Just yesterday, Silicon Valley was the epicenter of technological progress. Today, Africa will be the center of attention. How has it been able to catch up in just a few months?
Thanks to the rediscovery of the “knowledge and know-how of the creative process” born on the African continent, combining tradition and technology, Thierry Rayer, perpetual president of the Cercle d’Etudes Scientifiques Rayer, is on the way with young Africans to achieve the unimaginable with Univesae Analysis, that of the machine that learns, and that learns by itself.
Instead of carrying out the orders of a program, the machine can now acquire by itself, through experience, the necessary abilities to accomplish the tasks assigned to it, including those thought to be reserved for humans. The applications are immense: recognition of geometric shapes, of the symbolic forms of images, in order to understand the art and architecture of humanity.
Thierry Rayer is at the origin of this revolution. He is indeed the inventor of this application for deep learning, which characterizes a network of artificial neurons whose architecture and functioning are inspired by the brain. It is at the birth of this application of a new form of intelligence, at the emergence of an almost self-organizing system, that Thierry Rayer invites us. An approach that evokes the intellectual approach of an inventor at the crossroads of culture, human knowledge, art, computer science and neuroscience. A playful application that sheds light on the future of artificial intelligence without leaving aside the African origin of the culture of humanity. An exciting and accessible application that will take us to the heart of knowledge, to the secrets of humanity and will make us discover both our universal knowledge and a fascinating new world, which is already ours.
Artificial intelligence and pattern recognition from images, how does it work?
Artificial intelligence has proven its effectiveness in recent years in a growing number of applications. The recognition of shapes in images is possible thanks to a single type of algorithm that allows the detection of any shape, pattern or complex object in any type of image. Then a computer must be taught to imitate human vision to make it understand what it sees.
How can this be done?
The Artificial Neuron Network is in fact a real digital brain, or at least a simplified version inspired by the knowledge we have of our own! An ANN is a mathematical graph that links numerical values (neurons) by simple functions called activation functions (the equivalent of electrical impulses in our brain).
The input of the graph is a set of data to be analyzed, and its output is the response of the algorithm to our problem. Between the two, intermediate neurons are activated or not during a learning process: by traversing the network from input to output and vice versa, the algorithm finds correlations between the data, calculates and corrects its errors, and finally clears the roads leading to the solution of the problem.
The final response of a network will always depend on probability, the goal being to maximize the confidence of the algorithm (or minimize its error).
Recognition of geometric shapes in images: convolutional neural networks
A digital image is a set of pixels that contain a colour value, but not only that: their spatial organization is probably the most important information.
It is the set of local geometries that gives a general meaning to the image, a bit like assembling the pieces of a jigsaw puzzle.
The pattern recognition algorithms are thus composed of two successive RNAs. The first one aims at extracting the geometric attributes of an object using convolution operations. These attributes correspond to simple shapes (a right angle, a curve, …) translated into complex numerical values. Once extracted, they are used as input data for a second RNA specifically designed for image classification.
The two arrays are trained simultaneously using thousands of images annotated “with” or “without” the object to be detected. At the end of this process, the algorithm will be able to provide an answer as to the presence or absence of a shape in an image whose content we do not know a priori.
From theory to practice
For Universae Analysis it is the network architecture and the quality of the training data that will decide how complex the problem can be solved. In the case of pattern recognition, it is necessary to build a deep network, i.e. containing hundreds of thousands of neurons, to input several thousands of images.
The representativeness of the training images is essential. For example, if one wishes to detect a specific pattern in a table, one will have to present the algorithm with thousands of concrete examples, taking care to vary the orientation, style, size and illumination of the shape. In the absence of existing examples, data augmentation methods are used to generate new images synthetically.