Genetics

Open the episode, then seek to 0:00.

Marblestone proposes specialised learning objectives and architectures, then discusses prediction as a possible description of cortical computation.

Brain learning

Research note: A framework for investigation is not a decoded learning rule →

Open the episode, then seek to 5:00.

Marblestone considers how inherited circuitry could connect bodily signals and social concerns to learned representations.

Brain learning

Open the episode, then seek to 10:00.

Marblestone describes proposed mechanisms for assigning learned significance to abstract concepts and emphasises the uncertainty of these neuroscience models.

Brain learning

Open the episode, then seek to 15:00.

Marblestone discusses possible learning mechanisms and cites visual-cortex models as an example of biological constraints improving artificial systems.

Neuroscience and AI

Research note: A biological component can help on a defined benchmark →

Open the episode, then seek to 22:20.

Marblestone considers sampling, learned inference and what copying digital systems might change about learning.

Neuroscience and AI

Open the episode, then seek to 26:40.

Patel and Marblestone discuss how inherited information might specify architectures and learning objectives without encoding all acquired knowledge.

Genome and brain

Open the episode, then seek to 30:15.

Marblestone describes molecular cell-type diversity and discusses what it could reveal about specialised inherited brain structures.

Genome and brain

Research note: Cell diversity does not directly identify a behavioural objective →

Open the episode, then seek to 35:00.

Marblestone considers evolutionary changes in cortex, social learning and the relationships among neuron numbers, wiring and cognition.

Brain evolution

Open the episode, then seek to 39:50.

Marblestone discusses reinforcement learning in insects and contrasts learned associations with specialised circuitry for biologically relevant signals.

Brain evolution

Open the episode, then seek to 42:43.

Marblestone contrasts cached action values with planning and connects dopamine responses to reward-prediction errors.

Reinforcement learning

Research note: A prediction-error model has a specific explanatory role →

Open the episode, then seek to 46:50.

The speakers discuss model-based planning and cultural knowledge as a possible analogy for learned shortcuts.

Reinforcement learning

Open the episode, then seek to 50:32.

Marblestone compares biological and digital constraints, including energy use, copying, memory and the complexity of cellular machinery.

Neuroscience and AI

Open the episode, then seek to 55:00.

Marblestone uses cerebellar-cell timing to question simple neural models, then distinguishes intelligence from alignment with human values.

Cellular computation

Research note: One timing experiment constrains a model of memory →

Open the episode, then seek to 59:52.

Marblestone compares computational vocabulary, detailed simulation and empirical model testing as complementary approaches to neuroscience.

Brain learning

Open the episode, then seek to 1:04:00.

Marblestone considers what wiring diagrams could reveal about architectures, learning rules and acquired circuit details.

Connectomics

Open the episode, then seek to 1:08:10.

Marblestone offers uncertain timelines and discusses proposed cost reductions and funding needs for large-scale brain mapping.

Connectomics

Open the episode, then seek to 1:13:10.

Marblestone discusses molecularly annotated optical maps, their functional limits and the need for technology development and funding.

Connectomics

Research note: Imaging progress is a step toward a circuit map →

Open the episode, then seek to 1:17:30.

Marblestone discusses organisational challenges in neuroscience and possible uses of brain activity as additional training information for AI.

Scientific infrastructure

Open the episode, then seek to 1:21:40.

Marblestone considers learning from neural activity before turning to proof systems, verifiable mathematical tasks and conjecture generation.

Neuroscience and AI

Open the episode, then seek to 1:26:10.

Marblestone discusses automated mathematics and explains why a verified result still depends on choosing an appropriate specification.

Formal verification

Research note: A proof guarantees the property that was specified →

Open the episode, then seek to 1:30:45.

Marblestone discusses the scope of software guarantees and considers how automated mathematical tools could change research practice.

Formal verification

Open the episode, then seek to 1:35:23.

Marblestone describes ambitions for provable AI systems, then discusses structured representations and unresolved questions about cognitive models.

Formal verification

Open the episode, then seek to 1:39:57.

Marblestone acknowledges uncertainty about consciousness and considers memory consolidation, replay and possible biological analogues of attention.

Brain learning

Open the episode, then seek to 1:44:54.

Marblestone describes the Gap Map’s proposed research capabilities and discusses which kinds of projects might suit different institutions.

Scientific infrastructure

Research note: A research map is an invitation to investigate →