Abstract
The success of deep learning has proven the power of neural networks as a trainable computational artifact. While artificial neural networks (ANNs) are biologically inspired, they differ in many aspects from their biological counterparts. One salient aspect is that biological neurons communicate with short voltage pulses, so-called spikes. Models of such spiking neurons exist and the resulting networks are called Spiking Neural Networks (SNNs). SNNs have several potential advantages over ANNs: they are provably more computationally powerful and they can give rise to more energy-efficient implementations in specialized hardware. However, SNNs were notoriously hard to train, rendering these potential advantages insignificant. Another salient aspect of biological neuronal networks is that their dynamics are modulated by longer time constants on several temporal scales, for example, due to adaptation processes in neurons and plasticity processes in synapses. In this talk, I will discuss recent work in my lab on how learning can be implemented efficiently in recurrent networks of spiking neurons and I will show how longer time constants in neuronal dynamics can be utilized by SNNs. Finally, I will discuss how Hebbian plasticity can be utilized by neural networks to solve complex Question-Answering tasks on stories.
More Information & COVID-19 compliance:
- The "masked" participants are picked up at the door of the SR7.
- We ask the participants to keep a distance of 1 m.
- We ask the participants to sign the attendance list.
- We ask the participants to remove the MNS (Mouth and Nose Protection) only at the marked seat.
- The premises are cleaned and disinfected.
- In addition hand and surface disinfectants will be provided.