CRCNS AA1 Dataset

Dataset Source: AA1 Dataset

Citation:

Theunissen, FE; Hauber, ME; Woolley, SMN; Gill, P; Shaevitz, SS; Amin, Noopur; Hsu, A; Singh, NC; Grace, GA; Fremouw, T; Zhang, Junli; Cassey, P; Doupe, AJ; David, SV (2009): Single-unit recordings from two auditory areas in male zebra finches. CRCNS.org.
http://dx.doi.org/10.6080/K0F769GP

Papers Using the Dataset:

• “Tuning for Spectro-temporal Modulations: a Mechanism for Auditory Discrimination of Natural Sound” by Woolley S., Fremouw T., Hsu A. and Theunissen F.E.

• “Modulation and phase spectrum of natural sounds enhance neural discrimination performed by single auditory neurons” by Hsu A., Woolley S., Fremouw T. and Theunissen F.E.

• “Modulation spectra of natural sounds and ethological theories of auditory processing” by Singh N. and Theunissen F.E.

Dataset Details

Population fitting: ✅

Description of Stimuli:

• 10 clips of conspecific vocalizations and 20 clips of flat ripples, up to 5 s duration.

• Sample rate @ 32 kHz and 16 bit precision

• Up to 10 response trials for a given sound

Description of Neurons:

• Total Number of Neurons: 100 (50 MLd + 50 Field L)

• Extracellular single-unit recordings from anesthetized male zebra finches

Available data:

• Full Python preprocessing.

• One very simple .txt file for each cell (unit) response:

  • Spike timestamp relative to stimulus onset

  • Each line corresponds to one response trial

Processing needed (Dataset class init() method):

• Transforming the sound waveform (.wav file) into a 32-band spectrogram.

• Choosing neurons based on their recording site and stimulus type.

• Transforming the spike times of each repeat of each stimulus into PSTHs

  • Remove pre-onset spikes

  • Align trials temporally

  • Pad/cut to the right (present/future time steps) so that trials hvae the same duration

Benchmark results

Area	Model backbone	Rank	Remarks	Params / nrn	Perfs (CCraw / CCnorm) [%]	Paper (backbone)
Field L	StateNet	🥇	GRU, pop	24,900	/ 71.0	Rançon et al.
	Transformer	🥈	pop	29,109	/ 65.5	Rançon et al.
	2D-CNN	🥉	pop	26,915	/ 65.0	Pennington et al.
MLd	StateNet	🥇	Mamba, pop	32,334	/ 73.4	Rançon et al.
	2D-CNN	🥈	pop	29,109	/ 68.9	Rançon et al.
	Transformer	🥉	pop	34,475	/ 68.3	Pennington et al.

Setup

Requirements: a CRCNS account (the data host requires login).

Easiest path — auto-download via the CRCNS NERSC mirror:

from deepSTRF.datasets.audio import CRCNSAA1Dataset

ds = CRCNSAA1Dataset(
    download=True, dt_ms=5,
    crcns_username="your_username",
    crcns_password="your_password",
)

Alternatively, set $CRCNS_USERNAME / $CRCNS_PASSWORD in the env and omit the credential kwargs. Default cache dir is platformdirs.user_cache_dir('deepSTRF')/CRCNS_AA1, overridable via $DEEPSTRF_DATA_DIR. download=True is idempotent.

If you already have the data laid out manually:

1. Download crcns-aa1.zip at the original dataset repository.

2. Extract all_stims/, Field_L_cells/, MLd_cells/ into a data/ folder.

3. ds = CRCNSAA1Dataset('/path/to/data', dt_ms=5).

Filtering

Each stim_meta dict carries name, type ("conspecific" or "flatrip"), sample_rate, n_samples, duration_s. Each nrn_meta dict carries cell_id, area ("Field_L" or "MLd"), animal_id, cell_seq, rig. Combined with the base-class selection API:

ds.select_pop_by_nrn_attr("area", "MLd")           # only MLd cells
ds.select_stims_by_attr("type", "conspecific")     # only conspecific stims
                                                   # (auto-hides 2 cells with
                                                   # no conspecific data)