This project investigates the feature representations learned by convolutional neural networks (CNNs) by reconstructing images from the feature maps of different layers of a pre-trained AlexNet model, inspired by “Understanding Deep Image Representations by Inverting Them” (Dosovitskiy & Brox, 2016).

We:

• Use Tiny ImageNet, selecting a subset of classes for training and evaluation.
• Train separate reconstruction models for feature maps from conv2, conv5, fc6, and fc8.
• Base decoder architectures on the structures in the original paper, with reduced depth for computational efficiency.
• Produce loss plots and visualizations of reconstructed images.

👉 Full code and detailed description:
Image Reconstruction from Feature Maps on GitHub