Sure, there are lots of places where current models (ANNs or not) don’t explain the brain-details. Those seem like interesting areas for research. It’s better if we can quantify “don’t explain” with something like model-brain similarity. What’s the alternative?

You could try to collect data that would be especially hard for CNNs to model (same features at different points in an image, but driving very different responses.) That might help narrow down what non-convolutional mechanisms are at play.

There are lots of ways to make an ANN non-convolutional, so the conclusion for now might be that we haven’t found the right one; but convolutional is a good approximation for the data we have (similar features detected at different points in the visual field.)