Welcome! Please contribute your ideas for what challenges we might aspire to solve, changes in our community that can improve machine learning impact, and examples of machine learning projects that have had tangible impact.
Lacking context for this site? Read the original paper: Machine Learning that Matters (ICML 2012). You can also review the slides.
engage with domain-specific conferences
  • There are at least a couple of dozen NLP researchers who innovate in ML.  Surely this is also true in vision, robotics, etc.  Don't know about meteorology and other topics mentioned on this site.

    Many of us tend to publish our ML work in our domain-specific communities.  I'm not just talking about application papers, but papers where the main attraction is the novel learning settings, models, or inference methods.  Also, application papers that require interesting new twists on existing ML methods.  I can point to examples if people want.

    We publish at the domain conferences out of habit, but also in part because we've been burned in the past by ICML and NIPS reviewers who don't know enough about our domain ("what's a finite-state automaton?" "what's a dependency grammar?") or might not as easily appreciate what we're trying to achieve.  Of course that situation could be improved.  The ML conferences are indeed getting better about broadening their program committees.  But I'm not sure that having more application papers at ICML and NIPS is the right way to go. 

    Instead, why not try to infect the domain conferences with more ML papers?  This has already happened for the major branches of AI.  And there are people like Joshua Tenenbaum and Tom Griffiths who are doing their best to infect cognitive science with ML by publishing and evangelizing.  This is a great way to increase the impact of ML.  Publishing at a domain conference requires the paper to make a contribution to the domain (at least by the ivory-tower standards of the domain).  Furthermore, publishing there showcases the value of ML -- and the latest and most appropriate ML methods -- to people who may not yet be au courant, including students.  Conversely, the ML researchers who attend domain conferences, or at least read the proceedings, will get a better sense of the problems that need to be solved and where ML can help solve them better.  And they might find collaborators. 

    Kiri asked about the objective function of the field of ML.  But ML isn't the output layer of science; it's part of a hidden layer.  If you want to maximize your impact, you need to know how your work would be used downstream.  So, try to figure out what the recurring or important abstract problems are in the downstream fields, and back-propagate from that -- whether or not you directly run on their concrete problems.

    Of course, ICML/NIPS/COLT are still great places to describe general techniques that have been abstracted away from a particular problem.  That's why I'm at ICML right now. :-)
  • 4 Comments sorted by
  • As an example, some interesting features of the NLP domain are
    • structured random variables that range over strings, trees, grammars, automata, and terms of formal logic
    • lots of heterogeneous data resources that could be exploited with effort
    • estimation problems stemming from large vocabulary
    • huge unsupervised learning problems that have scientific or practical impact (some of these are solved by human language learners, some aren't)
    • domain knowledge of linguistic phenomena and linguistic formalisms
    • reviewers who expect you to tell a coherent story about why your approach fits the data, and to look carefully at how your approach interprets the data (is it behaving as you claimed it would?  what do the errors look like?)
    • real-world impact through machine translation, intelligent information access, dialogue systems, etc.
  • Jason, your idea of publishing in domain conferences is good but there is an issue of 1) outside CS conferences are generally 100% acceptance (so the publications don't count) and 2) many of them have so many talks that only the people who would already know about your stuff come to it.  The AMS meeting that I go to each year has around 15 parallel conferences running at once!
  • @amy -- sorry, I was thinking of within CS.  Is the correct mechanism outside CS to publish in domain-specific journals?  Are there other ways to organize ML discussion and education within an application community?
  • Yes, the talks at the conferences are often an advertisement for a journal paper.  Those are what count.

    I think the big issue is that at a CS conference you will get very few domain scientists.  If you reach out to their conferences, you will get fewer CS people but you will learn a lot more about what the scientists need/want from ML.


To post or add a comment, please sign in or register.

Tip: click the star icon to bookmark (follow) a discussion. You will receive email notifications of subsequent activity.
If search doesn't work, try putting a + in front of your search term.