Convolutional neural networks are very popular these days in computer vision. One of the things they do quite well is recognize objects with translation invariance. The way they do this is by "tying together" the strengths of certain synapses, so that they all change in the same way during learning. This is apparently not biologically plausible.

My question is this: what is known about how biological visual systems achieve translation invariance. For example, how is it that if we learn to recognize a new object with our left eye, we are automatically also able to recognize it with our right eye. What is the neuro-biological basis for this invariance? My field is mathematics/machine learning, not biology, so I would appreciate responses that are not overly technical (although I understand that might not be possible).


  • 1
    $\begingroup$ Great question - unfortunately I think what you are asking is actually quite broad. I'm hopeful someone might come along with a reasonable answer, but in fact this is a very open field in biology right now. We know a bit about what areas of visual cortex show some translation invariance in their tuning, but my understanding is that's about it, beyond computational modeling that shows feasible approaches. $\endgroup$
    – Bryan Krause
    Mar 21 '17 at 16:05
  • 1
    $\begingroup$ From the perspective of machine learning, consider that the visual cortex has many more layers than a typical artificial neural network, and the layers are interconnected in a much more complex and less strict hierarchy. This organization allows for incredibly complex computational abilities. Importantly, biological neural networks have simultaneous bottom-up and top-down processing, allowing brain areas with very abstract representations to make predictions about less abstract features. Additionally, attention and related processes allow for online reorganization of networks. $\endgroup$
    – Bryan Krause
    Mar 21 '17 at 16:08
  • $\begingroup$ @BryanKrause you gave a very good general insight to seemingly broad question. I would like to know if you pulled this insight from a paper? I would like to read it. Especially about "simultaneous bottom-up and top-down processing". $\endgroup$
    – gfdsal
    Sep 5 '20 at 18:46
  • 1
    $\begingroup$ @gfdsal I didn't have a single paper in mind, but Felleman, D. J., & Van Essen, D. C. (1991). Distributed hierarchical processing in the primate cerebral cortex. In Cereb cortex. is a good place to start. You could also look for recent reviews about "predictive coding", though this is only one of the schemes/frameworks used to think about hierarchical processing and it remains to be seen if it's the best description. I also think you'd find at least some discussion in an undergraduate neuroscience textbook like Purves or Kandel. $\endgroup$
    – Bryan Krause
    Sep 5 '20 at 19:34
  • $\begingroup$ +1 for providing credible references and direction to take. $\endgroup$
    – gfdsal
    Sep 5 '20 at 20:14

Your Answer

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

Browse other questions tagged or ask your own question.