Neural networks can learn from their experience how to learn faster next time. When we simulate this from simple principles, several brain-like properties emerge naturally.

The brain spontaneously retrieves sets of memories in fast bursts to reason about their joint implications.

Dopamine uses a clever trick to make multifaceted predictions of the future - simultaneously predicting the worst that could happen, the best, and many possibilities in between.

Zeb Kurth-Nelson
DeepMind, UCL
zebkurthnelson@gmail.com

Wang and Kurth-Nelson et al (2018)
Ritter et al (2018)
Dasgupta et al (2019)

Kurth-Nelson et al (2016)
Liu et al (2019)
Kurth-Nelson et al (2023)
Schwartenbeck et al (2023)

Dabney and Kurth-Nelson et al (2020)
Muller et al (2024)

Contact/Info
Google Scholar Profile