Making decisions with unknown sensory reliability.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Decision Science

Background:

• Accurate decision-making requires integrating sensory input, prior knowledge, and adjusting criteria.
• Standard diffusion models for decision-making face a challenge in determining optimal integration speed and thresholds without knowing sensory reliability beforehand.

Purpose of the Study:

• To develop a Bayesian decision model that simultaneously infers choice probabilities and estimates sensory information reliability within a single trial.
• To create a non-linear diffusion to bound model with dynamically changing sensory input weights and decision thresholds.

Main Methods:

• Utilized a Bayesian decision framework to model simultaneous inference of choice probabilities and sensory reliability.
• Developed a non-linear diffusion to bound model based on noisy sensory spiking neuron responses.
• Investigated model performance in a motion discrimination task.

Main Results:

• The model demonstrates dynamic adjustments in sensory input weighting and decision thresholds, adapting to trial difficulty.
• Difficult trials show faster, less accurate decisions due to collapsing thresholds and strong early sensory impact.
• Easy trials result in slower, more accurate decisions with increasing sensory weight and thresholds over time.

Conclusions:

• The adaptive sensory weights in this model provide accurate choice probability representations, superior to standard diffusion models.
• This model can integrate unreliable cues, like priors, effectively.
• The model accounts for empirical findings in motion discrimination and is implementable via Hebbian learning in neural architectures.