Monday, February 21, 2011
Sunday, February 20, 2011
Grand Challenge Questions
Grand Challenge Questions
1) Are we close to understanding systems-level neural computation? Why or Why not?
2) What are the key scientific challenges or technologies for achieving an understanding of neural computation?
3) Are there applications for neuromimetic processing that can lead to better technologies immediately? What are they?
These questions are deliberately left vague and their interpretation is left entirely to each speaker. Our hope is that plenary speakers will use this small workshop as an opportunity to present more speculative notions about how truly Neuromimetic Information Processing and (ultimately) Synthetic Cognition might be achieved and what it will look like.
Wednesday, November 18, 2009
Hiding in Plain Sight-cosyne
Hiding in Plain Sight: Effectively masking targets in speed of sight tasks
Humans exhibit remarkable accuracy detecting and recognizing visual objects in natural images, even when images are presented for very short durations (i.e. as little as 20 msec) followed by 1/f noise masks. Such results have been interpreted as evidence that object recognition is primarily mediated by a feed-forward pipeline passing upward through the visual processing hierarchically. However, all masks are not created equal. For example, it can easily be shown that scrambled natural scenes are much more effective masks than are solid gray backgrounds. Still, it is not well known what makes a mask effective. Understanding this phenomenon may be critical to building accurate models of the human visual system. To gain insight into the target/mask relationship, we conducted a series of Speed-of-Sight psychophysical experiments. Subjects were instructed to report whether an object (e.g. animal) was present in a briefly displayed image, quickly followed by a mask whose characteristics were varied. Performance accuracy was strongly influenced by the Stimulus Onset Asynchrony (SOA, the delay between the image and mask onsets, respectively) and the nature of the masks.
In animal/no-animal forced choice tasks, we observed a consistent pattern of mask effectiveness across subjects. Natural scenes and 1/f noise masked most effectively, reducing overall performance to approximately 60% to 65%, respectively, at an SOA of 20 msec, whereas white noise or gray backgrounds were considerably less effective, allowing performance to reach approximately 90% to 95%, respectively, at the same 20 msec SOA. To assess the role of spatial structure, which was present in both the natural scenes and 1/f noise masks but not in the white noise or gray background masks, we developed synthetic stimuli consisting of black & white line drawings; either containing a broken but smooth closed contour (amoeba) embedded in a field of curved line segments drawn from the same statistical distribution (clutter), or else containing just clutter alone. Interestingly, masks consisting of either amoeba+clutter or else containing clutter alone were intermediate in effectiveness, reducing performance to approximately 70% to 75%, respectively, at a 20 msec SOA.
Closer examination revealed that for certain animal-targets, specific amoeba+clutter-masks were especially effective. In particular, those images in which by chance the animal-target appeared in the same area of the visual field and possessed contours similar to the amoeba in the amoeba+clutter mask, reduced overall performance to near 50%. This observation led us to investigate whether an "optimal" mask for the animal/no-animal task might be constructed by using images containing objects at similar locations and possessing similar shapes and textures. We therefore constructed a data-base of images containing either cats (targets) or dogs (masks) and tested subjects in a cat/no-cat task. In these experiments, 1/f noise masks reduced overall performance to approximately 65% at a 20 msec SOA, as in the animal/no-animal task, whereas masks containing dogs reduced performance to approximately chance (50%) at the same SOA. Using images of dogs as a masks, overall performance on the cat/no-cat task require an SOA of nearly 80 msec to reach 80%.
To further investigate mask / target pairing efficacy, we designed an LED task for which the various parameters underlying mask effectiveness (occluding vs. non-occluding, agreement in the visual field) could be systematically controlled. Targets were numbers or patterns, depicted on a 7-segment LED-like display. Masks consisted of numbers, letters and non-meaningful subsets of the same 7 segments. As hypothesized, the numeral 8 (all segments on) served as an effective universal mask requiring long SOAs (80 ms or greater) for reliable detection of target digits. This long delay suggests that even basic object recognition might depend on interruptible feedback between successive visual cortical areas.
Subscribe to:
Posts (Atom)