Two paper solution idea

# Two paper solution idea Oct 14, 2024 I had a kind of crazy idea. Last week Marcelo mentioned that we could apply Sixing’s RNN to the data used in my PLoS paper (a one-shot preferential choice task). I'm wondering if it might actually make sense to write that paper first. More generally, maybe we should have an RNN-focused paper and a task-focused paper. ## Advantages of a two-paper solution: - It's usually best to introduce one major new idea/method per paper (unless they are inherently co-dependent, which I don't think is the case here) - Sixing will be first author on the first paper introducing the RNN method he developed - We know the optimal policy (without memory constraints) for preferential choice, so we can quantitatively evaluate the RNN's performance - We can develop the representation analysis methods in a simpler domain. Laying this groundwork could lead to a stronger analysis of the more complex planning case. - In the simple task, we can probably increase the size of the RNN to the point where it effectively doesn't have memory constraints. This would allow us to more definitively characterize the effects of memory constraints on the RNN behavior and representations - The RNN's representations could speak to the recent value-policy debate (e.g. [Hayden & Niv 2021](https://pubmed.ncbi.nlm.nih.gov/34060875/), which has received a lot of attention recently. This is also something I've wanted to explore for a while. ## Paper breakdown ### Plan A - **Paper A1** introduces the RNN method and applies it to preferential choice - Sixing first author, Fred unclear - NatNeuro if we can support the predictions with existing neural data - Otherwise, PLoS CB, Comp Brain Beh, Open Mind - Could do our own followup imaging paper - **Paper A2** introduces the planning task and applies the RNN to it - Fred first, Sixing co-first - NatHB (unless we get unexpectedly striking results) - **Papers A...** applying the RNN to other tasks? - e.g. multi-attribute/risky choice, memory recall - if things go well, we could write a TICs/BBS article overviewing all this work along with related approaches ### Plan B - **Paper B1** introduces the planning task with only symbolic modeling. - Fred first, Sixing unclear - NatHB - **Paper B2** applies the RNN to several tasks (preferential choice, multi-attribute choice, planning) - Sixing first, Fred unclear - Psych Rev, Nat HB, Science/Nature are all possible depending on how strong and compact the results are Order isn't critical because the cogsci paper already introduces the task, but B1 first would be ideal. ## Possible Pitfall - This plan could delay publication of the first paper - It's plausible to me either A1 and B1 could get to submission-ready shape before the joint paper (especially B1). - We can always fall back on the original one-paper plan if things don't look promising - For Plan A, if A2 is published before A1 is submitted, it would undercut novelty substantially. - We should try to submit A1 before or shortly after A2 - OR: we could de-emphasize the RNN approach in A2, for example by not presenting many representational analyses. - Plan B requires that I can figure out a decent symbolic model - However, Sixing can work on the choice RNN model immediately, which would be part of both plans - If we like B, Fred will prioritize the symbolic model to determine feasibility - Any paper that focuses on the RNN method could be seen as too similar to [Marcel Binz's work](https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/metalearned-models-of-cognition/F95059E07AE6E82AE56C4164A5384A18) - However, all of his work (that I'm aware of) applies meta-learning directly to *external* learning/decision problems. We're applying meta-learning to *internal* problems of how to direct mental effort. You might call this [metalearning to metareason](https://fredcallaway.com/pdfs/AAAI_19_Planning_workshop.pdf).