Open the neural network "black box"! Unveil the mystery of machine learning system with natural language

Neural networks are sometimes called black boxes , Because although they can surpass humans in some tasks , But even the researchers who designed them often don't understand how or why they work so well .

But if you use neural networks outside the Laboratory , It may be possible to classify medical images that are helpful in diagnosing heart disease , Understanding the working principle of this model will help researchers predict its performance in practice .

MIT researchers have now developed a system called MILAN（ Mutual information guidance language annotation for neurons ） The system of , It can reveal the internal working principle of black box neural network . Taking human brain as a model , Neural networks are arranged as interconnected nodes or 「 Neuron 」 layer . The new system can automatically generate descriptions of these individual neurons in English or other natural languages .

The team showed that , This method can be used to audit Neural Networks , To determine what it learned , You can even edit the network by identifying and then turning off useless or incorrect neurons .

The research 「 Natural Language Descriptions of Deep Visual Features」 entitled , On 2022 year 1 month 26 Published on the preprint platform   arXiv  On .

Thesis link ：https://arxiv.org/abs/2201.11114

Reference Content ：https://techxplore.com/news/2022-01-demystifying-machine-learning-natural-language.html