Coverage by Bhat Dittakavi of Variance.AI on “Journey of a Researcher” by a panel at IIITH during Computer Vision Summer School at IIIT Hyderabad on 6th July 2017.
Panel: Narendra Ahuja @ULUC, Gaurav Sharma @IITK, Karteek Alahari @InRia, Ajit Rao @Qualcom
Moderator: PJ Narayan
Q) How do you select a PH.D problem? Hoe do you select an approach to solve the problem?
Q) How are things diffeent today from the past in research?
Q) How do you do cutting edge research sitting in India?
Q) Advantages India has in doing such things?
Gaurav: You like an idea and hence you pick the problem. If you are interested in CV generally, then scan the research papers widely with abstracts and read them. From here you shortlist the ones you redo. Also you can find the problem that hasn’t been solved.
Race or not, choice had to be personal. The advantage is you have many solving the same problem and hence more awareness in a conference.
Karteek: The most important problem must speak to you. You should be able to come envision. Take big visionaries and hear the problem set they would like to be solved.
There is aspect of race to it as many groups may be working on the same problem. Each group proposes a different kind of solution. That si to our advantage to bring a variety of solutions.
Ajit: IIITH is the best place in India. Look at the problems attempted by IIITH.
Choose the problems that apply to the real world. Can avoid the problem is sometimes as good as can I solve the problem. So, take the compelling one.
Narendra: A fork got a resonating frequency, if you take it to a vessel, it resonates too the same way. Tune to the societal problem so empathy is there. You are made restless and you have this CV filter that you can use to solve it. Your approach is to try whatever gets in the nail, not the hammer. Today we are in a position to do more than ever as everything is a thing. Solve a local problem. Whatever you do will be cutting edge. You be the tuning fork that resonates with the society where people don’t insulate themselves.
PJN: Race with other group is relevant ad the things are changing much faster. Caffe or Theano is available to all. Code is readily available so one doesn’t have to reimplement from scratch. Tricks of the trade matter in the real life. There is more to it that is mentioned in the paper.
Q) Autonomous driving is a wrong problem for India. Taking that makes it harder for us more than expected. In the end researchers have to graduate too. Scientists need to pick the problrms that need to be solved.
Ajit) Advantage of indian problems is availability of indian data. Work we do is open for replication from the researchers view point. You can build on others work easily. Partition the problem well enough and solve your part. This I see as bug advantage.
Karteek) I see people as big advantage. Buying a tailor made suits is cheaper in India. Best algorithm is fully supervised data. Annotation of data is possible with the help we can get from several people. This is not technically super challenging but we bring the awareness. They csn be also a excited to get pulled into research on this information age.
Gaurav) There are local indian problems for which algorithms exist. The question is to implement. The disadvantage in Indian scenario is we are missing the middle layer that does engineering. Industry collaborates with academia and this is a full pipeline. Advantage is “people”. It is relatively easy to build and cultivate a broad set of competencies. This helps in panning projects that a real more diverse.
Internet era I find it more stressful. You see in archives others solving your problem that gives more stimulation and also stress. Because you can see what their next paper can be with the information available, you can differentiate yourself.
Q) New Era has come and people from signal processing domain say are finding it hard to adapt. Need for domain shift is critical. How did you tackle a domain shift?
Ajit) It is about learning to dig the hole. This is PH.D. If someone else is digging the hole with a spade so where else, that si what it is. Idea is to learn how to solve a prkblem.
Gaurav) Adaptability is the key. You get used to it and you make other feel comfortable.
Q) India is human rich country. How did few people crated ImageNet? How do you convince the people to annotate or create a dataset?
Gaurav) You pay them. Institutions invest in dataset. VGG dataset with 2.6M images got annotated in Hyderabad. Reward works. An undergrad or class 12 with internship can get exposed and create dataset. Motivated professionals do data creation.
Ajit) Crowd source like tge way Captcha does it. Csn you use it as incentive for done other transaction or application?
Q) How is it possible for a remote student in a remote college to get latest advances in the field?
PJN) You subscribe to archive and you are done. Fresh PH.D. student can’t comprehend the whole context of a paper right.
Gaurav) Barrier to enter is low. Use ImageNet. Struggle for a while.
Karteek) At IIIT students from elsewhere come and spend time on campus to learn.
Q) How do we balance the need for low level nuisances versus Hugh level things in Ph.D?
Karteek) Take the help of your supervisor or even senior researchers. Rely on peers and others.
Ajit) Intellectual property landscaping! Largely they can bring a whole 2D chart of who is doing what.
Gaurav) This is tree versus forest problem. We have to switch seamlessly over time as we get better. Incest heavily in trees in the beginning. As things get easier, you csn is late the stress of low level hard work that gives you time to think at high level.
PJN) PH.D. has three problems. Advisor gives you problem to find the solution. First step. Student defines the problem. Next step. Next is student finds the problem and also solution.
All major conference don’t merit the un-reviewed archives.
Jawahar) Be extremely happy that there is a paper like that. Great people think alike.
Gaurav) New problem, new solution or extensive empirical solution. At least diffeentiate in two of these three axes.
Q) Comparison between Ph.D. graduates and their papers?
Karteek) Contribution you have made to solve a problem. Did it stand the test of time? How important a particular paper had been? There awards people get for bring active paper for past 10 years.
PJN) Readiness to go forward in research is the real measure. Your past may not be indication of future.
Gaurav) In France, people don’t appreciate if you don’t have front math background. In USA, people appreciate a brilliant idea.
Q) Is it good to invest in two problems simultaneously?
Karteek) You will have a bigger goal in mind. There is diffeent set of prior knowledge foes into diffeent types of levels within the semantic segmentation as a global problem. So attach same domain problems simultaneously.
Gaurav) There is an assumption you are stuck with a problem. Don’t feel the way. Pick up some skills. It will come.
PJN) PhD is not an exercise of solving the already problems the same way. You may hit a block but there is golden pot at the end of the rainbow. Go in one direction and choose diversion.
Amar) It is not necessary that you have to be the first to solve. Qualcomm started as trucking industry and then pivoted. This is not a zerosum game.