Can a neural network learn to "see" grammar the way a fluent reader does?

In this second volume of the Applied Intelligence Series, computer scientist and educator Steeve Adjeh Venance turns a classroom observation from his own French workbook into a real, tested machine learning experiment. French adjectives and articles must agree with their noun in gender and number — masculine or feminine, singular or plural — and fluent speakers often recognize the correct form by its shape alone, before they could recite the rule.

This book formalizes that instinct. A compact convolutional neural network is trained to classify French adjectives purely from images of the printed word — no meaning, no context, just visual pattern. Tested on words it had never seen before, the network reaches 83% accuracy, and when applied directly to real questions from the author's own workbook, it gets four out of five right using sight alone.

Along the way, the book puts this result in conversation with the history of facial recognition — from Eigenfaces to DeepFace — to explain exactly why a tiny dataset was enough here, when face recognition needed millions of images. Real code, real results (including the honest failures), and a critical look at what this small experiment does and doesn't prove.