Participate in collaborative benchmarking and optimization of Deep Neural Network frameworks including Caffe, TensorFlow, etc and available models in terms of speed, accuracy, energy, size and various costs using your Android device. See shared results at http://cKnowledge.org/repo .
This app uses open-source Collective Knowledge workflow framework to implement public optimization scenarios from shared AI artifacts:
It is a part of our long-term community initiative to enable Open Science and help solve open challenges in computer systems' research via artifact sharing and reuse:
You can find sources of this application (BSD license) for collaborative development and extensions here:
If you have questions or comments, do not hesitate to get in touch with us via this public mailing list:
You may read more about our long-term vision in the following publications:
*** Motivation ***
We have been struggling with a lack of computational resources and diverse workloads/data sets/hardware for our own research to make faster, smaller, more power efficient reliable self-tuning software and hardware for more than a decade! Indeed, computer systems are becoming very inefficient - it is nowadays not uncommon to obtain 10x speedups, 2x size reduction and 40% energy reduction for popular algorithms (DNN, BLAS, video processing) on latest hardware using so-called autotuning of various algorithm parameters and compiler optimizations. However, this process is extremely time consuming due to very large design and optimization spaces.
Eventually, we developed this open source Collective Knowledge technology (CK) to let the community share workloads, data sets, tools and experimental workflows in an open CK format via GitHub or BitBucket, crowdsource experiments (such as multi-objective benchmarking and multi-dimensional autotuning) across diverse devices provided by volunteers, classify solutions on the fly (active learning), apply predictive analytics, exchange knowledge, and reproduce results.
Your participation supports open science and reproducible research initiatives such as Artifact Evaluation at leading conferences (sharing experimental workflows with all related artifacts and results along with publications in a reproducible and reusable way to be validated by the community):
This community-driven initiative is coordinated by:
* http://cTuning.org (non-profit research organization)
Thank you very much for participating in experiment crowdsourcing and enabling open science!