Rechercher des projets européens

A biologically inspired algorithm for training deep neural networks (DEEPLEARNING)
Date du début: 1 juin 2014, Date de fin: 31 mai 2015 PROJET  TERMINÉ 

In machine learning, deep neural networks are powerful computer-based models that use layers of computational units. Current commercial applications for these models include a wide array of software tasks such as image classification, identification of potential drugs, market predictions and speech recognition. Network models must be ‘trained’ using data, and their success hinges critically on the quality of the learning algorithm that is employed. We have recently discovered a novel, biologically inspired algorithm for training deep neural networks that is simpler to implement, more flexible and finds better solutions than existing techniques on bench-mark tests. Thus, our system has the potential to improve performance widely across the many fields that make use of machine learning in software tasks. Furthermore, the simplicity and flexibility of our method means that it could be more easily exploited in hardware devices such as mobile phones and cameras. The central aim of this proposal is to move our new algorithm to a stage where it is ready for commercialization. To do this we plan to accomplish two main areas of work. First, we will research the optimal way to employ our algorithm, establish its performance on a comprehensive set of industry-accepted bench-mark tasks, and compile our research into a manuscript for publication in a leading machine learning journal. Second, we will secure any arising intellectual property in line with the preliminary US patent application that we have already filed, assess application of the algorithm to the different commercial sectors identified through market research, and generate commercial interest in the technology through targeted marketing to relevant companies. This plan of work will confirm the innovation potential of our new algorithm and will establish the technical and commercial feasibility of our discovery.