Spiking Neural Network

Project THALAMUS

Statutory Activities for Young Staff No. 0311/SBAD/0714, Faculty of Computing and Telecommunications, Poznan University of Technology

The aim of the project is to develop an algorithm for the creation and training of an impulse neural network (SNN) with the omission of the weight mechanism based on the thalamus model and on the developmental jumps of infants. The development of a child in the period up to the age of two is characterized by a step learning curve. Based on the observation of this behavior, it can be concluded that in the learning curve of real neural networks (RNN) a long time of zero progression is visible, and the most important stage is a sudden increase in the dynamics of the curve. The authors analyze the possibility of translating this mechanism into pulsating neural networks. The aim of the project is to test the possibility of complete elimination of weights in the connections between neurons and the implementation of a neural network solely on the basis of connections between neurons.

Project SNNTOOLS

The aim of the project is to develop a software framework for simulation, prototyping and deployment of spiking neural networks on CPUs, GPUs and edge devices. These tools will contain most popular neuron and synapse models, STDP-based (Spike-Timing-Dependent Plasticity) learning algorithms and novel techniques for hiperparameter tuning and reduction of computational complexity of spiking neural networks. Framework will use Tensorflow as backend providing a lot of inbuilt utilities and possibility to design hybrid neural networks (second-third generation).