About ADAGOS - The story
Where do parsimonious neural networks come from?
ADAGOS, founded in 2011, is a spinoff of the IMT (Institut de Mathématiques de Toulouse).
ADAGOS breakthrough in deep learning techniques via introduction of parsimonious neural networks finds its origin in the outstanding scientific career of Mohamed MASMOUDI, Professor of Applied Mathematics at the University Paul Sabatier (France), where he participated to the creation of the Institute of Mathematics of Toulouse (IMT).
Mohamed MASMOUDI is indeed well known for his contribution to topological optimization techniques in optimal shape design used by the industry in various fields ranging from Mechanics to Electromagnetism. A main issue in topological optimization is to decide which nodes to remove from a mesh representing the shape of a given design.
Earlier in his scientific youth, he had also worked on neural networks, at the time where only single layers were used. As it often happens in human discoveries, the initiating sparkle came from the combination of two different domains, here topological shape optimization and neural network. The crucial idea introduced by Mohamed MASMOUDI was then to use topological optimization criteria for designing neural networks by providing a mean for selecting those nodes and links that can be added. The result is then no less than a parsimonious neural network, a powerful tool for solving large and complex problems, using less resources - by several orders of magnitude - than the usual neural networks, in terms of data and computation.