Unveiling the Secrets of Dopamine with Artificial Intelligence by DeepMind

2020-01-16

Demis Hassabis founded DeepMind with the goal of reimagining intelligence itself to tackle some of the world's most challenging problems. In this week's issue of Nature, DeepMind published a paper that addresses a daunting challenge in the field of biomedicine. The paper, authored by DeepMind's neuroscience team, proposes that the development of artificial intelligence research can serve as a framework for understanding how the brain learns. The research was conducted after DeepMind applied artificial intelligence to predict acute kidney injury (AKI) and challenging game environments such as Go, shogi, chess, dozens of Atari games, and Activision Blizzard's StarCraft II. In a paper on dopamine, researchers from DeepMind and Harvard University investigated whether the brain represents future potential rewards as probability distributions rather than single average values - a mathematical function that provides the probability of different outcomes occurring. They found evidence of "distributional reinforcement learning" from recordings in the ventral tegmental area. Recently, DeepMind shifted its focus to rational mechanisms and developed synthetic neural networks capable of applying human-like reasoning skills and logic to problem-solving. In 2018, DeepMind researchers conducted an experiment that demonstrated that the prefrontal cortex does not rely on changes in synaptic weights to learn rule structures, as previously believed, but instead uses direct encoding in dopamine. The researchers extended the canonical reward prediction error theory of dopamine using distributional reinforcement learning. Dopamine cells change their firing rates to represent prediction errors, meaning that prediction errors should be zero when rewards are received, indicating the size of the predicted cell. Taking this into account, the researchers identified reversal points for each cell - the reward size at which dopamine cells do not change their firing rates - and compared them to see if there were any differences. In the final experiment, the researchers attempted to decode reward distributions from the firing rates of dopamine cells. The results were successful: through inference, they successfully reconstructed a distribution that matched the actual allocation of rewards in a mouse experimental task.

RECENT AI NEWS

RECENT AI TOOLS