naturomics/capsnet-tensorflow

This repository is a code implementation of a research idea called CapsNet, or Capsule Networks, originally proposed by Geoffrey Hinton, a well-known figure in AI research. The implementation is built using TensorFlow, which is a popular Python library for running machine learning computations. The goal was to reproduce the results from Hinton's 2017 paper on a new kind of neural network architecture. Standard neural networks process information as individual numbers flowing between layers. Capsule Networks instead group those numbers into small clusters called capsules, where each capsule represents not just whether something is present but also its orientation and other properties. The author's notes describe it as a shift from single-value inputs and outputs to vector-based ones, and compare the routing mechanism to an attention system. The idea is that this approach could better understand spatial relationships in images. In practice, this implementation trains on two image recognition datasets: MNIST, which is handwritten digits, and Fashion-MNIST, which is clothing items. The training results shown in the README come close to but do not fully replicate the accuracy numbers reported in the original paper. The code supports multiple GPUs through a separate folder in the repository. To use it, you download the repository, fetch one of the datasets, and run the training script. A configuration file lets you adjust settings like batch size and how many training passes to run. TensorBoard, a visualization tool included with TensorFlow, can be used to monitor training progress. The author notes that for building new projects on top of capsule ideas, a separate library called CapsLayer is a better starting point than this repository, which was primarily a learning and reproduction exercise.