,
7 tweets,
3 min read
Read on Twitter
We're happy to announce that our work (with @satohirotajima, @Nisheet9, and @pouget_alex) on the optimal policy for multi-alternative decisions is finally out: nature.com/articles/s4159… (ReadCube link: rdcu.be/bM2Ej). 1/7
We asked how one would optimally trade off the speed and accuracy of a decision when having to decide between multiple alternatives. 2/7 
Even in simple setups, decisions among N alternatives require curved, time-dependent decision boundaries in the N-dimensional space where evidence is accumulated over time. This makes them more complex than a simple scaling-up of decisions among two alternatives. 3/7
A neural network with inhibitory cross-talk between different accumulators and an urgency signal implements the optimal policy with simple decision boundaries on individual accumulators - compatible with how cortex appears to accumulate such evidence and trigger choices. 4/7
Inhibition, which is frequently ascribed to suboptimalities, in this case turns out to be crucial. The network captures several key features of such decisions, like Hick's law. 5/7
For normalized inputs and accumulator noise, it additionally replicates several 'irrational' behaviors, like the similarity effect, and the violations of both the independence of irrelevant alternatives as well as the regularity principle. 6/7
For more details, check out the paper! We dedicate this work to Satohiro Tajima (@satohirotajima), who was the driving force behind most of its ideas, but tragically passed away in August 2017. 7/7
