Individual neurons in the songbird caudomedial nidopallium (NCM) respond not to a single spectrotemporal feature but to up to a dozen or more independent excitatory and suppressive features of natural song, i.e. they have composite receptive fields. A two-layer network trained with the sparse filtering algorithm — which enforces sparseness and divisive normalization — rediscovers the same feature statistics, suggesting but not demonstrating that these two canonical computations generate the biological fields. The source paper (Kozlov & Gentner, PNAS 2016) explicitly posed as a key question for future research the biological implementation of sparseness and normalization in sensory circuits and the functional significance of discontinuous, composite, apparently random receptive fields. The 2016–2026 literature indicates that composite receptive fields are a generic property of high-order auditory processing (mouse, cat, ferret, and even astrocytes), and that convergent indirect evidence supports an ensemble-coding function, but no work has identified the circuit or biophysical implementation of the sparseness-plus-divisive-normalization computation in any auditory circuit producing these fields, and no direct causal behavioral evidence establishes what the composite structure is for. The problem remains open.

Background

A classical way to characterize an auditory neuron is the spectrotemporal receptive field (STRF): a linear filter describing how features of the sound spectrogram drive the neuron's firing. Linear and quadratic statistical models, such as the maximum noise entropy (MNE) model — a logistic model of spike probability fitted so that its predictions match the neuron's firing rate and the first- and second-order correlations between spikes and the stimulus spectrogram — can recover several filters per neuron and thus reveal whether a cell is driven by one feature or by many. Using MNE analysis, Kozlov and Gentner (2016) showed that single neurons in the caudomedial nidopallium (NCM) — a high-order auditory area of the songbird forebrain, analogous to mammalian secondary auditory cortex — have composite receptive fields: each neuron is driven by up to a dozen or more independent excitatory and suppressive spectrotemporal features of comparable magnitude, so that any one of many distinct features of a natural song can elicit a response. Sparse, selective responses to vocalizations are well documented in the avian auditory pathway (Schneider & Woolley 2013), and the same group had shown that NCM neurons flexibly recombine multiple features with AND-like or OR-like logic (Kozlov & Gentner 2014). Two properties are near-universal in biological sensory circuits. Sparseness means each neuron responds to only a small fraction of all stimuli (lifetime sparseness) and each stimulus activates only a small fraction of neurons (population sparseness); it is theoretically linked to the expansion of sensory representations (Babadi & Sompolinsky 2014). Divisive normalization means a neuron's response is divided by a measure of the pooled activity of the surrounding circuit, rescaling responses across stimulus and contrast contexts; it is considered a canonical neural computation (Carandini & Heeger 2012). Kozlov and Gentner trained a two-layer artificial network on starling songs using sparse filtering, an unsupervised objective that enforces population sparseness and competition (normalization) between units; second-layer units of this network developed composite receptive fields whose feature-magnitude statistics matched the in vivo NCM data. This established computational sufficiency of sparseness plus divisive normalization for generating composite fields, but said nothing about how — or whether — biological circuits implement these computations. The paper also drew an explicit parallel with olfaction: in mammalian piriform cortex and the insect mushroom body, quasi-random projections from earlier stages are thought to provide an unbiased, expandable basis from which learning carves associations, and high-order olfactory neurons likewise have composite fields; it noted the related phenomenon of mixed selectivity in prefrontal cortex, where neurons are tuned to mixtures of task-relevant features. Against this background the paper posed the explicit future-research question that this problem formalizes: understand the implementation of sparseness and normalization in sensory neural circuits, and the functional significance of discontinuous, composite, and apparently random receptive fields in songbirds and other animals.

Problem Statement

The problem, as posed in the source (Kozlov & Gentner 2016): understand both (1) the biological implementation of sparseness and divisive normalization in sensory neural circuits — the circuit, synaptic, or biophysical mechanisms by which these computations are realized so as to generate composite receptive fields, the songbird NCM and homologous high-order auditory areas, where such fields are documented, providing the concrete setting — and (2) the functional significance of the discontinuous, composite, and apparently random receptive fields found in high-order auditory neurons of songbirds and other animals, that is, what this receptive-field structure is for, whether at the level of neural coding (for example the robust, statistically optimal representation of complex natural signals, or the ensemble-level advantage of allowing each neuron to participate in many different ensembles, as the source proposes) or at the level of perception and behavior. The source situates this question within a broader framing — the success of random projections for dimensionality reduction in compressed sensing and machine learning, and the parallel with the quasi-random projections of olfactory circuits (mammalian piriform cortex and the insect mushroom body) — suggesting that composite auditory receptive fields may reflect general principles of neural computation across sensory systems and animal classes; this framing is background context for the question rather than part of it. The two sub-questions are jointly posed in the source as one key question.

References

  1. Kozlov AS, Gentner TQ (2016) Central auditory neurons have composite receptive fields. Proceedings of the National Academy of Sciences USA 113(5):1441-1446. doi:10.1073/pnas.1506903113, https://doi.org/10.1073/pnas.1506903113
  2. Schneider DM, Woolley SMN (2013) Sparse and background-invariant coding of vocalizations in auditory scenes. Neuron 79:141-152. doi:10.1016/j.neuron.2013.04.038, https://doi.org/10.1016/j.neuron.2013.04.038
  3. Kozlov AS, Gentner TQ (2014) Central auditory neurons display flexible feature recombination functions. Journal of Neurophysiology 111:1183-1189. doi:10.1152/jn.00637.2013, https://doi.org/10.1152/jn.00637.2013
  4. Carandini M, Heeger DJ (2012) Normalization as a canonical neural computation. Nature Reviews Neuroscience 13:51-62. doi:10.1038/nrn3136, https://doi.org/10.1038/nrn3136
  5. Babadi B, Sompolinsky H (2014) Sparseness and expansion in sensory representations. Neuron 83:1213-1226. doi:10.1016/j.neuron.2014.07.035, https://doi.org/10.1016/j.neuron.2014.07.035
  6. Bjoring MC, Meliza CD (2019) A low-threshold potassium current enhances sparseness and reliability in a model of avian auditory cortex. PLoS Computational Biology 15:e1006723. doi:10.1371/journal.pcbi.1006723, https://doi.org/10.1371/journal.pcbi.1006723
  7. Shih JY, Yuan K, Atencio CA, Schreiner CE (2020) Distinct manifestations of cooperative, multidimensional stimulus representations in different auditory forebrain stations. Cerebral Cortex 30:3130-3147. doi:10.1093/cercor/bhz299, https://doi.org/10.1093/cercor/bhz299
  8. Lu S, Ang GWY, Steadman M, Kozlov AS (2023) Composite receptive fields in the mouse auditory cortex. The Journal of Physiology 601:4091-4104. doi:10.1113/JP285003, https://doi.org/10.1113/JP285003
  9. Mukherjee S, Babadi B, Shamma S (2025) Sparse high-dimensional decomposition of non-primary auditory cortical receptive fields. PLoS Computational Biology 21:e1012721. doi:10.1371/journal.pcbi.1012721, https://doi.org/10.1371/journal.pcbi.1012721
  10. Lu S, Schultz SR, Kozlov AS (2026) Astrocytes and neurons exhibit partially shared but distinct composite receptive fields for natural stimuli. Journal of Neurophysiology 135:382-393. doi:10.1152/jn.00505.2025, https://doi.org/10.1152/jn.00505.2025
  11. Robotka H, Thomas L, Yu K, Wood W, Elie JE, Gahr M, Theunissen FE (2023) Sparse ensemble neural code for a complete vocal repertoire. Cell Reports 42:112034. doi:10.1016/j.celrep.2023.112034, https://doi.org/10.1016/j.celrep.2023.112034
  12. Yu K, Wood WE, Theunissen FE (2020) High-capacity auditory memory for vocal communication in a social songbird. Science Advances 6:eabe0440. doi:10.1126/sciadv.abe0440, https://doi.org/10.1126/sciadv.abe0440
  13. Theilman B, Perks K, Gentner TQ (2021) Spike train coactivity encodes learned natural stimulus invariances in songbird auditory cortex. The Journal of Neuroscience 41:73-88. doi:10.1523/JNEUROSCI.0248-20.2020, https://doi.org/10.1523/JNEUROSCI.0248-20.2020
  14. Rudraraju S, Turvey ME, Theilman BH, Gentner TQ (2024) Predictive and error coding for vocal communication signals in the songbird auditory forebrain. bioRxiv preprint. doi:10.1101/2024.02.25.581987, https://doi.org/10.1101/2024.02.25.581987
  15. Moseley SM, Meliza CD (2025) A complex acoustical environment during development enhances auditory perception and coding efficiency in the zebra finch. The Journal of Neuroscience 45:e1269242024. doi:10.1523/JNEUROSCI.1269-24.2024, https://doi.org/10.1523/JNEUROSCI.1269-24.2024
  16. Spool JA, Macedo-Lima M, Scarpa G, Morohashi Y, Yazaki-Sugiyama Y, Remage-Healey L (2021) Genetically identified neurons in avian auditory pallium mirror core principles of their mammalian counterparts. Current Biology 31:2831-2843. doi:10.1016/j.cub.2021.04.039, https://doi.org/10.1016/j.cub.2021.04.039
  17. Velho TAF, Iancu D, Brenno Galvão de Miranda Costa R, Roberts P, Mello CV (2026) Auditory representation of vocal signals in a pallial cortical circuit. The Journal of Neuroscience 46:e0852252026. doi:10.1523/JNEUROSCI.0852-25.2026, https://doi.org/10.1523/JNEUROSCI.0852-25.2026
  18. Lian Y, Burkitt AN (2025) Relating sparse and predictive coding to divisive normalization. PLoS Computational Biology 21:e1013059. doi:10.1371/journal.pcbi.1013059, https://doi.org/10.1371/journal.pcbi.1013059
  19. Wingert J, Parida S, Norman-Haignere S, David SV (2026) Convolutional neural network models describe the encoding subspace of local circuits in auditory cortex. Nature Neuroscience 29:876-887. doi:10.1038/s41593-026-02216-0, https://doi.org/10.1038/s41593-026-02216-0
  20. Cazemier J, Douin L, Cornelisse L, Dubey M, Kole M, Heimel J (2026) Divisive normalization in mouse V1 is unchanged by increased thalamocortical short-term depression. bioRxiv preprint. doi:10.64898/2026.07.13.738133, https://doi.org/10.64898/2026.07.13.738133
  21. Vahidi NW (2021) Spatial and temporal organization of composite receptive fields in the songbird auditory forebrain. bioRxiv preprint. doi:10.1101/2021.07.21.453115, https://doi.org/10.1101/2021.07.21.453115