of elementary objects: points, sloping lines, fine detail resolutions, and so forth. As a result, one need not begin training the neural network from scratch every time, working instead from the last point.
For example, you can take a database of X-ray images of lung cancer and retrain only the last few layers of an existing neural network. And more importantly, additional training does not require powerful computing resources.
Each day brings with it more examples of the successful application of artificial intelligence in medicine: neural networks can already recognize malignant tumours, including skin neoplasms, blood clots, and visual impairments, and they are able to diagnose the condition of internal organs on the basis of ultrasound, x-rays, MRIs, etc.
Weak artificial intelligence technology has been an extremely popular topic lately, though some aspects continue to fly under the radar. In the next issue, we will take look at these issues together in more detail. Stay tuned for more!
Material prepared by:
Advisor to the President for Economy and Finance, JSC Business Alliance
PhD reseacher in Computer Science & Machine Learning