Mastodawn

Jeremie Jean-Nicolas Sep 26, 2023

Laurent Perrinet Sep 26, 2023

Studying classical #CNN architectures such as #VGG and #ResNet, we observed that they could be sensitive to simple rotations... and that incorporating a biologically inspired retinotopic mapping could alleviate this and also bring other nice features observed in the human visual system.

Hear Jean-Nicolas Jérémie aka @jnjer present these results today at the 32nd International Conference on Artificial Neural Networks (#ICANN 2023) in Heraklion, Greece.

The figure below shows the accuracy of different CNNs to the task "is there an animal in the image" - with a classical (Linear) or retinotopic (polar) mapping. Note that VGG may answer confidently the wrong answer for images rotated around 160°...

More in https://laurentperrinet.github.io/publication/jeremie-23-icann/

Retinotopy improves the categorisation and localisation of visual objects in CNNs | Novel visual computations

to be presented at the 32nd International Conference on Artificial Neural Networks (ICANN 2023) in Heraklion (Greece).

Novel visual computations

Jeremie Jean-Nicolas Jul 12, 2023

Laurent Perrinet Jul 12, 2023

#DeepLearning is fun sometimes, especially when you play with #ImageNet...

Here is one result of (our modified version of) ResNet which gives a wrong answer compared to the ground truth label, yet it is visually accurate.

A warning for us all that the objective is not just to reach the highest accuracy, more to better understand what is going on...

👉 This was a result obtained by Emmanuel Daucé from Aix Marseille Université, in a joint work with @jnjer and myself.

Jeremie Jean-Nicolas May 26, 2023

Laurent Perrinet May 26, 2023

#NewPaper on ultrafast #visualCategorisation in #biology and #NeuralNetworks 🚀

We used transfer learning to learn to detect if an image contains or not an animal. This a priori simple task is in fact not trivial as the animal can be of any species or in any configuration or pose. This showed as a simple perturbation of the image such as a rotation dropped the accuracy from 99% to a catastrophic 72% at an angle of 45° 😱

However, we found out that data augmentation allowed to get a robust response relative to rotation angle, similarly to what is observed in humans recognition abilities.

All the code is available #openSource at https://laurentperrinet.github.io/publication/jeremie-23-ultra-fast-cat/ (with extensive, reproducible supplementary material).

Check out more of the excellent work from PhD candidate Jean-Nicolas Jérémie !

Ultra-Fast Image Categorization in biology and in neural models | Novel visual computations

Humans are able to robustly categorize images and can, for instance, detect the presence of an animal in a briefly flashed image in as little as 120 ms. Initially inspired by neuroscience, deep-learning algorithms literally bloomed up in the last decade such that the accuracy of machines is at present superior to humans for visual recognition tasks. However, these artificial networks are usually trained and evaluated on very specific tasks, for instance on the 1000 separate categories of IMAGENET. In that regard, biological visual systems are more flexible and efficient compared to artificial systems on generic ecological tasks. In order to deepen this comparison, we retrained the standard VGG Convolutional Neural Network (CNN) on two independent tasks which are ecologically relevant for humans: one task defined as detecting the presence of an animal and the other as detecting the presence of an artifact. We show that retraining the network achieves human-like performance level which is reported in psychophysical tasks. We also compare the accuracy of the detection on an image-by-image basis. This showed in particular that the two models perform better when combining their outputs. Indeed, animals (e.g. lions) tend to be less present in photographs containing artifacts (e.g. buildings). These re-trained models could reproduce some unexpected behavioral observations from humans psychophysics such as the robustness to rotations (e.g. upside-down or slanted image) or to a grayscale transformation. Finally, we quantitatively tested the number of layers of the CNN which are necessary to reach such a performance, showing that a good accuracy for ultra-fast categorization could be reached with only a few layers, challenging the belief that image recognition would require a deep sequential analysis of visual objects. We expect to apply this framework to guide future model-based psychophysical experiments and biomimetic deep neuronal architectures designed for such tasks.

Novel visual computations