MinedDeep

---

Papers

• A Clinically Applicable Approach to Continuous Prediction of Future Acute Kidney Injury
• A Distributional Code for Value in Dopamine-Based Reinforcement Learning
• A General Reinforcement Learning Algorithm That Masters Chess, Shogi, and Go Through Self-Play
• Advancing Mathematics by Guiding Human Intuition with AI
• Applying and Improving AlphaFold at CASP14
• Big-Loop Recurrence Within the Hippocampal System Supports Integration of Information Across Episodes
• Clinically Applicable Deep Learning for Diagnosis and Referral in Retinal Disease
• Computations Underlying Social Hierarchy Learning
• Decoding Individual Episodic Memory Traces in the Human Hippocampus
• Decoding Neuronal Ensembles in the Human Hippocampus
• Deconstructing Episodic Memory with Construction
• Grandmaster Level in StarCraft II Using Multi-Agent Reinforcement Learning
• Highly Accurate Protein Structure Prediction for the Human Proteome
• Highly Accurate Protein Structure Prediction with AlphaFold
• Human-Level Control Through Deep Reinforcement Learning
• Human-Level Performance in First-Person Multiplayer Games with Population-Based Deep Reinforcement Learning
• Hybrid Computing Using a Neural Network with Dynamic External Memory
• Imagine All the People: How the Brain Creates and Uses Personality Models
• Improved Protein Structure Prediction Using Potentials from Deep Learning
• Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model
• Mastering the Game of Go with Deep Neural Networks and Tree Search
• Mastering the Game of Go without Human Knowledge
• Neural Mechanisms of Hierarchical Planning in a Virtual Subway Network
• Neural Scene Representation and Rendering
• Neuroscience-Inspired Artificial Intelligence
• Overcoming Catastrophic Forgetting in Neural Networks
• Patients with Hippocampal Amnesia Cannot Imagine New Experiences
• Prefrontal Cortex as a Meta-Reinforcement Learning System
• Protein Complex Prediction with AlphaFold-Multimer
• Protein Structure Predictions to Atomic Accuracy with AlphaFold
• Pushing the Frontiers of Density Functionals by Solving the Fractional Electron Problem
• Reinforcement Learning, Fast and Slow
• Semantic Representations in the Temporal Pole Predict False Memories
• The Construction System of the Brain
• The Future of Memory: Remembering, Imagining, and the Brain
• Tracking the Emergence of Conceptual Knowledge during Human Decision Making
• Using Imagination to Understand the Neural Basis of Episodic Memory
• Vector-Based Navigation Using Grid-Like Representations in Artificial Agents
• What Learning Systems Do Intelligent Agents Need? Complementary Learning Systems Theory Updated
• When Fear Is Near: Threat Imminence Elicits Prefrontal-Periaqueductal Gray Shifts in Humans

Lectures

• Accelerating Scientific Discovery with AI (Nobel Lecture)

Interviews

• Future of AI, Simulating Reality, Physics and Video Games (Lex Fridman Podcast #475)

Essays

• Chess Match of the Century
• DeepMind CEO Demis Hassabis Urges Caution on AI
• Demis Hassabis Is Preparing for AI’s Next Chapter (TIME100)

Posts

Themes

• AI for Science
• Deep Reinforcement Learning
• Game-Playing AI
• Hippocampal Construction
• Memory and Imagination
• Neuroscience-AI Bridge
• Protein Folding
• Self-Play

Projects

• AlphaFold2
• AlphaGo
• DQN
• Hippocampus Research Programme

Periods

• PhD Period (2007–2009)
• Postdoc Period (2009–2010)
• Early DeepMind (2010–2015)
• DeepMind Ascent (2015–2018)
• AlphaFold Era (2018–2022)
• Post-AlphaFold (2022–Present)

Collaborators

• David Silver
• Eleanor A. Maguire
• John Jumper

Venues

• Nature
• Neuron
• Science
• Current Biology
• Nobel Prize Lecture
• Other Venues

Claims

• Hippocampus as Construction System
• Learnable Nature Conjecture
• Neuroscience-AI Bidirectional Bridge
• Self-Play Sufficiency

Gaps

• Missing Paper: Learnable Nature Conjecture

Intersections

Epistemic tags: Grounded = both parents make specific conflicting claims, gap directly verifiable. Extrapolative = one parent’s mechanism extended to new domain. Conjectural = relies on mechanisms from other intersections, not directly from corpus papers.

Priority ranking (top 5)

1. ⬥ Self-Play Discovers Its Own Consolidation Conjectural — Decisive experiment: does meta-self-play discover CLS without biological priors?
2. ⬥ Memory as Query, Not Store Conjectural — Paradigm test: are hippocampal traces stable or do they shift with consolidation?
3. ⬥ EWC × Experience Replay Grounded — Trivially testable gap: same lab, same problem, never connected
4. Slow RL Grounded — Clean protocol: parametric distance + RL model fitting
5. ⬥ Social Hierarchy × Self-Play Grounded — Direct RL experiment: rank representations in multi-agent play

First-order (pairwise)

• Hippocampal Construction × Self-Play Extrapolative — Imagination and self-play both generate novel outputs by recombination
• Experience Replay × Hippocampal Replay Grounded — The most glaring silence: DQN’s replay and hippocampal SWR replay, same problem, never connected
• Learnable Nature × Hippocampal Construction Extrapolative — Both are structure prediction from compressed representations
• Meta-RL × Self-Play Grounded — A meta-learner that discovers its learning algorithm through self-play
• Slow RL × AlphaGo Zero Dynamics Grounded — The dual-system split may be emergent rather than designed
• Big-Loop Recurrence × Attention Grounded — Iterative integration as fixed-point iteration x_{t+1} = f(x_t)
• GQN × Hippocampal Construction Grounded — Scene representation + novel viewpoint synthesis; first/third-person as query variation
• Grid Cells × Self-Play Grounded — Do grid-like representations emerge differently under non-stationary self-play?
• EWC × Experience Replay Grounded — The second glaring silence: two biological forgetting solutions, same lab, never connected
• DNC × Complementary Learning Systems Grounded — DNC’s external memory is an incomplete hippocampus; CLS provides the missing consolidation
• Distributional RL × Hippocampal Construction Extrapolative — Imagined scenarios should produce distributional prediction errors
• Semantic False Memories × Construction Grounded — False memories are over-construction; the error term of the system
• Social Hierarchy × Self-Play Grounded — Multi-agent self-play should produce internal rank representations
• Density Functionals × Learnable Nature Grounded — The strongest evidence for learnable nature, never connected to the conjecture
• Slow RL Grounded — vmPFC→PAG shift is a parametric fast/slow transition
• Conceptual Emergence × Big-Loop Recurrence Extrapolative — Conceptual abstraction as phase transition from associations to categories
• MuZero × Construction Extrapolative — Both learn internal models without rules; MuZero predicts, construction generates
• Personality Models × Meta-RL Extrapolative — Personality knowledge conditions the learning algorithm

Second-order (intersection × intersection)

• Sleep Cycle Extrapolative — Construct during wake (self-play), consolidate during sleep (replay)
• Slow Extrapolative — The brain’s dual-system architecture may be a discovered optimum
• Episodic Consolidation as Structure Inference Extrapolative — Memory consolidation is big-loop attention over constructed scenes
• Meta-Replay Extrapolative — Replay past learning episodes to improve the learning algorithm itself
• Semantic Split Extrapolative — Episodic and semantic memory as emergent outputs of a single construction system
• Mathematics of Construction Extrapolative — Use AI-guided mathematical exploration to formalise episodic construction
• Mental Time Travel as Viewpoint Synthesis Extrapolative — Recall is re-rendering a cross-episode scene from a different query viewpoint
• Emergent Cognitive Maps in Adversarial Environments Extrapolative — Fast/slow split instantiated as spatial scale split over grid-like codes
• Error Term of Learnable Construction Extrapolative — False memories reveal where learned construction grammar breaks down
• Social Meta-Self-Play Extrapolative — Rank-conditioned meta-learning; different algorithms for different opponents
• AI-Discovered Laws of Mind Extrapolative — Discover correctness properties in cognitive architecture

Third-order (convergences)

• Convergences Conjectural — Four unified theories: AlphaFold for Imagination, Developmental Stack, Social Construction Engine, Self-Diagnostic Discovery Machine

Feedback loops (intersection × intersection, recursive)

• Self-Play Discovers Its Own Consolidation Conjectural — Meta-self-play discovers CLS without biological priors
• Memory as Query, Not Store Conjectural — Episodic “traces” are query parameters into scene grammars
• Self-Correcting Construction Conjectural — AI-discovered laws predict false memory locations; errors bootstrap theory
• Spatially Structured Mental Time Travel Conjectural — Recall is navigation over grid-like constructed scenes