Farah Baracat

Day 1 - Introduction to Neuromorphic

Morning Session

• Introduction
  • History
  • Logistics (website, Wiki,,,,)
  • Goal

The main goal is to discuss:

• what ideas/principles we can extract from ML and AI and implement them on the hardware?
• What can this community bring: what inspirations from the nervous system

Sebastian leading the introduction to the field:

• write down 6 words in the survey!
• Postdoc at Heidelberg, part of the brainscale
• Workshop started 1994:
  • Origins: Publication of Carver Mead in 1990
    • Shows the current of ion channels depends on the voltage
    • Measures I_syn (function of V)
    • Asks: why not use the transistor in all its complexity: analog to do physical omputation. Some simple computation primitives:
      • conservation of charge (summation)
      • I = V/R (multiplcation). This R also stores the value, “resistive memeory”? –> memeory computation: co-location fo computation and memeory. The computation is hence defined by the network structure. To be contrasted with the Von Neumann bottleneck: transferring the info to and from the memory (DRAM) requires a lot of energy
• Spikes: a nice analogue between engineered systems and nervous system: to transmit signals over distance we need to encode and decode
  • temporal sparsity: although it computes faster, it only spikes sparsily. The reason is that spikes are expensive (number of spikes prop to the energy consuumed) –> whihc means that each spike has a high information content –> so how to encode signals into spikes? what coding scheme should we implement in neuron-inspired signal?

2nd Morning Session

• Bio-inspired learning. A common theme is locality: local variables necessary for learning
  • STDP
  • Homeostatis
  • Structure plasticity

• Noise and computation in the brain

  • A crucial element in the brain
  • Maybe what we define as noise is actually computation. Noise as a ressorce
• Carver Mead recent article on analog computation - NECO [Tobi]
• precison (the components are different: ie mismath on transistors or memristors) vs. accuracy (on the system level)

Afternoon: Workgroups/ Discussions

• Discussion group by Micheal: Single-shot learning

• Workgroups
  • Denoising for speech enhancement: time series, generative network, unsolved in SNN yet
• Neuromorphic Chips
  • Analog: BrainScales, DYNASE, ALIVE,ROLLS
    • BrainScales: are accelerated in time (x1000), 512 AdExp
    • Spinaker: ARM 120 nm,
  • Digital: Loihi, Spinnaker, Xylo, Spech