This post is also available in: עברית (Hebrew)
A field of machine learning, deep learning is based on the assumption that computers can learn and teach themselves, aiming to mimic the brain’s activity in a computerized form. It enables machines to solve complex tasks even given a large and diverse set of unstructured data. A deep learning system can be found today in almost every technical field, from computer vision to bioinformatics and medical analysis.
As a part of deep learning, Deep Neural Networks (DNNs) are used, which are inspired by biological systems’ information processing and distributed communication. Training these networks isn’t cheap, and it’s quite complicated, too.
A deeper understanding of deep learning can help find new methods for reducing training costs, thus enhancing the effectiveness of machine learning. Weightwatcher is a new open-source Python tool that may be able to help. Participants can utilize the tool to evaluate the performance of their machine learning – information that is analyzed in depth, and which offers insights and data about the training process of the model, as well as warnings if anything goes wrong.
The new tool applies concepts from theoretical physics and Random Matrix Theory (RMT) models to measure correlation between data. Additionally, it can be used to estimate the model’s test accuracy without any test data. As a result, it can make it easier to fine tune pretrained models when applying