Bob just sent me the following:
Did you see the workshop I am organising in the Autumn?
I would really appreciate it if you could help advertise it. I am aiming to a good variety of presentations. So far, I have contributions from acoustic surveillance, finance, reproducible research software, and instagram like prediction. I have two keynotes already.....
Bob mentioned that impotant topic when he visited the Paris Machine Learning a few months ago. This is a fundamental question in Machine Learning these days. From the page:
On “Horses” and “Potemkin Villages” in Applied Machine LearningResearch workshop, QMUL, London
Monday 19 September 2016
Have you uncovered a “horse” in your domain?* Or perhaps discovered a “Potemkin village”?†
CALL FOR CONTRIBUTIONS
We invite presentations for this free one-day workshop (with free coffee & nibbles and a free lunch), which will explore issues surrounding “horses” and “Potemkin villages” in applied machine learning. One of the most famous “horses” is the “tank detector” of early neural networks research (https://neil.fraser.name/writing/tank): after great puzzlement over its success, the system was found to just be detecting sky conditions, which happened to be confounded with the ground truth. Humans can be “horses” as well, e.g., magicians and psychics. In contrast, machine learning does not deceive on purpose, but only makes do with what little information it is fed about a problem domain. The onus is thus on a researcher to demonstrate the sanity of the resulting model; but too often it seems evaluation of applied machine learning ends with a report of the number of correct answers produced by a system, and not with uncovering how a system is producing right or wrong answers in the first place.
The day will feature two keynote lectures — and free coffee & nibbles and a free lunch — but we are looking for contributions to the day in the form of 20-minute talk/discussions about all things “horse” and "Potemkin village." We seek presentations from both academia and industry. Some of the deeper questions we hope to explore during the day are:
- How can one know what their machine learnings have learned?
- How does one know if and when the internal model of a system are “sane”, or “sane enough”?
- Is a “general" model a “sane" model?
- When is a “horse" just overfitting? When is it not?
- When is it important to avoid “horses”? When is it not important?
- How can one detect a “horse” before sending it out into the real world?
- How can one make machine learning robust to “horses”?
- Are “horses” more harmful to academia or to industry?
- Is the pressure to publish fundamentally at odds with detecting “horses”?
Please submit your proposals (one page max, or two pages if you have nice figures) by July 1, 2016 to email@example.com, subject line: “On ‘horses', in memoriam Alan Young (1919-2016)”. Notification will be made July 7, 2016. Registration (free) will then be opened soon after.
This event is funded with support from the EPSRC through the Platform Grant on Digital Music (EP/K009559/1), and is co-organised with the QMUL Applied Machine Learning Lab and Machine Listening Lab.
* “As an intentional nod to Clever Hans, a 'horse' is just a system that is not actually addressing the problem it appears to be solving.” (B. L. Sturm, “A simple method to determine if a music information retrieval system is a 'horse',” IEEE Trans. Multimedia 16(6):1636–1644, 2014.)
† “[Our] results suggest that classifiers based on modern machine learning techniques ... are not learning the true underlying concepts that determine the correct output label. Instead, these algorithms have built a Potemkin village.” (I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessing adversarial examples,” in Proc. ICLR, 2015.)
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.