Caffe for High Performance

[Surzeet Saikia]

Hi All

I have been planning to work on Deep learning for Human Action recognition. For now my goal is to publish a paper in the student symposium at IEEE HiPC conference. 

I would like to know how the Caffe framework will be ideal for this task?

I have some specific doubts

1: For such Conferences what kind of work will be ideal if I am going to look for CNNs for recognition task?

2: What all things I can do using caffe so that I can come up with a good publication ?

Thanks

[Manuel Rodriguez]

Am Mittwoch, 26. August 2015 06:48:18 UTC+2 schrieb Surzeet Saikia:
>  I would like to know how the Caffe framework will be ideal for this

[IEEE HiPC] task?

The famous ImageNet Challange is a High-Performance-Challange by itself. The
abbreviation "ILSVRC 2015" means:
ImageNet Large Scale Visual Recognition Challenge

On last year, nearly all teams used GPU Clusters over many weeks to solve
the problems. That means, in the caffe-environment you won't find any
non-High-Performance-Computing. The only way to be to be more large that
large is to use quantum computers. But it is unclear, if Caffe-Framework
will run on dwave's systems. Good luck for writing your paper.

[Surzeet Saikia]

hi Manuel 

Thanks for your reply :)

Actually I was aiming for ILSVRC 2015 challenge, but i am new to caffe.

 I used to do lots of cuda programing, but in this imagenet challenge i am wondering how to use caffe. 

i mean what kind of optimization i can do using caffe for this challenge ?

thanks

[Manuel Rodriguez]

Am Mittwoch, 26. August 2015 11:45:01 UTC+2 schrieb Surzeet Saikia:
> i mean what kind of optimization i can do using caffe for
> this challenge ?

The number of projects who come to the conclusion that caffe and
deeplearning is well suited for nearly all kind of problems is
endless. There is no need for additional work in that space. The open
question is: What are the limits of deeplearning? What isn't possible to
implement with caffe?

There is a need for a DeepLearning AI Winter, like the
last one in 1969 where Marvin Minsky showed that it's
impossible to solve the XOR Problem with a neural network,
https://en.wikipedia.org/wiki/Feedforward_neural_network

[Mạnh Linh Trần]

Hello! I am also researching this field. Have you built any software yet? Could share code with me? My email address is: linht...@gmail.com or tranmanhl...@gmail.com. thank you