RTify: Aligning Deep Neural Networks with Human Behavioral Decisions.

Area of Science:

• Computational neuroscience
• Cognitive science
• Artificial intelligence

Background:

• Current neural network models of primate vision often overlook the dynamic nature of perceptual decisions, focusing primarily on behavioral accuracy.
• Modeling the temporal dynamics of human choices, such as reaction times, is crucial for understanding visual perception.

Purpose of the Study:

• To introduce a novel computational framework for modeling the dynamics of human behavioral choices in visual tasks.
• To align the temporal dynamics of recurrent neural networks (RNNs) with human reaction times (RTs).
• To optimize an ideal-observer RNN model for a speed-accuracy tradeoff without human data.

Main Methods:

• Developed an approximation to constrain RNN time steps based on human RTs.
• Evaluated the framework against psychophysics experiments.
• Trained a deep learning implementation of the Wong-Wang decision-making model integrated with a convolutional neural network (CNN).

Main Results:

• The framework successfully aligns RNN temporal dynamics with human RTs.
• The optimized ideal-observer RNN model demonstrated good accounting for human RT data.
• The integrated CNN-Wong-Wang model showed effectiveness with artificial and natural image stimuli.

Conclusions:

• The novel framework effectively aligns current vision models with human behavior.
• This approach advances the development of integrated models of human vision.
• The method allows for optimizing decision-making models for speed-accuracy tradeoffs.