Freezing feature extractor reduces forgetting in sound classification models

Investigating catastrophic forgetting in sound event classification

Artificial IntelligenceSound

Summary

When teaching computers to recognize new sounds one after another, the problem is they often forget what they learned before, which is called catastrophic forgetting. The authors looked at ways to stop this forgetting during sound recognition tasks. They found that freezing the parts of the model that identify general features while only adjusting the last part that decides the sound type helps prevent forgetting and keeps the model stable. This solution works well in their tests and balances keeping old knowledge with learning new sounds.

What this means in practice

Authors

Riccardo Casciotti, Annamaria Mesaros

Abstract

This work investigates a number of approaches to prevent catastrophic forgetting in class incremental learning scenarios for sound event classification tasks. We analyze the problem using architectural and regularization approaches, using FSD50K and AudioSet datasets. We design incremental stages and solutions that selectively protect the kernels of the network from weight updates to prevent catastrophic forgetting, and a dynamic head solution that expands itself each time a new task is learned. The findings show that catastrophic forgetting mainly happens in deeper layers, in particular in the classifier head. For the studied in-domain sound classification problem, the solution that seems to alleviate catastrophic forgetting and is the most efficient is a full freezing of the feature extractor with a fine-tuning of the dynamic head classifier, showing little to no forgetting and great training stability, and a good balance between memory-stability and learning plasticity.