Search Papers | Poster Sessions | All Posters
Poster A1 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Reinforcement-Based Control of Information Processing in Recurrent Neural Networks Produces Optimal Speed-Accuracy Tradeoff
Ivan Grahek1,2 (), Alekh Karkada Ashok1,2, Atsushi Kikumoto1,2, Thomas Serre1,2, Michael J. Frank1,2; 1Brown University, 2Carney Institute for Brain Science
Optimal decision-making entails not only arriving at the best choice but doing so in the most efficient way possible. Critically, humans and other animals adjust the speed and accuracy of their decisions to the demands of the current task. Recurrent neural networks (RNNs) can process noisy sequential bits of evidence and are used as models of decision-making. However, they are typically trained on input sequences of fixed length, and thus have no notion of decision time. Here, we develop an RNN with a separate controller network that adjusts the number of RNN steps taken in a decision-making task. Using reinforcement learning in the controller, this architecture optimally trades off decision time and accuracy. In this way, it aligns with normative models of human decision-making, and produces a natural notion of decision time.
Keywords: decision-making recurrent neural networks decision time reinforcement learning