Understanding Higher-Level Intelligence from AI, Psychology, and Neuroscience Perspectives
2024-06-24..2024-06-28
Berkeley, United States
who it's for
cognitive scientists, ai researchers, neuroscientists, developmental psychologists, machine learning engineers, multimodal researchers, comparative cognition researchers
topics
language models as world models; REMEDI representation editing; neuronal assembly sequence models; NEMO cortical assembly model; kid-inspired visual analogies (KiVA); SAYCam and BabyView egocentric corpora; multimodal grounding and contrastive vision; benchmarks for social and physics cognition; training from continuous video streams; passive learning for causal intervention; multi-agent learning and sociality; consciousness and metacognition in AI
| title | author | length |
|---|---|---|
| Language Models as World Models? | Jacob Andreas | 46m |
Evidence shows transformer LMs store rudimentary, linearly decodable representations of entities and dynamic state: linear probes recover properties from… | ||
| Computation with sequences of neuronal assemblies | Max Dabagia | 12m |
A concrete claim: a NEMO-style network with random connectivity, k-winners-take-all inhibition and Hebbian plasticity can form assemblies that encode symbols… | ||
| Bridging the data gap between LLMs and children | Michael Frank | 48m |
Claim: LLMs need 3–6 orders of magnitude more language tokens than children (child upper bound ~60M words; GPT-3 ~500B, Chinchilla ~1T, GPT‑4 reports ~15T)… | ||
| How grounded are ungrounded LLM models | Trevor Darrell | 49m |
Finds that strong text-only embeddings often remove the added benefit of pixels on parsing benchmarks, shows a pipeline where an LLM generates scene graph… | ||
| What neural machinery is needed for language acquisition (Virtual Talk) | Christos Papadimitriou | 01h 13m |
A minimal neural model called NEMO postulates cortical assemblies, discrete timesteps, random sparse synapses, Hebbian updates and long-range inhibitory… | ||
| KiVA: Kid-inspired Visual Analogies for Large Multimodal Models | Eunice Yiu | 20m |
Introduces KiVA, a kid‑inspired benchmark with ~3,000 stimuli of common 3D objects and simple transformations, tested on 40 adults and 30 children (ages 3–7)… | ||
| A comparative cognition approach to evaluating AI capabilities | Lucy Cheke | 45m |
A comparative cognition pipeline treats AI like another species: a Unity-based Animal-AI environment hosts ~300 adapted animal-cognition tasks for a 2019… | ||
| Lightning Talk: The Perception Test | Viorica Patraucean | 21m |
DeepMind builds a new benchmark and films 11,600 videos with 100+ crowdworkers, rich annotations (VQA, object and point tracks, action and sound segments) to… | ||
| 10 AIs or 10B AIs | Joao Carreira | 44m |
Proposes training from a single continuous video stream (concatenating ~10 walking‑tour videos of ~1h40 each, ~700 object classes per video) using a DINO… | ||
| Linguistic Interaction in Humans and Machines | Yoav Artzi | 51m |
CerealBar is a procedurally generated leader–follower game used to train instruction generation by deploying, observing user executions, aggregating outcomes… | ||
| What can be passively learned about causality? | Andrew Lampinen | 56m |
Argues that models trained by behavioral cloning on expert interventional data can learn actionable causal strategies: a simulated agent with a five‑variable… | ||
| Lightning Talks | Mirabel Reid, Paul Liang | 01h 01m |
Short study defines a five-level hierarchical scale for algorithm understanding and runs a five-question survey comparing students to GPT-3.5/4/4o on two… | ||
| Space is a latent sequence: A theory of hippocampus and PFC | Dileep George | 01h 22m |
Claim is that space is a sequence and hippocampal phenomena arise from sequence learning via CSCG clone-structured cognitive graphs, an overcomplete HMM with… | ||
| Discussion | Jacob Andreas | 43m |
Panelists argue that lots of capabilities emerge from models trained on code+text, with concrete examples: variable binding and long-range reference appearing… | ||
| Educability (Virtual Talk) | Leslie Valiant | 54m |
Proposes "educability" as a concrete, computational capability composed of three parts: learning from examples (standard PAC-style generalization with error… | ||
| Cultural Evolution | Alison Gopnik | 55m |
Argument claims that LLMs as cultural technologies — they compress and retrieve centuries of human text and pictures rather than act as independent agents —… | ||
| Social Behavior Prediction from Video Observations | Shiry Ginosar | 41m |
Proposes predicting 3D human behavior by regressing off-the-shelf 3D parametric body models and using a discrete motion vocabulary learned with a VQ‑VAE motion… | ||
| Retrospective on multi-agent learning as a path to intelligence | Igor Mordatch | 38m |
Large-scale RL runs on a 2D physics simulator with transformer policies and LiDAR-style observations produce staged behaviors: random motion, predator–prey by… | ||
| Emerging interactions in humans and machines: from Social Physiology to Social Neuro-AI | Guillaume Dumas | 45m |
Claim that inter-brain synchrony reflects more than coincidental similarity and can causally reshape social behavior, supported by multi-brain hyperscanning… | ||
| The Development of Sociality in Language and Thought | Justine Cassell | 59m |
Claim: everyday rapport mechanisms (hedges, turn-taking, pitch/speech-rate alignment) materially change task outcomes. Methods: dense video annotation with… | ||
| AI and Emotions: opportunities and challenges (Virtual Talk) | Àgata Lapedriza | 37m |
Project started in 2016–17 builds a 23,000‑image corpus with 34 annotators and shows context plus body improves recognition over face alone; models use… | ||
| What does it all mean anyway? Decoding ape communication. (Remote from the Ugandan forest) | Catherine Hobaiter | 34m |
Ape gesture is intentional and language-like with over 100 gesture types and a decoded meaning method that uses recipient response to stop signaling. A… | ||
| Theoretic analyses of internal model dynamics in the brain | David Chalmers | 32m |
Argues that canonical internal world models are unnecessary and often suboptimal because time is the scarce resource for biological agents and chronologically… | ||
| AI Consciousness is Inevitable: A Theoretical Computer Science Perspective | Lenore Blum | 01h 14m |
Claims a Conscious Turing Machine that is a simple, finite formal machine with STM holding one chunk, LTM of >10 million processors, an up-tree/down-tree… | ||
| On the Hard Problem of Pain | Manuel Blum | 59m |
Argues that agony requires two specific mechanisms: two generators of pain — a global broadcast that forces processors to interrupt tasks (attention) and a… | ||
| How Can We Detect Consciousness in Infants? | Claudia Passos Ferreira | 42m |
Argument claims newborns and late-term fetuses already show neural and behavioral signatures consistent with conscious perception. Evidence uses adapted… | ||
| Quantitative research programs relating metacognition and consciousness | Megan Peters | 48m |
Argues for a measurable “type-two psychophysics” by forcing dissociations between task accuracy and confidence using visual masking, TMS, ECoG and large… | ||
| From Simulated Subjectivity to Collective Consciousness in Large Language Models | James Evans | 40m |
Claims that large language models can simulate individual and collective subjectivity are backed by concrete methods and numbers: a hypergraph-embedding… | ||
| Coin-Flipping in the Brain: Probabilistic Computation and Learning in the Assembly Model | Santosh Vempala | 43m |
Model uses a directed Erdős–Rényi connectome with constant brain areas of n neurons where k out of n fire (k ≈ √n) via winner‑take‑all inhibition, Hebbian… | ||
| Consciousness and Large Language Models | David Chalmers | 53m |
Argues current transformer LLMs lack several plausible consciousness X‑factors—robust self‑models, recurrence, a global workspace and unified agency—though… | ||
| nothing matches | ||