Really efficient face detection

Do you remember when OpenAI published "Multimodal Neurons in Artificial Neural Networks"? If you don't know what this is, it's basically the "essence" of prompts for CLIP (Image identifier neural network).

CLIP's rendition of Summer Nature

That got me thinking. What would be the easiest way to recreate this? Well, i created a neural network with only an input and output layer. Then i used the FFHQ Dataset, Natural images Dataset and CIFAR 10 in black and white 128x128. After 10000 epochs i got this:

Pretty cool, right? So actually, since they are actually weights, you can multiply any image by this one, pass it though a sigmoid function and you will get a result from 0-1! This is so easy that i made a p5js sketch with face identification.

Probably i will try this on other datasets. If you download the weights, you can download them here

Download files

Filename	Size
weights.txt	128 KB