edge-ai for intelligent user experience kal mos
TRANSCRIPT
Behavior Modeling
Action Automation
Adaptive UI
User Sensing
User Identification
Intelligence on the Edge
Analytics for BigData
Connectivity is not stable: bandwidth, latency, availability..
Battery Power
Limited resources (Processing, Memory,…)
Algorithm in the wild (self training & decision making)
Challenges at the Edge
S1
a2-‐n
a1
Deciding to Automate
S2
S3-‐n
automate
don’t automate
rc = -‐1*(C.L.+Annoyance) = 0
rw = <<0
E[r]!automate = <0
P(A=ac|S1)
1 -‐ P(A=ac|S1)
Deciding to Automate
If then automate the action
This only happens when is high enough to counteract rw
Uncertainty with Drop Out
http://mlg.eng.cam.ac.uk/yarin/blog_3d801aa532c1ce.html
1. Apply Drop Out at inference time
2. Doing this many times will create samples from the learned posterior distribution
3. Estimate the moments of the posterior from these samples (Monte Carlo Integration)
4. The variance represents an estimate of uncertainty
• Having model uncertainty is important for instant decision making in the Edge
• Getting model uncertainty estimates in deep systems can be based on Monte Carlo sampling methods using dropout
• Doing this with parallelization works well on a GPU.