We are developing an open-source Collective Knowledge framework aka CK -
a cross-platform customizable Python framework used by leading universities, Fortune 50 companies and non-profit organizations
to share artifacts as reusable components with JSON API;
quickly prototype portable experimental workflows (such as multi-objective DNN optimization);
automate package installation;
crowdsource and reproduce experiments across diverse hardware;
unify predictive analytics;
enable interactive articles,
and develop sustainable and customizable research software.
We participate in international research projects
with the leading universities, companies
and international non-profit organizations to help scientists use our open-source
Collective Knowledge framework (CK)
and implement sustainable research software, share reusable artifacts and workflows,
and crowdsource their experiments across diverse platforms provided by volunteers
similar to SETI@home.
We use CK to enable interactive and reproducible articles with reusable artifacts and experimental workflows which have unified JSON API and meta information!
We help conferences, workshops and journals including CGO, PPoPP, PACT and SC
to develop a common experimental methodology and framework
for artifact evaluation
and digital libraries such as ACM DL.
We also promote open, collaborative, reproducible and reusable research
as well as our new publication model with the community-driven reviewing
and validation of results.
We are developing portable, customizable and multi-objective autotuning workflow powered by CK
to help the community automatically crowd-tune and crowd-fuzz compilers such as GCC and LLVM in terms of execution time,
code size, compilation time and bugs. We also use it crowd-tune OpenCL, CUDA and other libraries,
and co-design various DNN engines and models.
For example, such autotuning workflow can help gain up to 10x performance speedups and 30% energy savings
on various OpenCL/CUDA/CPU libraries without sacrificing accuracy (or up to 20x speed-ups with a small accuracy degradation)
as shown in a CK-powered interactive graph below
for a popular SLAM algorithm
(all points can be reproduced/validated on practically any platform using CK):
Having a common experimental infrastructure allows us to
build reusable, realistic, diverse,
and continuously evolving training sets in a common format
(programs, data sets, models, unexpected behavior, mispredictions)
with the help of our partners and the community.
See the following examples of shared training sets: