Exploiting Developments in Auto-Machine Learning to Advance the Field of Audio-Machine Learning

Abstract:
Audio neural network design requires specialist expertise that few practitioners possess. Can graph-theoretic methods and automated search reduce the manual effort involved whilst maintaining competitive performance? After establishing that traditional classifiers cannot match neural performance on audio benchmarks, three graph-theoretic approaches are investigated: centrality-based network compression, differentiable architecture search adapted for audio, and graph neural network methods for architecture prediction, generation, and search. On compression, convolutional filters are modelled as nodes in a similarity graph and ranked by centrality to identify structural redundancy, requiring no training data. On the Pre-trained Audio Neural Networks (PANNs) 14-layer convolutional variant (CNN14), this achieves 43.58% mean Average Precision (mAP) at 78% compression, approximately matching the unpruned baseline. On architecture search, Differentiable Architecture Search (DARTS) is adapted with audio-specific search spaces and evaluated across two audio classification benchmarks against expert-designed baselines and architectures randomly sampled from the same search space; the cell-based architectures outperform the compact expert baseline, although the DARTS search method does not show a clear advantage over random sampling in these runs. On architecture analysis, weight-agnostic networks evolved for audio tasks reach above-chance classification with a single shared untrained weight, and the evolved topology used as initialisation keeps a 6.5 percentage point advantage over a random-topology baseline. Graph neural network surrogates trained on a controlled pool outperform scalar centrality and a parameter-count regression for architecture ranking. These results indicate that graph-theoretic analysis and automated search can reduce the manual effort required to design and compress audio neural networks whilst maintaining competitive performance.