Skip to main content

Theories

Accepted sessions on the schedule. Please join the google calendar via the link in the discord.

📄️ Stephen Mell

Stephen Mell is a PhD student in computer science (programming languages and machine learning). He thinks that some ideas from linear logic and type theory could inform the theoretical discussion of abstractions. He is working on a type theory for abstractions (see termgraph workshop paper) that can represent rich structured data (like Church encodings, but beyond trees). Concerns have been raised that this is hazardous, as it might lead to novel neural network architectures on code (RNNs and tree RNNs are uses of Church-encoded lists and trees, and thus might produce an analogous architecture on structured representations of code). He’s looking to get feedback on (1) whether this is promising and (2) whether this is dangerous.

📄️ Vanessa Kosoy

The learning-theoretic research agenda is a programme for creating a mathematical theory of intelligent agents, and using this theory to solve the AI alignment problem. This programme draws on several areas of mathematics / theoretical computer science / decision theory, such as statistical learning, computational complexity, algorithmic information and imprecise probability. This talk will give you a small taste of what problems the agenda is concerned with, and how it might all come together to solve alignment.

📄️ NARS

Logic should return its focus to valid reasoning in real-world situations. Since classical logic only covers valid reasoning in a highly idealized situation, there is a demand for a new logic for everyday reasoning that is based on more realistic assumptions, while still keeps the general, formal, and normative nature of logic. NAL (Non-Axiomatic Logic) is built for this purpose, which is based on the assumption that the reasoner has insufficient knowledge and resources with respect to the reasoning tasks to be carried out. In this situation, the notion of validity has to be re-established, and the grammar rules and inference rules of the logic need to be designed accordingly. Consequently, NAL has features very different from classical logic and other non-classical logics, and it provides a coherent solution to many problems in logic, artificial intelligence, and cognitive science.

📄️ Alex Havrilla

Controlled automated story generation seeks to generate natural language stories satisfying constraints from natural language critiques or preferences. Existing methods to control for story preference either rely on labor intensive or often inconsistent prompt engineering. They may also use logit-manipulation methods which require annotated datasets to exist for the desired attributes. To address these issues, we first train a contrastive bi-encoder model to align stories with corresponding human critiques, named CARP, building a general purpose preference model. This is subsequently used as a reward function to fine-tune a generative language model via PPO-based reinforcement learning toward a particular set of preferences. However, simply fine-tuning a generative language model with a contrastive preference model does not reliably result in a story generation system capable of generating stories that meet user preferences. To increase story generation robustness we further fine-tune the contrastive model using a prompt-learning technique. A human participant study is then conducted comparing generations from our system, ablations, and a prompted LLM baseline. We show that the full fine-tuning pipeline results in a story generator preferred over a LLM 20x as large.