community-driven ai/sw/hw co-design and optimisation · 10/3/2017 · ai/sw/hw co-design...
TRANSCRIPT
Community-Driven AI/SW/HW Co-design and Optimisation
or how to win in a Cambrian in AI/SW/HW explosion…
Cambridge Wireless Technology & Engineering Conference, 3 October 2017
Anton Lokhmotov, CEO and co-founder, dividiti
… with cKnowledge.org and open co-design competitions
cKnowledge.org : helping industry and academia adapt to a Cambrian AI/SW/HW explosion via open co-design competitions (2 of 24)
A race to develop innovative AI products and systems (SW & HW) …
Various form factors: IoT, mobile, data centers, supercomputers
Various constraints:speed, energy, accuracy, size, resiliency,
costs
cKnowledge.org : helping industry and academia adapt to a Cambrian AI/SW/HW explosion via open co-design competitions (3 of 24)
… leads to a Cambrian AI/SW/HW explosion and technological chaos
cKnowledge.org : helping industry and academia adapt to a Cambrian AI/SW/HW explosion via open co-design competitions (4 of 24)
Which AI/SW/HW solutions will survive?
AI users
We at dividiti.com performcompetitive analysis
and optimization of the whole AI/SW/HW stack for various realistic scenarios
(object detection, image classification, etc)
cKnowledge.org : helping industry and academia adapt to a Cambrian AI/SW/HW explosion via open co-design competitions (5 of 24)
AI researchers should care about real constraints
cKnowledge.org : helping industry and academia adapt to a Cambrian AI/SW/HW explosion via open co-design competitions (6 of 24)
Optimising for benchmarks means:
High benchmark scores (maybe).
Low relevance of optimizations to real workloads.
Lost opportunities to make more competitive products.
Optimising for real workloads implies the need for:
Systematic experimentation across many workloads, platforms, system
conditions, input data, user inputs, etc.
Systems designers should care about real workloads
A. Krizhevsky et al.
“ImageNet classification
with Deep Convolutional
Neural Networks” (2012)
cKnowledge.org : helping industry and academia adapt to a Cambrian AI/SW/HW explosion via open co-design competitions (7 of 24)
Crowdsource benchmarking across Android devices provided by volunteers
Continuously collect statistics, bugs and misclassifications at cKnowledge.org/repo
The number of distinct participated platforms:800+
The number of distinct CPUs: 260+
The number of distinct GPUs: 110+
The number of distinct OS: 280+
Power range: 1-10W
Also collecting real images from users for misclassifications to build an open
and continuously updated training set!
Winning solutionson various frontiers
Tim
e p
er
imag
e
(se
con
ds)
Cost (euros)
cKnowledge.org : helping industry and academia adapt to a Cambrian AI/SW/HW explosion via open co-design competitions (8 of 24)
Caffe, TensorRT 1.0 EA (released with NVIDIA & GM permission)
cKnowledge.org : helping industry and academia adapt to a Cambrian AI/SW/HW explosion via open co-design competitions (9 of 24)
Caffe, TensorRT 1.0 EA (released with NVIDIA & GM permission)
F.N. Iandola et al. “SqueezeNet:
AlexNet-level accuracy with 50x
fewer parameters and <0.5MB
model size” (2016)
cKnowledge.org : helping industry and academia adapt to a Cambrian AI/SW/HW explosion via open co-design competitions (10 of 24)
● Bring together industry and academia to participate in open and reproducible AI/SW/HW co-design competitions using the CK framework.
● Share more artifacts, workflows and results in a reusable and customizable CK format (common JSON API and meta description).
● Collaboratively improve models and find missing features.● Gradually expose more design and optimization knobs at all AI/SW/HW levels.● Enable distributed online learning for self-optimizing and self-learning systems.
http://cKnowledge.org/partners http://cKnowledge.org/publications
Join the growing Collective Knowledge community!