97% de précision (sur 43 items …) mais est ce que corriger les 3% d'erreurs ne prennent pas un temps fou ?
They set about developing a new AI model called GarbageNet, which relies on a type of deep neural network called a convolutional neural network commonly used to analyze images. It uses an existing dataset with labelled images to understand different kinds of garbage.
But the team wanted to build upon the AI further. “The model needs to be flexible enough in order that new categories of garbage can be recognized by the model without much effort,”
Therefore, the team added a second module to the model that’s able to memorize mysterious, new items of garbage and categorize them based it on their similarities to items the AI is already familiar with. In this way, GarbageNet can expand upon the categories of garbage it recognizes without extra training. Lastly, a third component of the model helps the AI learn better when multiple pieces of garbage are present, reducing the “background noise” of the other items.
via Deep Learning Weekly : lire l’article source