For example, not all large-scale data processing problems parallelize well in the current state-of-the-art. There are problems that would greatly benefit from being able to throw thousands of machines at them, except that attempting to do so hits bottlenecks of an algorithmic nature. Finding better algorithms to get past those bottlenecks would be worth a lot of money to the right people.
I think the paradigm of large scale data processing is itself uninteresting to me. I want to study problems in machine vision and perception that are not well-solved simply by processing large amounts of data … i.e. problems where if you give me a large amount of data, it is not at all known what I should do with that data to produce a “good” solution to the problem. Once you know what you’re supposed to do with the data, I feel like the rest is just engineering, which is uninteresting to me. After 3 years of reading huge chunks of the vision literature, and contemplating this and discussing this with many other faculty and researchers, the consensus seems to be that even if such a problem did exist, no one would be interested in publishing results on it or giving you money to study it. This is why I need to hack myself to cause myself to want to study the “uninteresting” engineering / data processing aspect.
Surely there are mathematical research questions about how to do large-scale data processing? Research them.
For example, not all large-scale data processing problems parallelize well in the current state-of-the-art. There are problems that would greatly benefit from being able to throw thousands of machines at them, except that attempting to do so hits bottlenecks of an algorithmic nature. Finding better algorithms to get past those bottlenecks would be worth a lot of money to the right people.
I think the paradigm of large scale data processing is itself uninteresting to me. I want to study problems in machine vision and perception that are not well-solved simply by processing large amounts of data … i.e. problems where if you give me a large amount of data, it is not at all known what I should do with that data to produce a “good” solution to the problem. Once you know what you’re supposed to do with the data, I feel like the rest is just engineering, which is uninteresting to me. After 3 years of reading huge chunks of the vision literature, and contemplating this and discussing this with many other faculty and researchers, the consensus seems to be that even if such a problem did exist, no one would be interested in publishing results on it or giving you money to study it. This is why I need to hack myself to cause myself to want to study the “uninteresting” engineering / data processing aspect.