*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Name: Farshad Rafiei
Dissertation Proposal Meeting
Date: Friday, December 3rd, 2021
Time: 1:30 PM
Location: https://bluejeans.com/403600291/8380
Advisor: Dobromir Rahnev, PhD (Georgia Tech)
Dissertation Committee Members:
Daniel Spieler, PhD (Georgia Tech)
Sashank Varma, PhD (Georgia Tech)
Paul Verhaeghen, PhD (Georgia Tech)
Mark Wheeler, PhD (Georgia Tech)
Title: Explaining choice, reaction time and confidence for novel images using convolutional neural networks
Abstract
Feedforward neural networks exhibit excellent object recognition performance and currently provide the best models of biological vision in neuroscience. Despite their remarkable performance in recognizing unseen images, their behavior in making decisions is different from human decision-making. Standard feedforward neural networks spend an identical number of computations (i.e., time) to process a given stimulus and always land on the same response for that stimulus. Human decisions, in contrast, occur over time and are stochastic (i.e., the same stimulus may elicit different responses on different trials). The goal of this project is to build a neural network model of human decision-making which accounts for some of the ubiquitous aspects of choice data and is able to make accurate behavioral predictions regarding an unseen image. Specifically, I aim to train a Bayesian neural network which enables sequential sampling of evidence in favor of choice options until a predefined threshold is met. We expect that the resulting model will generate its decisions over time and react to the stimulus like a human agent. More specifically, I expect that the model will: (1) predict for which images the humans will have higher accuracy, higher RT and higher confidence; (2) explain the overall distribution of RTs; (3) provide less accurate responses when decisions need to be made fast and vice versa (i.e., it will exhibit speed-accuracy trade-off); and (4) produce higher confidence for correct trials. To evaluate the model’s behavior, we will compare it with data which will be acquired from twenty healthy human subjects. They will be asked to discriminate the handwritten digits selected from the publicly available MNIST dataset with two different levels of noise/contrast (i.e., low/high) in two different conditions that emphasize speed or accuracy. We expect that the model will account for choice, reaction time and confidence under different conditions, and show behavior which is highly correlated with human choice data. The proposed study will provide a model of human decision-making that not only accounts for standard behavioral effects but is also capable of making predictions regarding novel images.
