Publications by authors named "Alessandro Barp"
Article Synopsis
- Deep neural networks (DNNs) simplify complex data through geometric and topological transformations, which are not fully understood for non-smooth activation functions like ReLU.
- This study suggests that the changes DNNs make during classification are similar to Hamilton's Ricci flow, a mathematical method used to analyze the shape and topology of spaces.
- By analyzing over 1500 DNN classifiers, the research found that the effectiveness of these geometric transformations is linked to the classifiers' accuracy, highlighting the potential of applying geometric concepts to enhance understanding in deep learning.
