Isshori Gurung¹, Matthew Parrella¹, Nicholas Crotty¹, Michael Grubb¹; ¹Trinity College

An abrupt onset in the visual periphery (a task-irrelevant “disk”) elicits the reflexive allocation of spatial attention without observable eye movements. Transient, peripheral disks also modulate pupil size, which is observable with an eye-tracker. It can be difficult, however, to disentangle the pupillary response to the disk, from the pupillary response to subsequently presented stimuli (ie, those needed to observe the impact of attention). In the service of an ongoing investigation about the interaction between expectation and exogenous attention, we used a 2000ms stimulus-onset-asynchrony (SOA) between a briefly-presented peripheral disk and a simple visual target to isolate the pupillary response to the disk alone. Across ~10,000 trials, a disk was presented half the time. For each trial, we collected 1000 samples of pupil area during the SOA and recorded whether a disk was presented or omitted. Motivated by the popularity of deep learning, here we asked: Can a fully connected neural network (NN) categorize a trial’s disk status (presented or omitted) using the pupil timeseries alone? We built a NN that takes in trial-level pupil data and passes them through three fully connected hidden layers, each with 512 neurons. The NN returns a single value indicating its prediction of whether a disk was presented or omitted on that trial. The network underwent a 3/4:1/4, training:validation split with an early stopping procedure. After 18 epochs, the NN predicted disk status well above chance (72% vs. 50%) for the ~2500 validation trials. In sum, a NN can reliably predict the presentation/omission of a task-irrelevant peripheral disk using trial-level pupil data alone. This project demonstrates the utility of using a NN as a complementary analysis technique for pupil data, and a shareable, interactive python notebook will make our pipeline accessible beyond our lab.

Acknowledgements: NSF-2141860 CAREER Award to Michael Grubb