Online registration is open at http://www.regonline.com/sss15
Slides are here.
Symposium location: History Building (200), Room 002, lower level
Poster sessions: History Building (200), Room 002 and foyer, second level
Early work on knowledge representation and inference, which was done in the AI community back in the 1980s, was primarily symbolic. Later, symbolic approaches fell out of favor, and were largely supplanted by statistical methods. This symposium will try to close the gap between these two paradigms, and aim to formulate a new paradigm that is inspired by our current understanding of how humans solve these tasks.
Both symbolic / structured approaches and distributed / statistical approaches have a long history, and both have strengths and weaknesses. For example, symbolic systems are able to represent and reason with crisp rules, and distributed systems are able to represent (and to a much lesser extent, reason with) fuzzy concepts. It is widely believed that "general" AI systems will need both forms of functionality. This dichotomy was widely debated during the first "connectionist revolution" in the 1980s. We feel the time is ripe to revisit this discussion, based on the development and wide availability of massive symbolic knowledge bases (e.g., Freebase) on the one hand, and recent advances in deep learning on the other. While historically this research has been conducted in the computer science community, we would like to bridge between this work and the study of human cognition.
We will encourage symposium participants to explain how they would approach certain primitive, but central, kinds of reasoning patterns in their systems. For example,
1. Given a rule such as p(x, y) ^ q(y, z) => r(x, z), and observations that p(a,b) ^ q(b,c), how would you conclude r(a,c)? (Note that the method should work for any x, y, z).
2. If you were told that Beer is to Germany as X is to England, how would you find plausible values for X?
3. Short/long-term memory: We give the system a set of (person, phone number) pairs - how does it store and retrieve them? And how does it not confuse such verbatim memory with cases where we want it to generalize?
We will also encourage participants to discuss ways to evaluate their knowledge representation / inference systems and to demonstrate improved performance on one or more practical tasks.
Please register before Feb 27 at http://www.regonline.com/sss15.
When registering, please choose symposium #4.
If you need accommodation for your stay, please note the following cut-off dates for the hotels. Reservations made after the dates below are subject to availability and then-current room rates:
Cardinal Hotel reservation cut-off
Creekside Inn reservation cut-off
Further information about the AAAI spring symposium series (registration, accommodation etc.) can be found at http://www.aaai.org/Symposia/Spring/sss15.php
The symposium will include a mix of invited and submitted (peer-reviewed) contributions, which will be presented as formal talks as well as posters. The symposium will be highly interactive, with longer Q&A sessions after each talk and plenty of time for unstructured discussion. The last day of the workshop will be devoted to summarizing the various approaches presented at the symposium, and fusing them to formulate a hybrid research agenda for the field. We will also formulate a list of tasks that are expected to benefit from the improvements in knowledge representation and reasoning in the short to medium term, so that new approaches can be measured and evaluated on a common set of tasks.
Antoine Bordes (Facebook)
Leon Bottou (Microsoft)
Geoffrey Hinton (Google & U Toronto)
Jerry Hobbs (USC/ISI)
Doug Lenat and Michael Witbrock (CYC)
Percy Liang (Stanford)
Submissions should be up to 4 pages in PDF format, due on October 20, 2014 [UPDATED!]
Submissions should be made via the Easychair site at https://easychair.org/conferences/?conf=krr2015 ; no email submissions will be accepted. Submissions should not be anonymized, and the author names and affiliations should be displayed on the first page.