Upcoming Dovetail fellow talks & discussion

As the current Dovetail research fellowship comes to a close, the fellows are giving talks on their projects. All are welcome to join!

The easiest way to keep track of the schedule is to subscribe to the public Dovetail google calendar. I’ll also list them here in this post, which I’ll update as more talks get scheduled.

All talks will be on Zoom at this link.

August 7 (Friday) 1400 GMT/0700 PT: Ayur Pulle—Congruences and Obstructions in Predictive Agents
August 8 (Saturday) 1600 GMT/0900 PT: Neal Batra—From Optimal Actions to World Models: Identifiability of Transition Kernels in Discounted MDPs
- Abstract: We study what can be recovered about the transition probabilities of a Markov decision process from optimal actions alone. This is closely related to the inverse problem considered by Letcher et al., who ask when the dynamics can be recovered from numerical Q-values. Here the numerical values themselves are not observed; only the optimal actions are known, for every reward in a given class.
August 10 (Monday) 1500 GMT/0800 PT: Orhan Bayraktutan—Optimal Policies and Policy Orderings
- Counterexamples to a conjecture in theoretical reward learning about when reward functions with the same optimal policies must induce the same policy ordering.
August 11 (Tuesday) 1400 GMT/0700 PT: Jennifer Benedict—Which Abstractions can be Transported Across World Models?
- Two agents can look at the same system with totally different internal representations and still agree on how it behaves. So, it is useful to understand what types of coarse-grainings do not depend on the description of the world model. This talk will present an attempt to answer that question for models of computation.
August 11 (Tuesday) 1800 GMT/1100 PT: Adithya Shenoy—Touchette-Lloyd for KL divergences
- A generalisation of the Touchette-Lloyd theorem to policies which optimise towards a goal distribution, as measured by KL divergence.
August 12 (Wednesday) 1700 GMT/1000 PT: Aurelio Carlucci—World models under coarse-graining do not generally admit categorical products
- Abstract: In a class of objects there may be various ways to combine them; moreover, for a chosen notion of transformation, we would like the combination to be well behaved with respect to these. In a nutshell, that is the idea behind categorical products, which generalise cartesian products, intersections, greatest common divisors, and more.
- If we fix an abstraction encoding world models, together with coarse-graining transformations between them, we would like a mereology of world models compatible with coarse-graining: a notion of “greatest common part” of two models. Categorical products are a candidate universal property for one particular way of combining world models, and are the natural formalisation of such a part.
- Unfortunately, no such product exists in general. The issue is already visible before any dynamics or probability enters the picture, and it recurs once additional structure is included: the common refinement of two world models is not itself a world model (in one case the map onto it fails to be a coarse-graining, in another the refinement is no longer compatible with the dynamics). Both failures say that the theories in play, while shaped like lattices of congruences, are not lattices of set-partitions. The absence of the product is itself informative, telling us which strategies for combining models fail to yield a valid model, and which transformations fail to count as coarse-grainings.
- This talk presupposes no prior knowledge of categories or world models, although familiarity with sets, probability and graphs is desirable.
August 13 (Thursday) 1600 GMT/0900 PT: Sumayya Manji—When does an agent learn its dynamics? An extension to Touchette-Lloyd
- If a sighted and adaptive agent reduces entropy by more than a sighted, non-adaptive agent could, does that mean it has learned something about the hidden dynamics, not just the state?