Thanks! :)

Yes! Next Wednesday! 🙏

Congratulations! 🎉🎉 And welcome to Frankfurt!

Yes indeed. It probably has something to do with learning dynamics that favors increasing the complexity gradually. Or it could be that the loss landscape has edges between high and low complexity volumes

In AlexNet, however, the first layers are the most predictive. That's because they have bigger filters at earlier layers (see Miao and Tong 2024)

V1 is usually predicted by more intermediate layers than early layers but it depends on the architecture of the model. In Cadena et al 2019 block3_conv1 in VGG19 was the most predictive. Early layers in VGG have very small receptive fields which makes it difficult to capture V1-like features.

This was the most predictive layer of V1 in the VGG16 model. Same for IT, it was block4_conv2.

and then starts increasing again with further training to fit the target function. This is the most likely explanation for the initial drop in V1 prediction.

We also observed in separate experiments on the simple CNN models that the complexity of the models "resets" to a low value (lower than its random-weight complexity) after the first training epoch (likely using the linear part of the activation function)

Thanks for your interest! Object recognition performance increases directly starting from the first training epoch and nevertheless V1 prediction drops considerably so this drop supports the non significance of object recognition training for V1.

The legend of the left plot was missing!

read more here
www.biorxiv.org/content/10.1...

15/15
It is also important to use various ways to assess model strengths and weaknesses, not just one like prediction accuracy.

14/15

Our results also emphasize the importance of rigorous controls when using black box models like DNNs in neural modeling. They can show what makes a good neural model, and help us generate hypotheses about brain computations

13/15
Our results suggest that the architecture bias of CNNs is key to predicting neural responses in the early visual cortex, which aligns with results in computer vision, showing that random convolutions suffice for several visual tasks.

12/15
We found that random ReLU networks performed the best among random networks and only slightly worse than the fully trained counterpart.

11/15
Then we tested for the ability of random networks to support texture discrimination, a task known to involve early visual cortex. We created Texture-MNIST, a dataset that allows for training for two tasks: object (Digit) recognition and texture discrimination

10/15
We found that trained ReLU networks are the most V1-like concerning OS. Moreover, random ReLU networks were the most V1-like among random networks and even on par with other fully trained networks.