Perceptual Stability
We move our eyes about 3 times per second. Imagine what a video would look like if you moved the camera that often. How does the brain create a stable percept from its ever changing input?
Saccadic Suppression
When you look into a mirror and move your eyes from left to right and back, you will see that you cannot observe your own eye-movements. This simple experiment demonstrates the phenomenon of saccadic suppression: during saccadic eye movements, visual sensitivity is much reduced. Given that humans make more than 100.000 eye movements each day, it is clear why such a suppression mechanism is needed: without it, the continuous barrage of visual motion on the retina would prevent us from seeing anything at all.
Using psychophysics we have argued that saccadic suppression is dominated by a change in gain of visual detectors; and electrophysiological recordings in the macaque brain show that such gain changes can be found in dorsal areas (MT, MST,VIP, LIP) but also as early as V1. Functional imaging in humans also demonstrated a clear reduction of brain activity in the period surrounding rapid eye movements. This neural correlate of saccadic suppression was clearest in an area (hMT+)
Eye Position Signals
Every time you move your eyes, the retinal image of an object in the world changes dramatically. How then, do we know where things are, even when we move our eyes?
Our work supports the view that the answer lies in the presence of eye position signals in early visual areas. After all, if visual cortex knows where the eyes are pointing, and it knows where an object is on the retina, then it is only a small step to combine these pieces of information and infer where an object is relative to the head and body.
We have shown that the eye position signals present in a distributed manner across areas of the dorsal visual system are precise , accurate, and nimble enough to support perceptual stability most of the time. Sometimes (right around a rapid eye movement) these neural eye position signals are incorrect, and this explains why people report mislocalization errors at these times.
Recurrent Networks
Most cortical processing is dominated by local recurrent feedback; what does this feedback do?
Motion
Most motion models rely on feedforward connectivity. In real cortical tissue, however, neurons are part of many recurrent feedback loops. The model that we are developing starts from this recurrent connectivity and asks how such a network could become direction and speed selective.
To our surprise, we found that it is not only possible to train such a recurrent neural network, the network actually behaves like MT, and like a simple feedforward model in many unexpected ways. We believe this is an interesting example of a system that is inherently strongly nonlinear, but behaves like a linear system when probed in a limited manner.
Orientation
The importance of recurrent connectivity for orientation sensitivity in V1 is relatively well-established, but the consequences of this recurrency for the dynamics of V1 response has received relatively little attention. Our modeling efforts show that recurrence can generate effects that look a lot like short-term adaptation. This is important to better understand adaptation, and may also provide a way to measure effective recurrence in a network