COMPETITION FOR N.1 RESEARCH CONTRACT, GROUP 02/PHYS-02 THEORETICAL PHYSICS OF FUNDAMENTAL INTERACTIONS, MODELS, MATHEMATICAL METHODS and 01/INFO-01 INFORMATICS, DPT. COMPUTING SCIENCES, UNIVERSITÀ COMMERCIALE “LUIGI BOCCONI” DI MILANO

The project aims to develop and validate learning rules that are local in space and time for recurrent neural networks, based on the geometry of stable states and compatible with low-power neuromorphic processors. Activities include theoretical analysis for continuous and finite-precision units, convergence analysis, assessment of scalability and performance in online, continual and few-shot learning, quantization of weights and activations, and validation on emulators and physical hardware. Accuracy, energy consumption and memory usage will be measured. In collaboration with the Principal Investigator and other collaborators, the researcher will develop and implement algorithms, conduct and document experiments, analyse results and compare them with the state of the art. The researcher will also adapt the algorithms to hardware constraints and prepare the release of documented, reproducible open-source software.