what is frontier-ai-autoencoder fr?

This project is a research experiment that measures how much a neural network can compress an image before the quality of the reconstructed image noticeably suffers. It uses an autoencoder, a type of neural network that learns to squeeze an image down into a smaller representation and then rebuild the original image from that compressed form, trained on the CIFAR-10 dataset of small 32 by 32 pixel images. The experiment trains the same network architecture multiple times, each time using a different size for the compressed representation, ranging from 64 values up to 3072 values. Smaller representations compress the image more but lose more visual detail, while larger ones keep more detail but compress less. The project measures how good each reconstruction is using two standard image quality metrics, mean squared error and peak signal to noise ratio, and plots the results as a tradeoff curve between compression ratio and quality. Under the hood, the network shrinks the image through several convolutional layers that also use residual connections, a technique that helps preserve useful features as data passes through many layers, before compressing everything down to the chosen representation size. A matching set of layers on the other side expands that representation back out into a full image. Results show that quality improves steadily as the representation size grows, but the improvement levels off around a representation size of 2048, suggesting the network's design, not the amount of compressed data it is allowed to keep, becomes the limiting factor at that point. A representation size of 512 achieves the project's target reconstruction quality while still compressing the image sixfold. The project is written in Python using PyTorch. To run it, you set up a virtual environment, install the listed dependencies, download the CIFAR-10 dataset with a provided script, and then run the main training script, which trains and evaluates every tested representation size and saves the resulting models, plots, and sample reconstructions. A separate script lets you feed in your own image and see how well the trained model can reconstruct it, though results will be less accurate for images very different from CIFAR-10 photos.