Training deep spiking networks with the N2D2 deep learning framework
The N2D2 deep learning library is a new open-source C++ based deep learning framework launched by CEA.
Despite its lightweight implementation (it should run on most Ubuntu systems without installing additional dependencies), N2D2 is capable of implementing all state-of-the art networks for image recognition (LeNet, GoogleNet). Most architectures can be implemented without touching the source code by a simple plain text configuration file. N2D2 supports acceleration with CUDA and is available for Windows, Ubuntu, RHEL and Mac.
In contrast to its other deep learning counterparts, N2D2 fully embraces spike coding and implements spike-transcoding of trained deep networks. Additionally, it contains an event-based simulator for training networks directly with spikes (for instance with STDP). This makes it in particular interesting for the simulation of large-scale neuromorphic hardware implementations of deep learning architectures.
This workshop will give a short introduction to the main functionalities of the N2D2 library and its main functionalities. The source code and a manual can be downloaded on Github:
Although it is recommended to download the files before the workshop, a copy of the open-source version will be distributed in the workshop.
The source code and a tutorial with several examples are available under:
For a demonstration of training with STDP, see for instance:
No timetable published yet.