A recent study shows that individual neurons in a deep network for audio classification can be tied to clear symbolic concepts, such as ‘high‑pitch’ or ‘rhythmic’. The researchers examined the final hidden layer of a 32‑neuron multilayer perceptron (MLP) that classifies music by genre or instrument. 57 numeric sound descriptors were extracted from each audio clip and thresholded at the 25th and 75th percentiles. The lower extreme becomes a low‑level tag and the upper extreme a high‑level tag; values in the middle are ignored. With this tag set, each audio clip becomes an ontology individual, and each tag becomes an ontology class.

For each neuron the team built an OWL ontology and then used DL‑Learner to induce a concept expression that covers clips strongly activating the neuron (activation ≥ 0.60 × max) and excludes weakly activating clips (activation ≤ 0.20 × max). The learned expressions are statistically supported by Mann‑Whitney U tests. The process uncovered several neurons whose induced concepts are linguistically interpretable.

The MLP achieved 0.94 overall accuracy on genre classification and 0.81 on instrument classification, confirming that the network had learned useful audio representations. Out of 32 investigated neurons, 7 of 9 genre‑neurons (21.9 %) and 9 of 10 instrument‑neurons (28.1 %) were validated with meaningful concepts.

The study also highlights limitations. The tags rely on fixed thresholds and may not fully capture the richness of acoustic phenomena. The ontology used is flat and lacks hierarchical structure, which could support more expressive concepts. Finally, only the final layer of a single MLP was examined; other architectures, deeper layers, and validation sets remain to be explored.

These findings point to a promising direction for explaining latent audio representations, but further work is needed to refine tag selection, expand knowledge bases, and test across different model types.