SYmbIONT - Synthetic Images Of Neural neTworks
The project was aimed at creating a prototype for the generation, through the visual reality, of synthetic images for the training and testing of neural networks and application to a case of real use, identifying market prospects in order to increase the probability successfull.
The project intends to bring an innovation in the ambit of AI (Artificial Intelligence) or software of machine learning through the execution of a program that foresees the acquisition of two qualified services(B.1.1 e B.1.2):
- study and design of a prototype system for creating artificial images of various kinds, using Virtual Reality development environments, for training neural networks to recognize objects, events, real life situations (B.1.1, servizio di supporto alla ricerca e sviluppo e all’innovazione di prodotto e/o processo nella fase di concetto);
- study of application scenarios of the system most requested by the market and Proof of Concept to be proposed to potential customers (B.1.2, servizio di supporto all’introduzione di nuovi prodotti).
The aim of the project is to demonstrate that a Convolutional Neural Network CNN system can be developed that is trained on synthetic data capable of achieving a test error similar to a network that is trained on real-world data.
Synthetically generated RGB images can provide similar or better results than real-world data sets if a simple domain adaptation technique is applied. Furthermore, the study of application scenarios of the system in the market, from the identification of one of the most promising scenarios in terms of necessity and business to the realization of a Proof of Concept to be proposed to potential customers, constitute the guide for the development of a prototype capable of meeting the needs of the market