Summary of Round Table

Dynamical Systems and Brain Inspired Computing

Brussels : 31 May – 2 June 2017

Summary of the Round Table on 1st June 2017

Around the Table: Raul Vicente, Ingo Fisher, Herbert Jaeger, Joni Dambre, David Wolpert.

Moderator: Serge Massar.

  • What is the key/most important cost measure (energy, speed, ..) of (analog) computing systems?
    • Depends on application, each platform will have specific applications. No general-purpose device
    • Find similarities between different systems -> use synergies
  • Is reservoir computing a niche or is there place to grow (esp. considering Deep Learning)
    • Try to find underlying (common) information processing requirements
  • Why does reservoir computing work so well (in different systems/platforms)?
    • It's a cheap trick, that works in limited fashion only
    • Disagreement on amount of non-linearity needed
    • Can use almost any non-linearity to get decent results
  • Description in terms of non-linear system:
    • Reservoir Computing needs negative conditional Lyapunov exponents (with respect to external drive)
  • How to define computation? What is computation?
  • Where can reservoir computing / analog computing beat state of the art technology?
    • Different platforms have different advantages
    • Hardware limits number of trainable parameters
    • Having a completely autonomous hardware for analog computation is tricky
  • Is there a niche where analog can beat digital?
    • Not obvious
    • How to implement error correction?
    • Not necessarily needed, digital too precise for many task requirements
  • Can we change the methodology/paradigms of computing?
  • How large is "large enough"? How large has the reservoir to be?
    • Can we go from 3 bit header recognition to 40 bit header recognition?
  • Can we use reservoir computing to explore thermodynamics? Improve energy efficiency for ExaScale computing
  • How can we extend the uses/impact of analog hardware platforms?
    • At the moment: Lack of communication, should reach out to more communities
  • Comparison to Quantum computing:
    • It is as far away from usable computing platform as reservoir computing
    • QC: Promise of exponential speed up, in contrast to polynomial speed up learning
    • D-Wave is/is not a quantum computer?
  • Reservoir Computing is too simple an idea to found a community
    • Reservoir Computing is a tool
  • Brains show the possibility of efficient/analog/flexible computing systems. But we are very far away from understanding
    • We need to accept that we can't understand the brain soon
    • start with smaller brains (insects)
    • The important questions are bigger than reservoir computing