BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Deep Learning meets Control Theory: Research at NNAISENSE - Marco 
 Gallieri\, NNAISENSE
DTSTART:20190215T140000Z
DTEND:20190215T150000Z
UID:TALK118564@talks.cam.ac.uk
CONTACT:Alberto Padoan
DESCRIPTION:NNAISENSE inherits IDSIA’s long lasting track record of grou
 nd-breaking results in artificial intelligence (AI). From perception to re
 inforcement learning\, the company’s legacy of super-human performance p
 laces them in the right position to take AI technology into everyday contr
 ol systems. While AI approaches control problems from an information theor
 etical and statistical perspective\, control theory addresses issues conce
 rning the physical world with a strong focus on safety\, hard constraints 
 and theoretical guarantees. While control approaches can be very robust th
 ey can seldom suffer from conservativeness of their assumptions. This is b
 elieved not to be the case for AI\, where non-conservative results can be 
 achieved. AI performance depends mainly on the quality and the amount of d
 ata and no unified framework exist for the analysis of stability and robus
 tness. For this reason\, while deep learning is becoming the industry stan
 dard for perception\, its use in control is mostly limited to simulated or
  non-critical tasks. Combining the fields of control and AI has the potent
 ial for retaining best of both Worlds. The first part of the talk will bri
 efly introduce NNAISENSE’s research in this direction.  \n \nIn the seco
 nd part\, we will then introduce Non-Autonomous Input-Output Stable Networ
 k (NAIS-Net): a very deep architecture\, developed in collaboration with P
 olimi\, where each stacked processing block is derived from a time-invaria
 nt non-autonomous dynamical system. Non-autonomy is implemented by skip co
 nnections from the block input to each of the unrolled processing stages a
 nd allows stability to be enforced so that blocks can be unrolled adaptive
 ly to a pattern-dependent processing depth. NAIS-Net induces non-trivial\,
  Lipschitz input-output maps\, even for an infinite unroll length. We prov
 e that the network is globally asymptotically stable so that for every ini
 tial condition there is exactly one input-dependent equilibrium assuming t
 anh units\, and multiple stable equilibria for ReL units. An efficient imp
 lementation that enforces the stability under derived conditions for both 
 fully-connected and convolutional layers is also presented. Experimental r
 esults show how NAIS-Net exhibits stability in practice\, yielding a signi
 ficant reduction in generalization gap compared to ResNets. 
LOCATION:Cambridge University Engineering Department\, Lecture Room 5
END:VEVENT
END:VCALENDAR