Archive for November, 2015

Graph Isomorphism is “just” Quasipolynomial

Well if Graph Isomorphism was a person you will feel for that person, its no more special. For long we have been saying its NP complete or not and now its no more ­čśŽ If it was person I will say Sorry Mate!

Anyhow, its wonderful news in theoretical computer science but  just realized its going to take a long time to become practical in computer vision applications. However it does require long thinking.

TensorFlow: another Deep Learning library?

The week starts with interesting but predictable news. There is one more player there in Deep Learning. We had Caffe, Torch, Theano and now we have TensorFlow vying for our deep learning solutions.

I have questions, for example why its called Tensor Flow? might be because they are not presenting themselves as another Deep Learning library but as a library that can represent flow of Data in a graph. Those graph nodes could be computational or just data-pushers. That makes me think of them different from other deep learning libraries, atleast from Caffe.

Let me confess, I have only used Caffe. And I have found it really easy to use, however getting around setting up specific network with different form of learning looks really difficult thing to do. Torch, I heard lets you setup your own learning mechanism, however I have never worked with it. As per recent┬áTombone’s┬ásmall survey more people are inclined to use Caffe, might be because its easy to use.

TensorFlow looks exciting, their tutorial section is also well curated. It has shine and smoothness of the any product made by the big corporate rather than roughness we see in most of academic based products. However how much its “open source” is really an open-source (easy to change, easy to update, etc…) and how much we can play with the code rather than just use it as API is going to define where it will be used. My feeling is that TensorFlow’s API will be in use pretty quickly however whether it will be used to re-imagine existing Deep Learning Paradigm is still a question.