How Do Neural Networks Learn?

Neural networks are responsible for many AI breakthroughs, including large language models used in a wide range of applications. However, engineers and scientists struggle to understand their inner workings. Data and computer scientists at the University of California San Diego gave neural networks the equivalent of an X-ray to uncover how they actually learn. 

The scientists found that a formula used in statistical analysis provides a streamlined mathematical description of how neural networks learn relevant patterns in data (AKA features). With the formula, the scientists can easily interpret which feature the network uses to make predictions.

The challenge is that it’s currently difficult to understand the mechanism neural networks use to make decisions and biases in the training data that might impact this. Understanding how the networks learn is critical to knowing that they produce reliable, accurate, and appropriate responses.

The team also showed that the formula they used, known as Average Gradient Outer Product (AGOP), could be applied to improve performance and efficiency in other types of machine learning architectures that do not include neural networks.

The machine learning systems they envisioned would need less computational power and less power from the grid to function. They would also be less complex and, therefore, easier to understand. 

The team published their findings in the March 7 issue of the journal Science. You can find it at Mechanism for feature learning in neural networks and backpropagation-free machine learning models