[jcranendonk]

My take on this: if we want vision ML to succeed at recognition in the same way as humans, perhaps we need to pre-process and present visual information in the same way as the human vision system? As far as I'm aware, we get a lot of info from our eyes about lines and orientation that assists in recognizing shapes.

I'm not well-informed about the current state of visual recognition DL, perhaps someone who is can tell us more about whether that approach makes sense.

[tveita]

When you train a deep convolutional neural network, the first couple of layers appear to take on this role, detecting simple features like edges and textures, which the higher layers build upon to see more complex objects.

For example https://www.researchgate.net/figure/Visualization-of-example..., where you can see (somewhat, if you zoom in) that layer 1 neurons are interested in very simple features, like strong horizontal edges, or particular gradients.