DeepMind Ascent (2015–2018)

Type: period Slug: period—deepmind-ascent Sources: a-general-reinforcement-learning-algorithm-that-masters-chess-shogi-and-go—hassabis, artificial-intelligence-chess-match-of-the-century—hassabis, clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease—hassabis, computations-underlying-social-hierarchy-learning—hassabis, human-level-performance-in-3d-multiplayer-games-with-population-based-rl—hassabis, hybrid-computing-using-a-neural-network-with-dynamic-external-memory—hassabis, mastering-the-game-of-go-with-deep-neural-networks-and-tree-search—hassabis, mastering-the-game-of-go-without-human-knowledge---hassabis, neural-mechanisms-of-hierarchical-planning-in-a-virtual-subway-network—hassabis, neural-scene-representation-and-rendering—hassabis, neuroscience-inspired-artificial-intelligence—hassabis, overcoming-catastrophic-forgetting-in-neural-networks—hassabis, prefrontal-cortex-as-a-meta-reinforcement-learning-system—hassabis, semantic-representations-in-the-temporal-pole-predict-false-memories—hassabis, what-learning-systems-do-intelligent-agents-need-complementary-learning-systems-theory-updated—hassabis Last updated: 2026-05-13

---

Summary

The most prolific and intellectually diverse period in the corpus, with 15 publications across game-playing AI, neuroscience theory, clinical AI, and cognitive neuroscience. The era is anchored by three game-playing milestones — AlphaGo (2016), AlphaGo Zero (2017), and AlphaZero (2018) — but its distinctive intellectual contribution is the neuroscience-AI bridge programme: four papers (Botvinick/Kumaran/Dayan/Dolan collaborations) that explicitly connected hippocampal and prefrontal neuroscience to RL architectures, culminating in the canonical Neuron review “Neuroscience-Inspired Artificial Intelligence” (2017).

Core content

The Go sequence (2016–2018): Three papers trace a rapid evolution. AlphaGo (paper—mastering-the-game-of-go-with-deep-neural-networks-and-tree-search) combined deep neural networks with Monte Carlo tree search, supervised from human expert games, to defeat Fan Hui and Lee Sedol. AlphaGo Zero (paper—mastering-the-game-of-go-without-human-knowledge) removed human data entirely, learning from self-play alone to surpass AlphaGo in 40 days. AlphaZero (paper—a-general-reinforcement-learning-algorithm-that-masters-chess-shogi-and-go) generalised the self-play approach to chess and shogi, achieving superhuman play in all three games with a single architecture. The intellectual arc is from human-guided → self-taught → general.

The neuroscience-AI bridge (2016–2017): Four papers with the Botvinick/Kumaran/Dayan/Dolan group. The Neuron review (paper—neuroscience-inspired-artificial-intelligence) is the canonical statement, arguing that neuroscience can contribute specific, implementable ideas to AI (not just vague inspiration). Supporting papers: meta-RL in prefrontal cortex (paper—prefrontal-cortex-as-a-meta-reinforcement-learning-system), complementary learning systems updated for AI agents (paper—what-learning-systems-do-intelligent-agents-need-complementary-learning-systems-theory-updated), and hierarchical planning in the hippocampus (paper—neural-mechanisms-of-hierarchical-planning-in-a-virtual-subway-network).

Continued hippocampal research (2016): Two papers extending the PhD programme with new collaborators — social hierarchy learning (paper—computations-underlying-social-hierarchy-learning) and false memory prediction from temporal pole semantics (paper—semantic-representations-in-the-temporal-pole-predict-false-memories).

AI systems advances (2016–2017): DNC (paper—hybrid-computing-using-a-neural-network-with-dynamic-external-memory) added differentiable external memory to neural networks. Population-based RL (paper—human-level-performance-in-3d-multiplayer-games-with-population-based-rl) achieved human-level play in Quake III capture-the-flag. GQN (paper—neural-scene-representation-and-rendering) learned to render 3D scenes from novel viewpoints without explicit 3D representation — a construction-system-like approach to visual scenes.

Continual learning (2017): Elastic weight consolidation (paper—overcoming-catastrophic-forgetting-in-neural-networks) addressed the catastrophic forgetting problem by selectively slowing learning on important weights, drawing on synaptic consolidation in neuroscience.

Clinical AI (2016, 2018): Retinal disease diagnosis (paper—clinically-applicable-deep-learning-for-diagnosis-and-referral-in-retinal-disease) demonstrated that deep learning could match expert ophthalmologists on referral decisions — DeepMind’s first clinical deployment paper.

Public writing (2016): A Nature book review (paper—artificial-intelligence-chess-match-of-the-century) on the AlphaGo vs Lee Sedol match, framing it as a test of human creativity.

Connections

• Themes: theme—game-playing-AI, theme—go, theme—chess, theme—neuroscience-AI-bridge, theme—deep-RL, theme—reinforcement-learning, theme—hippocampal-construction, theme—self-play
• Projects: project—AlphaGo, project—AlphaGo-Zero, project—AlphaZero, project—DNC, project—GQN, project—hippocampus-research, project—DeepMind-general
• Collaborators: David Silver (5 papers), Timothy Lillicrap (4), Koray Kavukcuoglu (3), Dharshan Kumaran (4), Matthew Botvinick (3), Peter Dayan (3), Raymond Dolan (3), Eleanor A. Maguire (2)
• Venues: venue—Nature (5), venue—Neuron (3), venue—Science (2), venue—PNAS (2), venue—Current-Biology (1), venue—Nature-Medicine (1)
• Succeeds: period—early-deepmind — DQN’s deep RL paradigm explodes into multiple domains
• Precedes: period—alphafold-era — the self-play and generalisation ideas mature into AlphaFold

Honest Gaps

• Several 2018 papers (neural-scene-representation, vector-based-navigation) are classified as “alphafold-era” in metadata but chronologically belong here — the era boundary (2018) is fuzzy.
• The neuroscience-AI bridge papers are authored primarily by Botvinick/Kumaran/Dayan/Dolan with Hassabis as senior author — the degree of Hassabis’s direct intellectual contribution vs. lab-leadership role is unclear.
• No sources from this period document the internal debate about pursuing games vs. protein folding vs. clinical applications.
• The Lee Sedol match (March 2016) was a major cultural event but only a book review covers it — no primary account from Hassabis.
• The clinical AI paper (retinal disease) sits thematically apart from the rest of the era — no bridge paper connects it to the game-playing or neuroscience work.