Abstract
Sound Event Detection (SED) in real-world, noisy environments typically requires large, human-designed Deep Neural Networks (DNNs). This paper presents a pilot study on applying neuroevolution to this domain via J-NEAT (Joint Neuro-Evolution of Augmenting Topologies). By combining a highly compressed feature representation with dynamic cooperative co-evolution, the authors successfully evolved remarkably small neural networks that outperformed standard deep learning baselines, securing first place in the F1 metric for the DCASE 2017 Task 3 challenge.
Introduction
Neuroevolution—the process of evolving neural network topologies and weights using genetic algorithms—is typically reserved for simulated environments like video games or virtual robotics where data is clean and fitness evaluations are fast. Applying these techniques to noisy, real-world classification tasks like Sound Event Detection (SED) is computationally prohibitive because standard neuroevolution struggles to scale to the massive network sizes typically used in modern deep learning.
The authors investigate whether they can evolve small, highly efficient networks capable of competing with large, human-designed DNNs on a complex, polyphonic real-world audio task. If successful, these tiny networks could be deployed on hardware with extreme resource constraints, such as autonomous micro-robots.
Methodology
The authors developed a two-part system: a highly compressed feature pipeline and a novel neuroevolution algorithm called J-NEAT.
Compact Feature Extraction
To keep the input size manageable for evolution, the researchers employed a wavelet-based deep scattering transform. This preserves high-frequency transient details (like the "attack" of a sound) while maintaining translation invariance. Because this transform outputs 520 coefficients per time slice, k-means clustering was applied to reduce the complex audio into just 17 centroid-based features, plus one feature representing stereo channel differences.
J-NEAT Algorithm
The J-NEAT algorithm modifies the standard NEAT framework to handle real-world SED:
- Dynamic Cooperative Co-evolution: Three separate populations evolve concurrently, each receiving a random third of the input features. Ensembles are formed dynamically by picking one network from each population to make a joint prediction.
- Evolvable Activation Functions: J-NEAT allows the evolutionary process to select the activation function for each individual node (e.g., ReLU, tanh, sigmoid).
- Destructive Mutations: The algorithm can remove nodes and synapses, actively promoting parsimonious network structures.
- Fixed Recurrence: A corrected layer-wise recurrence mechanism was implemented to handle the temporal nature of audio data.
Results and Discussion
The system was evaluated on the DCASE 2017 Challenge Task 3 dataset, which focuses on sound event detection in real-life, noisy audio environments.
Performance Metrics
On the development set, the J-NEAT ensemble matched the baseline's segment-based Error Rate (ER) (0.73 vs. 0.72) but drastically improved the event-based ER, dropping it from 3.30 to 1.46. In the official DCASE 2017 Challenge, the J-NEAT ensemble achieved 1st place overall on the F1 metric with a score of 44.9%.
Key Insights
- Parsimony: The study proves that massive, deep networks aren't strictly necessary for complex audio detection. Neuroevolution can discover highly parsimonious networks that perform at state-of-the-art levels, opening doors for edge-computing and micro-robotics.
- Ensemble Robustness: In polyphonic audio, the dynamic ensemble approach allowed individual networks with high confidence to override uncertain peers, effectively "tuning into" specific audio characteristics while ignoring overlapping noise.
- Task Decomposition: By forcing different populations to look at different slices of the input features, the algorithm naturally decomposes the difficult SED task into manageable sub-tasks without human-engineered heuristics.
Conclusion
This pilot study demonstrates that neuroevolution is a viable alternative to standard deep learning for real-world audio tasks, particularly when resource efficiency is paramount. While the random splitting of features among populations is a current limitation, future work will explore providing the full feature set to all populations and allowing symbiotic synaptic connections across population boundaries.
Summary updated