From 2438d99a94496db4fb0503fb350a2a330a094460 Mon Sep 17 00:00:00 2001 From: Ulysse Date: Tue, 2 Jun 2026 15:28:58 +0200 Subject: [PATCH 1/2] docs: fix fabricated citations; add Transformer ref + RTD API links MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The Transformer/StateNet/AdapTrans docstrings carried a hallucinated author list (Rançon, Bornschein, King, Schnupp, Willmore). Corrected to the real authorship (verified online): - Rançon, Masquelier & Cottereau (2025), Communications Biology 8:1456, doi:10.1038/s42003-025-08858-3 (Transformer + StateNet) - Rançon, Masquelier & Cottereau (2024), PLOS Comput. Biol. 20(8):e1012288, doi:10.1371/journal.pcbi.1012288 (AdapTrans; also fixed the title to the published one) README_models.md: - added the missing Transformer reference (Vaswani et al. 2017 architecture; designed as the attention baseline in Rançon et al. 2025) - linked every model's class to its RTD API page - fixed two wrong years (DNet: Rahman 2016->2019; ConvNet2D: Pennington & David 2023), both verified online - de-linked two broken empty () links in the LN section Co-Authored-By: Claude Opus 4.8 --- deepSTRF/models/audio/audio_zoo.py | 14 ++++++++------ deepSTRF/models/prefiltering.py | 8 ++++---- docs/_source/md/README_models.md | 18 +++++++++--------- 3 files changed, 21 insertions(+), 19 deletions(-) diff --git a/deepSTRF/models/audio/audio_zoo.py b/deepSTRF/models/audio/audio_zoo.py index a094a8a..f77612b 100644 --- a/deepSTRF/models/audio/audio_zoo.py +++ b/deepSTRF/models/audio/audio_zoo.py @@ -548,9 +548,10 @@ class Transformer(AudioEncodingModel): References ---------- - Rançon, Bornschein, King, Schnupp, Willmore (2025). "Temporal - recurrence as a general mechanism to explain neural responses in - the auditory system." Comm. Bio. (preprint on BioRxiv). + Rançon, Masquelier & Cottereau (2025). "Temporal recurrence as a + general mechanism to explain neural responses in the auditory + system." Communications Biology 8:1456. + https://doi.org/10.1038/s42003-025-08858-3 Vaswani et al. (2017). "Attention Is All You Need." NeurIPS. @@ -710,9 +711,10 @@ class StateNet(AudioEncodingModel): References ---------- - Rançon, Bornschein, King, Schnupp, Willmore (2025). - "Temporal recurrence as a general mechanism to explain neural - responses in the auditory system." Comm. Bio. (preprint on BioRxiv). + Rançon, Masquelier & Cottereau (2025). "Temporal recurrence as a + general mechanism to explain neural responses in the auditory + system." Communications Biology 8:1456. + https://doi.org/10.1038/s42003-025-08858-3 Notes ----- diff --git a/deepSTRF/models/prefiltering.py b/deepSTRF/models/prefiltering.py index 77f77fe..6596125 100644 --- a/deepSTRF/models/prefiltering.py +++ b/deepSTRF/models/prefiltering.py @@ -259,10 +259,10 @@ class AdapTrans(nn.Module): References ---------- - Rançon, Bornschein, King, Schnupp, Willmore (2024). "A general - theoretical framework unifying the adaptive, transient and - sustained properties of ON and OFF auditory responses." BioRxiv. - https://doi.org/10.1101/2024.01.17.576002 + Rançon, Masquelier & Cottereau (2024). "A general model unifying + the adaptive, transient and sustained properties of ON and OFF + auditory neural responses." PLOS Computational Biology + 20(8):e1012288. https://doi.org/10.1371/journal.pcbi.1012288 Notes ----- diff --git a/docs/_source/md/README_models.md b/docs/_source/md/README_models.md index 32bfff8..04bafbb 100644 --- a/docs/_source/md/README_models.md +++ b/docs/_source/md/README_models.md @@ -46,22 +46,22 @@ separable kernels. Similar to AdapTrans, prefiltering, hyperparameters for param to the class constructor.* -Torch class: `Linear(...)`; Parameterization available +Torch class: [`Linear`](https://deepstrf.readthedocs.io/en/latest/_source/deepSTRF.models.audio.html#deepSTRF.models.audio.audio_zoo.Linear); Parameterization available ### Linear-Nonlinear (LN) *Consists of a Linear model, with an added output activation which makes it nonlinear. The latter often takes the form of -a sigmoid or parameterized function (see e.g. [Rahman et al.]() or [Willmore et al.]()).* +a sigmoid or parameterized function (see e.g. Rahman et al. or Willmore et al.).* -Torch class: `LinearNonlinear(...)`; Parameterization available +Torch class: [`LinearNonlinear`](https://deepstrf.readthedocs.io/en/latest/_source/deepSTRF.models.audio.html#deepSTRF.models.audio.audio_zoo.LinearNonlinear); Parameterization available ### Network Receptive Field (NRF) *In a nutshell, a LN model with several hidden units.* -Torch class: `NetworkReceptiveField(...)`; Parameterization available; Original paper: [Harper et al. (2016)](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1005113) +Torch class: [`NetworkReceptiveField`](https://deepstrf.readthedocs.io/en/latest/_source/deepSTRF.models.audio.html#deepSTRF.models.audio.audio_zoo.NetworkReceptiveField); Parameterization available; Original paper: [Harper et al. (2016)](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1005113) @@ -72,7 +72,7 @@ Torch class: `NetworkReceptiveField(...)`; Parameterization available; Original *In a nutshell, a NRF model in which hidden and output units follow leaky dynamics (as in LIF spiking neurons, but without spikes), with learnable time constants.* -Torch class: `DNet(...)`; Parameterization available; Original paper: [Rahman et al. (2016)](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1006618) +Torch class: [`DNet`](https://deepstrf.readthedocs.io/en/latest/_source/deepSTRF.models.audio.html#deepSTRF.models.audio.audio_zoo.DNet); Parameterization available; Original paper: [Rahman et al. (2019)](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1006618) @@ -83,7 +83,7 @@ Torch class: `DNet(...)`; Parameterization available; Original paper: [Rahman et kernels do not entirely span all frequencies of the input spectrogram. Fully connected prediction head after a convlutional extraction stage. -Torch class: `ConvNet2D(...)`; Original paper: [Pennington et al.](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1011110) +Torch class: [`ConvNet2D`](https://deepstrf.readthedocs.io/en/latest/_source/deepSTRF.models.audio.html#deepSTRF.models.audio.audio_zoo.ConvNet2D); Original paper: [Pennington & David (2023)](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1011110) ### Recurrent / state-space network (StateNet) @@ -92,7 +92,7 @@ Torch class: `ConvNet2D(...)`; Original paper: [Pennington et al.](https://jour / S4 / LMU). It captures long-range temporal dependencies through the recurrent state and is the strongest model in the zoo on NS1.* -Torch class: `StateNet(...)`; Original paper: [Rançon et al. (2025)](https://doi.org/10.1038/s42003-025-08858-3) +Torch class: [`StateNet`](https://deepstrf.readthedocs.io/en/latest/_source/deepSTRF.models.audio.html#deepSTRF.models.audio.audio_zoo.StateNet); Original paper: [Rançon et al. (2025)](https://doi.org/10.1038/s42003-025-08858-3) ### Transformer @@ -101,7 +101,7 @@ Torch class: `StateNet(...)`; Original paper: [Rançon et al. (2025)](https://do self-attention mask, so it generalizes to any sequence length. An optional finite `context_window` makes attention band-causal.* -Torch class: `Transformer(...)` +Torch class: [`Transformer`](https://deepstrf.readthedocs.io/en/latest/_source/deepSTRF.models.audio.html#deepSTRF.models.audio.audio_zoo.Transformer); Architecture: [Vaswani et al. (2017)](https://arxiv.org/abs/1706.03762); designed as the attention baseline in [Rançon et al. (2025)](https://doi.org/10.1038/s42003-025-08858-3) ### ICNet @@ -112,7 +112,7 @@ audio directly (no precomputed spectrogram). Designed for midbrain (IC) recordings; it ports cleanly into deepSTRF but is oversized for small cortical datasets like NS1.* -Torch class: `ICNet(...)`; Original paper: [Drakopoulos et al. (2025)](https://doi.org/10.1038/s42256-025-01104-9) +Torch class: [`ICNet`](https://deepstrf.readthedocs.io/en/latest/_source/deepSTRF.models.audio.html#deepSTRF.models.audio.ICNet); Original paper: [Drakopoulos et al. (2025)](https://doi.org/10.1038/s42256-025-01104-9) --- From c34c0928219a5a6ef79a84b13fe4d7174d99b692 Mon Sep 17 00:00:00 2001 From: Ulysse Date: Tue, 2 Jun 2026 16:06:13 +0200 Subject: [PATCH 2/2] docs: fix two citation inaccuracies found in the full audit MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - Hsu et al. (2004) title (crcns_aa1): 'Modulation and phase spectrum … enhance neural discrimination' -> the published title 'Modulation power and phase spectrum … enhance neural encoding' (J. Neurosci. 24(41):9201-9211). - ns1: the DNet companion paper (doi pcbi.1006618) was labelled 'Rahman et al. 2018 PLoS Comp Biol'; the PLOS CB paper is 2019 (the 2018 was the bioRxiv preprint). Now consistent with the rest of the file. Full citation audit done this session: ~25 references verified online; all others (Machens 2004, Asari & Zador 2009, Lopez-Espejo 2019, Hsu/Borst 2004, Woolley 2005, Gill 2006, Amin 2010, Singh & Theunissen 2003, Elie & Theunissen 2015/2019, Rahman 2020 PNAS, Harper 2016, Brennan 2019, Brodbeck 2023, Bhattasali 2020, Pennington 2022/2023, Glasberg & Moore 1990, SincNet 2018, LEAF 2021, Wingert 2026, Drakopoulos 2025, Le 2025) confirmed correct. Co-Authored-By: Claude Opus 4.8 --- deepSTRF/datasets/audio/crcns_aa1.py | 4 ++-- deepSTRF/datasets/audio/ns1.py | 4 ++-- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/deepSTRF/datasets/audio/crcns_aa1.py b/deepSTRF/datasets/audio/crcns_aa1.py index 01258b2..06ac122 100644 --- a/deepSTRF/datasets/audio/crcns_aa1.py +++ b/deepSTRF/datasets/audio/crcns_aa1.py @@ -133,8 +133,8 @@ class CRCNSAA1Dataset(AudioNeuralDataset): Woolley et al. (2005). "Tuning for Spectro-temporal Modulations: a Mechanism for Auditory Discrimination of Natural Sound." - Hsu et al. (2004). "Modulation and phase spectrum of natural sounds - enhance neural discrimination performed by single auditory neurons." + Hsu et al. (2004). "Modulation power and phase spectrum of natural + sounds enhance neural encoding performed by single auditory neurons." Singh & Theunissen (2003). "Modulation spectra of natural sounds and ethological theories of auditory processing." diff --git a/deepSTRF/datasets/audio/ns1.py b/deepSTRF/datasets/audio/ns1.py index 8e8539f..a51c7f9 100644 --- a/deepSTRF/datasets/audio/ns1.py +++ b/deepSTRF/datasets/audio/ns1.py @@ -30,7 +30,7 @@ # The precomputed mel-spectrogram tensor (X_nfht: S=20, F=34, hopdim=1, T=999) # used by the original Harper/Rahman analyses is NOT on OSF, but it IS in the -# DNet companion repo (Rahman et al. 2018 PLoS Comp Biol, doi: 10.1371/ +# DNet companion repo (Rahman et al. 2019 PLoS Comp Biol, doi: 10.1371/ # journal.pcbi.1006618 — github.com/monzilur/DNet). The 5 ms version is # ``test_data_5ms.mat`` (5.2 MB); the 1 ms version (``test_data.mat``, 52 MB) # is also there but we don't use it. The same file also contains a ``y_nt`` @@ -93,7 +93,7 @@ def download_ns1(dest: Optional[str] = None) -> str: per-neuron metadata (.mat), and the spike + wav zip (~155 MB total). - **DNet GitHub** (https://github.com/monzilur/DNet, master branch): the precomputed 5 ms mel-spectrogram tensor ``test_data_5ms.mat`` - (5.2 MB) accompanying Rahman et al. 2018 PLoS Comp Biol. NOT on OSF. + (5.2 MB) accompanying Rahman et al. 2019 PLoS Comp Biol. NOT on OSF. Idempotent: skips files that already exist; returns the destination path.