Efficient coding of spectrotemporal binaural sounds leads to emergence of the auditory space representation

To date a number of studies have shown that receptive field shapes of early sensory neurons can be reproduced by optimizing coding efficiency of natural stimulus ensembles. A still unresolved question is whether the efficient coding hypothesis explains formation of neurons which explicitly represent environmental features of different functional importance. This paper proposes that the spatial selectivity of higher auditory neurons emerges as a direct consequence of learning efficient codes for natural binaural sounds. Firstly, it is demonstrated that a linear efficient coding transform - Independent Component Analysis (ICA) trained on spectrograms of naturalistic simulated binaural sounds extracts spatial information present in the signal. A simple hierarchical ICA extension allowing for decoding of sound position is proposed. Furthermore, it is shown that units revealing spatial selectivity can be learned from a binaural recording of a natural auditory scene. In both cases a relatively small subpopulation of learned spectrogram features suffices to perform accurate sound localization. Representation of the auditory space is therefore learned in a purely unsupervised way by maximizing the coding efficiency and without any task-specific constraints. This results imply that efficient coding is a useful strategy for learning structures which allow for making behaviorally vital inferences about the environment.

💡 Research Summary

The paper investigates whether the efficient coding hypothesis can account for the emergence of neurons that explicitly encode spatial information in the auditory system. By applying Independent Component Analysis (ICA), a linear efficient‑coding transform, to spectro‑temporal representations of natural binaural sounds, the authors demonstrate that spatial selectivity arises automatically, without any task‑specific supervision. Two datasets are used: (1) simulated binaural speech generated by convolving short speech segments with head‑related transfer functions (HRTFs) for 24 azimuthal positions, providing labeled spatial information; and (2) a real‑world binaural recording of three people conversing while moving freely in an echo‑free room, offering uncontrolled but fully natural spatial cues. Each 216 ms sound segment is transformed into a left‑right ear spectrogram (25 overlapping 16 ms windows, 256 logarithmically spaced frequency channels), concatenated into a 12 800‑dimensional vector. After dimensionality reduction with PCA (preserving >99 % variance, reduced to 324 dimensions), ICA learns a set of basis functions (the “dictionary”) that maximally decorrelates the data while enforcing sparsity via a logistic prior.

The learned basis functions are examined for binaural similarity (Pearson correlation between left and right ear parts, termed Binaural Similarity Index, BSI). Functions with BSI near –1 exhibit opposite signs in the two ears, indicating strong spatial selectivity. To quantify how well each component encodes azimuth, the authors compute Fisher information based on the derivative of the mean activation with respect to angle, assuming Gaussian noise. A small subset (≈5 % of the 324 components) yields Fisher information peaks that allow azimuth estimation within 5° error, matching physiological observations that only a minority of higher‑order auditory neurons display sharp spatial receptive fields.

A hierarchical extension is proposed: the sparse activations from the first ICA layer serve as inputs to a second ICA, which directly decodes sound position. This two‑stage model demonstrates that “where” information can be extracted in an unsupervised manner, supporting the claim that redundancy reduction alone can separate “what” (spectral content) from “where” (spatial cues).

The authors also provide a theoretical justification based on cochlear processing. The cochlea performs a Fourier transform and a logarithmic compression; because log(ab) = log a + log b, the spectro‑temporal representation becomes a linear sum of the raw sound spectrum and the HRTF spectrum. Consequently, ICA can separate these additive components, effectively isolating spatial information encoded in the HRTFs.

Overall, the study shows that efficient coding via ICA can automatically give rise to spatially selective auditory representations, both in simulated and natural binaural recordings. The findings suggest that the brain may exploit statistical redundancy reduction to develop a neural map of auditory space without explicit training for localization, offering a unified account of early auditory processing and higher‑order spatial selectivity. This work bridges sensory neuroscience and unsupervised machine learning, providing testable predictions for neurophysiology and a principled framework for designing biologically inspired auditory algorithms.