Denosing Images of Cats and Dogs with Autoencoders (mayberay.bearblog.dev)

The post demonstrates practical autoencoder experiments: first a simple fully connected autoencoder that compresses MNIST 28x28 images to a 2‑D latent vector (Flatten → Linear(784→latent) → ReLU → Linear(latent→784) → Sigmoid) trained with BCE loss, and then a much deeper convolutional RED‑NET (symmetric encoder–decoder with residual skip connections) applied to denoise cats and dogs corrupted with Gaussian noise (σ=1). Code and full training details are available in the linked Kaggle notebook. The RED‑NET mirrors each convolutional layer with a corresponding deconvolutional block and adds skip activations back into the decoder (x + skip), similar in spirit to UNet/ResNet but without downsampling concatenations. This is significant because it ties foundational autoencoder concepts (encoder, bottleneck, decoder, latent space) to practical image restoration: showing how symmetric skip connections preserve detail across deep encoder–decoder stacks and how loss choice influences perceptual quality. Key takeaways include why Sigmoid+BCE was used for binary image outputs, why MSE/BCE tend to produce blurry reconstructions, and how low-dimensional latent spaces limit fidelity and sampling diversity (random sampling from a 2‑D latent produced repetitive digits). The author concludes by motivating Variational Autoencoders to impose structure on the latent space for more diverse, controllable generation—covered in the next post.