Learning by Glueing

Wikipedia: (emphasis mine)

Microsaccades are a kind of fixational eye movement. They are small, jerk-like, involuntary eye movements, similar to miniature versions of voluntary saccades. They typically occur during prolonged visual fixation (of at least several seconds), not only in humans, but also in animals with foveal vision (primates, cats, etc.). Microsaccade amplitudes vary from 2 to 120 arcminutes.
The role of microsaccades in visual perception has been a highly debated topic which is still largely unresolved …

I have a (new?) hypothesis about the origin and necessity of microsaccades for mammalian vision. In a hierarchical cascade of pattern engines, the upper layers may starve from lack of input because the lower layers do their job — construction of locally constant functions — too well. If this happens, Microsaccades provide the necessary glueing data for the upper levels. This mechanism became necessary when the mammalian brain evolved from its reptilian ancestors.

For a more conventional view on the benefits of microsakkades that links them to hyperacuity,
see the Vision & Embedded DSP Group   or   Journal of Statistical Mechanics .

This animation shows sixteen different images of the letter ‘e’, scaled down to 8×8 pixels and quantized with three levels. Prior to scaling and quantization, the very small shifts corresponding to microsaccades are barely visible, but the discretization maps squash some differences and expand others.
Letters like ‘c’ and ‘e’ are difficult to separate at low resolution. Sometimes, they get mapped to the same point in lo-res space. It is quite unlikely,however, that such a singular collision persists under slight variation.

Post a Comment