Autotelic Agents that Use and Ground Large Language Models (Oudeyer)

Pierre-Yves Oudeyer, Inria, Bordeaux, 21 mar 2024

ABSTRACT: Developmental AI aims to design and study artificial agents that are capable of open-ended learning. I will discuss two fundamental ingredients: (1) curiosity-driven exploration mechanisms, especially mechanisms enabling agents to invent and sample their own goals (such agents are called ‘autotelic’; (2) language and culture enabling enabling agents to learn from others’ discoveries, through the internalization of cognitive tools. I will discuss the main challenges in designing autotelic agents (e.g., how can they be creative in choosing their own goals?) and how some of them require language and culture to be addressed. I will also discuss using LLMs as proxies for human culture in autotelic agents, and how autotelic agents can leverage LLMs to learn faster, but also to align and ground them on the dynamics of the environment they interact with. I will also address some of the current main limitations of LLMs.

Pierre-Yves Oudeyer and his team at INRIA Bordeaux study open lifelong learning and the self-organization of behavioral, cognitive and language structures, at the frontiers of AI and cognitive science. In the field of developmental AI, we use machines as tools to better understand how children learn, and to study how machines could learn autonomously as children do and could integrate into human cultures. We study models of curiosity-driven autotelic learning, enabling humans and machines to set their own goals and self-organize their learning program. We also work on applications in education and assisted scientific discovery, using AI techniques to serve humans, and encourage learning, curiosity, exploration and creativity.

Colas, C; T Karch, C Moulin-Frier, PY Oudeyer (2022) Language and Culture Internalisation for Human-Like Autotelic AI Nature Machine Intelligence 4 (12), 1068-1076 https://arxiv.org/abs/2206.01134

Carta, T., Romac, C., Wolf, T., Lamprier, S., Sigaud, O., & Oudeyer, P. Y. (2023). Grounding large language models in interactive environments with online reinforcement learning. ICML https://arxiv.org/abs/2302.02662

Colas, C., Teodorescu, L., Oudeyer, P. Y., Yuan, X., & Côté, M. A. (2023). Augmenting Autotelic Agents with Large Language Models. arXiv preprint arXiv:2305.12487. https://arxiv.org/abs/2305.12487

« Algorithmes de Deep Learning flous causaux » (Faghihi)

Usef Faghihi , Informatique, UQTR, 16-Nov 2023

RÉSUMÉ : Je donnerai un bref aperçu de l’inférence causale et de la manière dont les règles de la logique floue peuvent améliorer le raisonnement causal (Faghihi, Robert, Poirier & Barkaoui, 2020). Ensuite, j’expliquerai comment nous avons intégré des règles de logique floue avec des algorithmes d’apprentissage profond, tels que l’architecture de transformateur Big Bird (Zaheer et al., 2020). Je montrerai comment notre modèle de causalité d’apprentissage profond flou a surpassé ChatGPT sur différentes bases de données dans des tâches de raisonnement (Kalantarpour, Faghihi, Khelifi & Roucaut, 2023). Je présenterai également quelques applications de notre modèle dans des domaines tels que la santé et l’industrie. Enfin, si le temps le permet, je présenterai deux éléments essentiels de notre modèle de raisonnement causal que nous avons récemment développés : l’Effet Causal Variationnel Facile Probabiliste (PEACE) et l’Effet Causal Variationnel Probabiliste (PACE) (Faghihi & Saki, 2023).

Usef Faghihi est professeur adjoint à l’Université du Québec à Trois-Rivières. Auparavant, Usef était professeur à l’Université d’Indianapolis aux États-Unis. Usef a obtenu son doctorat en Informatique Cognitive à l’UQAM. Il est ensuite allé à Memphis, aux États-Unis, pour effectuer un post-doctorat avec le professeur Stan Franklin, l’un des pionniers de l’intelligence artificielle. Ses centres d’intérêt en recherche sont les architectures cognitives et leur intégration avec les algorithmes d’apprentissage profond.

Faghihi, U., Robert, S., Poirier, P., & Barkaoui, Y. (2020). From Association to Reasoning, an Alternative to Pearl’s Causal Reasoning. In Proceedings of AAAI-FLAIRS 2020. North-Miami-Beach (Florida).

Faghihi, U., & Saki, A. (2023). Probabilistic Variational Causal Effect as A new Theory for Causal Reasoning. arXiv preprint arXiv:2208.06269.

Kalantarpour, C., Faghihi, U., Khelifi, E., & Roucaut, F.-X. (2023). Clinical Grade Prediction of Therapeutic Dosage for Electroconvulsive Therapy (ECT) Based on Patient’s Pre-Ictal EEG Using Fuzzy Causal Transformers. Paper presented at the International Conference on Electrical, Computer, Communications and Mechatronics Engineering, ICECCME 2023, Tenerife, Canary Islands, Spain.

Zaheer, M., Guruganesh, G., Dubey, K. A., Ainslie, J., Alberti, C., Ontanon, S., . . . Yang, L. (2020). Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33, 17283-17297.

The Debate Over “Understanding” in AI’s Large Language Models (Mitchell)

Melanie Mitchell , Santa Fe Institute, 19-Oct

VIDEO

ABSTRACT: I will survey a current, heated debate in the AI research community on whether large pre-trained language models can be said — in any important sense — to “understand” language and the physical and social situations language encodes. I will describe arguments that have been made for and against such understanding, and, more generally, will discuss what methods can be used to fairly evaluate understanding and intelligence in AI systems. I will conclude with key questions for the broader sciences of intelligence that have arisen in light of these discussions.

Melanie Mitchell is Professor at the Santa Fe Institute. Her current research focuses on conceptual abstraction and analogy-making in artificial intelligence systems. Melanie is the author or editor of six books and numerous scholarly papers in the fields of artificial intelligence, cognitive science, and complex systems. Her 2009 book Complexity: A Guided Tour (Oxford University Press) won the 2010 Phi Beta Kappa Science Book Award, and her 2019 book Artificial Intelligence: A Guide for Thinking Humans (Farrar, Straus, and Giroux) is a finalist for the 2023 Cosmos Prize for Scientific Writing.

Mitchell, M. (2023). How do we know how smart AI systems are? Science, 381(6654), adj5957.

Mitchell, M., & Krakauer, D. C. (2023). The debate over understanding in AI’s large language models. Proceedings of the National Academy of Sciences, 120(13), e2215907120.

Millhouse, T., Moses, M., & Mitchell, M. (2022). Embodied, Situated, and Grounded Intelligence: Implications for AI. arXiv preprint arXiv:2210.13589.

Rethinking the Physical Symbol Systems Hypothesis (Rosenbloom)

Paul Rosenbloom , Computer Science, USC, 12-Oct 2023

VIDEO

ABSTRACT: It is now more than a half-century since the Physical Symbol Systems Hypothesis (PSSH) was first articulated as an empirical hypothesis. More recent evidence from work with neural networks and cognitive architectures has weakened it, but it has not yet been replaced in any satisfactory manner. Based on a rethinking of the nature of computational symbols – as atoms or placeholders – and thus also of the systems in which they participate, a hybrid approach is introduced that responds to these challenges while also helping to bridge the gap between symbolic and neural approaches, resulting in two new hypotheses, one – the Hybrid Symbol Systems Hypothesis (HSSH) – that is to replace the PSSH and the other focused more directly on cognitive architectures. This overall approach has been inspired by how hybrid symbol systems are central in the Common Model of Cognition and the Sigma cognitive architectures, both of which will be introduced – along with the general notion of a cognitive architecture – via “flashbacks” during the presentation.

Paul S. Rosenbloom is Professor Emeritus of Computer Science in the Viterbi School of Engineering at the University of Southern California (USC). His research has focused on cognitive architectures (models of the fixed structures and processes that together yield a mind), such as Soar and Sigma; the Common Model of Cognition (a partial consensus about the structure of a human-like mind); dichotomic maps (structuring the space of technologies underlying AI and cognitive science); “essential” definitions of key concepts in AI and cognitive science (such as intelligence, theories, symbols, and architectures); and the relational model of computing as a great scientific domain (akin to the physical, life and social sciences).

Rosenbloom, P. S. (2023). Rethinking the Physical Symbol Systems Hypothesis. In Proceedings of the 16^th International Conference on Artificial General Intelligence (pp. 207-216). Cham, Switzerland: Springer.

Laird, J. E., Lebiere, C. & Rosenbloom, P. S. (2017). A Standard Model of the Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics. AI Magazine, 38, 13-26.

Rosenbloom, P. S., Demski, A. & Ustun, V. (2016). The Sigma cognitive architecture and system: Towards functionally elegant grand unification. Journal of Artificial General Intelligence, 7, 1-103.

Rosenbloom, P. S., Demski, A. & Ustun, V. (2016). Rethinking Sigma’s graphical architecture: An extension to neural networks. Proceedings of the 9^th Conference on Artificial General Intelligence (pp. 84-94).

Symbols and Grounding in LLMs (Pavlick)

Ellie Pavlick , Computer Science, Brown, 05-Oct 2023

VIDEO

ABSTRACT: Large language models (LLMs) appear to exhibit human-level abilities on a range of tasks, yet they are notoriously considered to be “black boxes”, and little is known about the internal representations and mechanisms that underlie their behavior. This talk will discuss recent work which seeks to illuminate the processing that takes place under the hood. I will focus in particular on questions related to LLM’s ability to represent abstract, compositional, and content-independent operations of the type assumed to be necessary for advanced cognitive functioning in humans.

Ellie Pavlick is an Assistant Professor of Computer Science at Brown University. She received her PhD from University of Pennsylvania in 2017, where her focus was on paraphrasing and lexical semantics. Ellie’s research is on cognitively-inspired approaches to language acquisition, focusing on grounded language learning and on the emergence of structure (or lack thereof) in neural language models. Ellie leads the language understanding and representation (LUNAR) lab, which collaborates with Brown’s Robotics and Visual Computing labs and with the Department of Cognitive, Linguistic, and Psychological Sciences.

Tenney, Ian, Dipanjan Das, and Ellie Pavlick. “BERT Rediscovers the Classical NLP Pipeline.” Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019. https://arxiv.org/pdf/1905.05950.pdf

Pavlick, Ellie. “Symbols and grounding in large language models.” Philosophical Transactions of the Royal Society A 381.2251 (2023): 20220041. https://royalsocietypublishing.org/doi/pdf/10.1098/rsta.2022.0041

Lepori, Michael A., Thomas Serre, and Ellie Pavlick. “Break it down: evidence for structural compositionality in neural networks.” arXiv preprint arXiv:2301.10884 (2023). https://arxiv.org/pdf/2301.10884.pdf

Merullo, Jack, Carsten Eickhoff, and Ellie Pavlick. “Language Models Implement Simple Word2Vec-style Vector Arithmetic.” arXiv preprint arXiv:2305.16130 (2023). https://arxiv.org/pdf/2305.16130.pdf

Grounding in Large Language Models: Functional Ontologies for AI (Mollo)

Dimitri Coelho Mollo. Philosophy of AI, Umeå University, 21 sept 2023

VIDEO

ABSTRACT: I will describe joint work with Raphaël Millière, arguing that language grounding (but not language understanding) is possible in some current Large Language Models (LLMs). This does not mean, h that the way language grounding works in LLMs is similar to how grounding works in humans. The differences open up two options: narrowing the notion of grounding to only the phenomenon in humans; or pluralism about grounding, extending the notion more broadly to systems that fulfil the requirements for intrinsic content. Pluralism invites applying recent work in comparative and cognitive psychology to AI, especially the search for appropriate ontologies to account for cognition and intelligence. This can help us better understand the capabilities and limitations of current AI systems, as well as potential ways forward.

Dimitri Coelho Mollo is Assistant Professor with focus in Philosophy of Artificial Intelligence at the Department of Historical, Philosophical and Religious Studies, at Umeå University, Sweden, and focus area coordinator at TAIGA (Centre for Transdisciplinary AI), for the area ‘Understanding and Explaining Artificial Intelligence’. I am also an external Principal Investigator at the Science of Intelligence Cluster, in Berlin, Germany. My research focuses on foundational and epistemic questions within artificial intelligence and cognitive science, looking for ways to improve our understanding of mind, cognition, and intelligence in biological and artificial systems. My work often intersects issues in Ethics of Artificial Intelligence, Philosophy of Computing, and Philosophy of Biology.

Coelho Mollo and Millière (2023), The Vector Grounding Problem

Francken, Slors, Craver (2022), Cognitive ontology and the search for neural mechanisms: three foundational problems

Conscious processing, inductive biases and generalization in deep learning (Bengio)

Yoshua Bengio, MILA, Université de Montréal 17 feb 2023

VIDEO

Abstract: Humans are very good at “out-of-distribution” generalization (compared to current AI systems). It would be useful to determine the inductive biases they exploit and translate them into machine-language architectures, training frameworks and experiments. I will discuss several of these hypothesized inductive biases. Many exploit notions in causality and connect abstractions in representation learning (perception and interpretation) with reinforcement learning (abstract actions). Systematic generalizations may arise from efficient factorization of knowledge into recomposable pieces. This is partly related to symbolic AI (aas seen in the errors and limitations of reasoning in humans, as well as in our ability to learn to do this at scale, with distributed representations and efficient search). Sparsity of the causal graph and locality of interventions — observable in the structure of sentences — may reduce the computational complexity of both inference (including planning) and learning. This may be why evolution incorporated this as “consciousness.” I will also suggest some open research questions to stimulate further research and collaborations.

Yoshua Bengio, Professor, University of Montreal, founder and scientific director of Mila – Institut québécois d’AI, and co-director CIFAR’s Machine Learning, Biological Learning program as a Senior Fellow. He also serves as scientific director of IVADO.

Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., … & VanRullen, R. (2023). Consciousness in artificial intelligence: insights from the science of consciousness. arXiv preprint arXiv:2308.08708.

Zador, A., Escola, S., Richards, B., Ölveczky, B., Bengio, Y., Boahen, K., … & Tsao, D. (2023). Catalyzing next-generation artificial intelligence through neuroai. Nature Communications, 14(1), 1597.

Active inference and artificial curiosity (Friston)

Karl Friston, UCL, 8 December, 2022

VIDEO

Abstract: This talk offers a formal account of insight and learning in terms of active (Bayesian) inference. It deals with the dual problem of inferring states of the world and learning its statistical structure. In contrast to current trends in machine learning (e.g., deep learning), we focus on how agents learn from a small number of ambiguous outcomes to form insight. I will use simulations of abstract rule-learning and approximate Bayesian inference to show that minimising (expected) free energy leads to active sampling of novel contingencies. This epistemic, curiosity-directed behaviour closes `explanatory gaps’ in knowledge about the causal structure of the world, thereby reducing ignorance, in addition to resolving uncertainty about states of the known world. We then move from inference to model selection or structure learning to show how abductive processes emerge when agents test plausible hypotheses about symmetries in their generative models of the world. The ensuing Bayesian model reduction evokes mechanisms associated with sleep and has all the hallmarks of aha moments.

Karl Friston, Professor, Institute of Neurology, UCL, models functional integration in the human brain and the principles that underlie neuronal interactions. His main contribution to theoretical neurobiology is a free-energy principle for action and perception (active inference).

Medrano, J., Friston, K., & Zeidman, P. (2024). Linking fast and slow: the case for generative models. Network Neuroscience, 8(1), 24-43.

Pezzulo, G., Parr, T., & Friston, K. (2024). Active inference as a theory of sentient behavior. Biological Psychology, 108741.

Grounded Language Learning in Virtual Environments (Clark)

Stephen Clark, U Cambridge and Quantinuum, 19 November, 2020

VIDEO

Abstract: Natural Language Processing is currently dominated by the application of text-based language models such as GPT. One feature of these models is that they rely entirely on the statistics of text, without making any connection to the world, which raises the interesting question of whether such models could ever properly “understand” the language. One these models can be grounded is to connect them to images or videos, for example by conditioning the language models on visual input and using them for captioning. In this talk I extend the grounding idea to a simulated virtual world: an environment which an agent can perceive, explore and interact with. A neural-network-based agent is trained — using distributed deep reinforcement learning — to associate words and phrases with things that it learns to see and do in the virtual world.The world is 3D, built in Unity, and contains recognisable objects, including some from the ShapeNet repository of assets. One of the difficulties in training such networks is that they have a tendency to overfit to their training data.We demonstrate how the interactive, first-person perspective of an agent helps it to generalize to out-of-distribution settings. Training the agents typically requires a huge number of training examples. We show how meta-learning can be used to teach the agents to bind words to objects in a one-shot setting. The agent is able to combine its knowledge of words obtained one-shot with its stable knowledge of word meanings learned over many episodes, providing a form of grounded language learning which is both “fast and slow”. Joint work with Felix Hill.

Clark, S., Lerchner, A., von Glehn, T., Tieleman, O., Tanburn, R., Dashevskiy, M., & Bosnjak, M. (2021). Formalising Concepts as Grounded Abstractions. arXiv preprint arXiv:2101.05125.

Tull, S., Shaikh, R. A., Zemljic, S. S., & Clark, S. (2023). From Conceptual Spaces to Quantum Concepts: Formalising and Learning Structured Conceptual Models. arXiv preprint arXiv:2401.08585.