Neural Computation and Decision-Making in Saccadic Eye Movements
Neural Computation of Log Likelihood in Control of Saccadic Eye Movements
Latency is related to the evaluation of the existence of a signal. Neural integration of the signal’s existence: if the threshold is reached (→), the signal exists. Changing the prior by changing how many times it goes left/right changes the latency. For low latency, we assume express saccades. The prior changes, so the slope changes, converging to the same threshold.
Neuronal Correlates of a Perceptual Decision
Decision-making is reflected in neurons. Recorded from the Middle Temporal (MT) area. Neuronal thresholds are lower than perceptual thresholds (→) using neurons gives a better estimate. A pooled signal from several neurons gives a better estimate, but if sources are not independent, pooling provides no benefit.
Glimcher Review
- MT (Middle Temporal area): neuronal rate reflects saccade probability in that direction. Electrical stimulation alters the probability of a particular saccade.
- Mutual inhibition (→) LIP (Lateral Intraparietal area): the interface between sensory and motor systems integrates the decision.
- (→) FEF (Frontal Eye Fields): integrative signals trigger a saccadic eye movement.
A gradual process of neural decision-making fires if the threshold is reached.
Representation of a Perceptual Decision in Developing Oculomotor Commands
The monkey will move depending on the coherence of the dots and the time they watch the stimulus. Electrostimulation shifts the saccade slightly.
From Grid Cells to Place Cells
Place cells discharge when in the place field. Place cells have a single firing field. Medial Entorhinal Cortex (MEC). Grid cells are universally applicable. Dendritic summation of grid cell afferents into place cells, producing Gaussian place fields.
Taxi Drivers
The posterior hippocampus is larger for taxi drivers; the anterior hippocampus is larger for the control group. Presumably, the posterior is for remembering past maps, whereas the anterior is for the current map.
Motion Planning in Visually Perceived Space
We learn in terms of visual perception. If joint space is visually represented, we make straight lines in joint space. If hand space is visually represented, we make straight lines in hand space. The linearity ratio quantifies path curvature; a straight line is 0.
Rote Learning
We learn an internal model to adapt to a force field. Adaptation to a disturbing force field is in parallel with the development of aftereffects. If the force field is learned for A, then rote would imply it is not learned for B, but it is.
Natural Auditory Scene Statistics
Frequency Elevation Mapping (FEM) for natural auditory scenes and Head-Related Transfer Functions (HRTFs). Priors match FEM. Tilting allows to separate bias and FEM. When info from HRTF becomes unreliable (narrow spectrum), the perceived elevation is determined by the prior.
Correlation Causes Causation
We can’t see causation, so correlation implies causation: temporally correlated perceptions are assumed to come from the same source. Time-shift (→) no synchrony. Multimodal > unimodal. The black dot is the Maximum Likelihood Estimation (MLE) prediction.
Accuracy-Precision Trade-Off in Human Sound Localization
Response gain (accuracy) and variability (precision) of elevation response for Signal-to-Noise Ratio (SNR). Gain = 1 is perfect. Variability = 0 is perfect. More noise (→) worse performance. With a lot of noise, only use the prior. Elevation is worse than azimuth. MAP takes posterior max, AS takes random around the peak of the posterior. MLE represents azimuth data best, AS represents elevation data best.