Does anyone else get the impression that given "the set of C2 features that is passed to the final classifier is very redundant, and probably more redundant than other approaches" it shouldn't be "[interesting that] the number of features needed to reach the ceiling (about 5000 features) is much larger than the number used by current systems"? Doesn't the redundancy in features mean that they by definition carry less novel information per feature than a less redundant set?

Training seems to consist entirely of randomly sampling patches from seen images. It seems both unlikely that this is the best thing to do or that it is at all an approximation of biological vision.

First, the authors claim that it's agreed on that early vision "follows a mostly feedforward hierarchy." The impression I get is that this is far from agreed upon, that there are many feedback connections to V1 and the LGN, and that there's little agreement about how these feedback connections work.

Edited by author 10-17-2006 03:03 PM
I wonder if shape features would improve recognition in this approach, although they say that biology is unlikely to use geometrical information (why?). It would have been interesting to know the performance on computing the features for the detection part (single class and multiclass experiments). I didn't understand very well what is max pooling.

The authors make the seemingly bold claim that "with such degree of invariance, it is unlikely that the SIFT-based features could perform well on a generic object recognition task." Is this a fair assessment? Has SIFT been applied with better results in the past to these datasets? If they are correct, I would think they could have found a better performing target for comparison to their own features.

When they compare their method to SIFT, they use SIFT in a very funny way... They select 1000 keypoints at random from the training data, and then the feature vector for each novel image is the min distance from each selected training keypoint to some keypoint in the novel image (if I understand correctly). I've never heard of this being done before, and I wonder how it compares to the more standard ways of using SIFT for object recognition.

Also, I wonder how fast their implementation is.

Q: SVM usually has better performance using non-linear kernels such as RBF. Why do this paper and most its citations use a linear kernel?

Is it because that
(a) Linear kernel is more similar to huamn neural network.
(b) Non-linearity has been introduced at S2 step.
(c) #feature is so large so that the problem is linear separable

In page 2, column 2:
When determining which filters to use for S1, a large number of filters is created by sampling the parameter space, then applying those filters to commonly used V1 stimuli and removing those which are incompatible with biological cells.
How is it determined if a filter is incompatible with biological cells? and what effect would using an incompatible filter have on the overall process?

1. On page 3, top of 2nd column:

    "The large pool of K patches are of various sizes"

    Do "K" and the "patch size" have a biological interpretation?


2. Figure 6 shows the S2 features. What do the C2 features look like since they are the final features being used?