WeightScale: Interpreting Weight Change in Neural Networks
This provides interpretability tools for neural network researchers, though it appears incremental as it builds on existing weight analysis methods.
The researchers tackled the problem of interpreting neural network learning dynamics by measuring relative weight changes per layer and aggregating trends with dimensionality reduction and clustering, enabling analysis of very deep networks. They applied this approach to vision tasks with state-of-the-art networks, revealing insights such as how task complexity affects deeper layer learning.
Interpreting the learning dynamics of neural networks can provide useful insights into how networks learn and the development of better training and design approaches. We present an approach to interpret learning in neural networks by measuring relative weight change on a per layer basis and dynamically aggregating emerging trends through combination of dimensionality reduction and clustering which allows us to scale to very deep networks. We use this approach to investigate learning in the context of vision tasks across a variety of state-of-the-art networks and provide insights into the learning behavior of these networks, including how task complexity affects layer-wise learning in deeper layers of networks.