This document discusses a project using convolutional neural networks to recognize handwritten digits from the MNIST dataset. It proposes a hierarchical convolutional neural network approach with two levels - an initial CNN to make preliminary predictions and additional CNNs to further classify ambiguous digits. The model is trained and tested on MNIST data, achieving an error rate of 0.82%. Key aspects covered include CNNs, hierarchical networks, and training/testing a model for handwritten digit recognition.