Please can you give me a very brief explanation about all functions in the ventral stream architecture summarized in this figure: enter image description here

This figure is from Serre et al.'s A quantitative theory of immediate visual recognition. Prog Brain Res. 2007.

I read multiple articles about this model, but I still don't understand the basic aim, especially behind the two operations (Gaussian-like and max-like operations).. So please can someone explain to me in details the ventral stream pathway (from V1-V2-V4-IT-PFC) including the two operations in this model.

For example : I don't understand how the cells in S1 are constructed...

  • $\begingroup$ The title is misleading. This is just ONE model of the ventral stream processing among many possibilities. Can you be more as to which references you read? Looks like a convolutional deep learning architecture, but you need to tell us more. $\endgroup$ – Memming Oct 19 '13 at 18:20
  • $\begingroup$ thank you for your response. Can you tell me in details what's happen in this picture ? (in S1, C1, etc. ) . Thank you in advance. $\endgroup$ – Christina Oct 19 '13 at 18:27
  • $\begingroup$ The picture doesn't tell much. It seems like it's alternating finding local linear features then aggregating same features across space. You need to point us to where you got the picture, and if you could also write down the equations that would help. Not enough information from the picture! $\endgroup$ – Memming Oct 19 '13 at 19:19
  • $\begingroup$ i edited my picture.. i need only the concept $\endgroup$ – Christina Oct 19 '13 at 20:25
  • $\begingroup$ where did you get the picture? $\endgroup$ – Memming Oct 20 '13 at 3:23

This is a typical architecture of computation proposed as a model for ventral stream of visual processing in primates. It has a long history (e.g., Neocognitoron by Fukushima was 1980) and still widely accepted in machine learning (e.g., deep learning) and neuroscience.


It is motivated by the organization of V1 simple cells and complex cells. Simple cells in V1 can be approximately thought of as edge detectors at a specific retinal location. This is why on the figure you cite, they are represented as a circle with a bar (a cartoon receptive field). The simple cells can only detect things very locally, meaning if the edge appears at a different location in your field of view, it will not respond.

Mathematically, you can think of a spatial filter that detects an edge (e.g., oriented Gabor patch) multiplied to your retinal image, and summed. For example the filter below will detect match a 45 degree bar aligned on the hot-colored area, but will have less activity if the bar is shifted out of the specific position.

oriented Gabor patch

The complex cells in V1, on the other hand is still an edge detector, but has some location invariance. In other words, when the edge is slightly displaced, the response of complex cells does not seem to change. It is believed that this is because complex cells pull from multiple simple cells with the same orientation. This is what you see in your figure where a single complex cell pulls information from the same orientation simple cells but at different locations.

Mathematically, a soft-max operation or a max operation over the simple cell outputs can lead to a good complex cell model. But, it is not limited to such operations. In fact, quadratic or other nonlinear models are also widely used in computational neuroscience.

The full hierarchy for ventral stream is then simply obtained by extending repeatedly using the simple-cell-complex-cell analogy. For each stack simple cell layer extracts some local feature (by computing on the previous layer's complex cell's output), and complex cell layer makes it invariant over space. From edges in V1, one can get corners on the next layer, then complex contours, and all the way up to objects. At least that's how the story goes.

  • $\begingroup$ I don't know how to thank you about this response. I really appreciate it very much... But i still don't understand some details, i will tell you my opinion in the second comment. $\endgroup$ – Christina Oct 21 '13 at 6:50
  • $\begingroup$ I know that in area v1, we have simple and complex cells. each simple cell receive some inputs from the Lateral Geniculate Nucleus (LGN).These inputs are combined with a bell-shaped tuning (gaussian-Like tuning) with the preferred orientation. And all we know that each simple cell can response to a specific oriented bar (in the case of gaussian-like tuning, the response of the cell must be optimal?? because we are making a tuning with the preferred orientation of the receptive field of the cell ?). $\endgroup$ – Christina Oct 21 '13 at 7:05
  • $\begingroup$ However, if we consider such an image, please can you explain to me in details what happen in order to get the tuned simple cells in v1 ? In other words, and in case of the considered image, what we mean about the inputs of such a simple cell? is there a segmentation of the image before the tuning operation? i didn't understand the phenomenon...is there a convolution of the image with a specific filter in order to obtain the image as a form of bars ?? please i need your appreciated help :) and thank you very very much Dear. $\endgroup$ – Christina Oct 21 '13 at 7:06
  • 1
    $\begingroup$ @Liszt You're welcome. The orientation tuning curve can be thought of as a consequence of LN model. If you ask this in a separate question, I could write down the equations for you. $\endgroup$ – Memming Oct 21 '13 at 14:18
  • 1
    $\begingroup$ @Liszt In fact, the tuning curve would have a cosine shape, not exactly Gaussian or von Meises...but they look very similar. This comes from the relation between dot product and cosine. $\endgroup$ – Memming Oct 21 '13 at 14:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.