[ipunchghosts]

Why is this not a typical data augmentation method for training deep networks? It would allow the networks to learn invariants that's align with humans. Currently, I suppose none of these dithered images would be correctly classified.

[dist-epoch]

When someone says AI can't do something, I like to try it:

> This image features a close-up abstract scene styled with a classic halftone (Ben-Day dot) pop-art filter ...

https://share.gemini.google/O5eqZOKxwzw3

> This image is a halftone/thermal print styled photograph. The photo features a group of four people posing closely together indoors or in front of a windowed backdrop:

https://share.gemini.google/KJOxBDXYsFT4

[ipunchghosts]

This is a good way to think.

[dahart]

Perhaps because inference on dithered images is pretty rare, not a common or popular thing to do? Otherwise, people would train on dithered images. NNs are trained on noisy images. Training with dither would not learn any other invariants besides how to see through dither, it doesn’t automatically generalize to anything else.

[ipunchghosts]

I am not sure that's true. It would learn to not focus in texture which is what current nns do.

[dahart]

Which NNs and what tasks are you thinking of? DLSS, for example, reconstructs texture, along with other denoisers, as well as all video models. Are you thinking only of object recognition?

Why do you believe training on dithered images would learn how to see through any other kinds of image modification? In my experience, image NNs learn only what you explicitly show them, and they don’t learn concepts they’re not trained on. All the denoisers I’ve ever seen have trouble handling subtle changes in pixel filtering they haven’t seen in the training data, and they don’t easily learn how to denoise things like clouds or reflections. I’d be willing to bet that if you trained on halftoned images, NNs would not be able to handle Floyd Steinberg dither. They just don’t generalize automatically.

[ipunchghosts]

Eero simmoncellis group showed that specific network modifications allow the networks to extrapolate beyond noise they were trained in.

NNs can certainly learn beyond what they were trained for as new architectures have demonstrated. See benchmarks for any if the dan hendryks imagenet corruption dataset benchmarks (imagenet c r and a). Data + architecture work together to enable.

I have work using neural networks trained on synthetic aperture sonar SAS imagery using language only. The SAS datsets are extremely hard to come by. The dataset I used was from parts if world I can assure you we're never in the training set. My approach exceeds what clip models can do zero shot wise. Therefore, the network classified imagery if a modality and environment 100 percent not guaranteed to be in the training data.

[trollbridge]

I’m sorry, but I can’t understand what you’re saying at all. Explain?

[ipunchghosts]

There are many features in an image dataset which a can correctly classify the dataset. Neural nets usually focus in texture which we human clearly don't, we focus on a higher level abstraction and usually ignore texture. Using either would encourage the network to use broader features and ignore texture