GTC Silicon Valley-2019: Training Spiking Neural Networks on GPUs with Bidirectional Interleaved Complementary Hierarchical Networks
Note: This video may require joining the NVIDIA Developer Program or login
GTC Silicon Valley-2019 ID:S9151:Training Spiking Neural Networks on GPUs with Bidirectional Interleaved Complementary Hierarchical Networks
Brent Oster(ORBAI),Anusha Swamy(ORBAI)
We will discuss using spiking neural networks on GPUs to train audio and vision networks with our new BICHNN architecture. Spiking neural networks are more brain-like and have richer time-domain signal processing behavior than traditional feed-forward networks, but back propagation and gradient descent don't work on spiking neural nets. We'll describe how we train our models using our Bidirectional Interleaved Hierarchical Neural Networks. We will show how we construct these spiking neural networks in our NeuroCAD visual design tool, then connect them up with parameter-driven probability maps to give us the BICHNN architecture. Then we will turn them loose and train them in real-time and demonstrate how we use genetic algorithms and massive amounts of GPU simulation time to optimize the networks to the specified task.