OSCR

Noisy models of the ventral stream reveal the impact of recurrence and learned representations on information processing timescales.

Code ↔ Paper

1 match between paragraphs of the paper and lines of its authors' code, computed by the harvester (lexical-v1). Click a colored paragraph or line to see its counterpart.

The 1 match
  1. [1] § Methods › Network architecture ↔ cornet/cornet_rt.py, lines 28–97 · score 0.53 · Gaussian noise, CORnet RT, pre, model

Paper

Loaded from Europe PMC by your browser, not stored by OSCR: doi.org · Europe PMC

The paper is loaded when this pane is shown.

The authors' code

Python · 170 lines · 6.6 KB · no license · 1 match

  2. from collections import OrderedDict
  3. import torch
  4. from torch import nn
  5. import numpy as np
  6. HASH = '933c001c'
  7. torch.backends.cudnn.deterministic = False
  8. torch.backends.cudnn.enabled = False
  10. class Flatten(nn.Module):
  11. """
  12. Helper module for flattening input tensor to 1-D for the use in Linear modules.
  13. """
  15. def forward(self, x):
  16. return x.view(x.size(0), -1)
  19. class Identity(nn.Module):
  20. """
  21. Helper module that stores the current tensor. Useful for accessing by name.
  22. """
  24. def forward(self, x):
  25. return x
  28. class CORblock_RT(nn.Module):
  29. def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, out_shape=None, α=0.9, β=0.5, γ=0.9, noise_scale=0):
  30. super().__init__()
  31. self.in_channels = in_channels
  32. self.out_channels = out_channels
  33. self.out_shape = out_shape
  34. self.noise_scale = noise_scale
  36. self.conv_input = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size,
  37. stride=stride, padding=kernel_size // 2)
  38. self.norm_input = nn.GroupNorm(32, out_channels)
  39. self.nonlin_input = nn.ReLU(inplace=True)
  41. self.conv1 = nn.Conv2d(out_channels, out_channels,
  42. kernel_size=3, padding=1, bias=False)
  43. self.norm1 = nn.GroupNorm(32, out_channels)
  44. self.nonlin1 = nn.ReLU(inplace=True)
  45. self._last_pre_relu_state = None
  46. self._last_adaptive_state = None
  47. self.output = Identity() # for easy access to this block's output
  49. self.β = β
  50. self.γ = γ
  51. self.α = α
  53. def forward(self, inp=None, state=None, s=None, timestep=0, output=None, batch_size=None, alpha=0.00001):
  54. # Ensure all tensors are on the same device as the model
  55. device = next(self.conv_input.parameters()).device
  57. if inp is None: # at t=0, there is no input yet except to V1
  58. inp = torch.zeros([batch_size, self.out_channels, self.out_shape, self.out_shape], device=device)
  59. else:
  60. inp = inp.to(device)
  61. inp = self.conv_input(inp)
  62. inp = self.norm_input(inp)
  63. inp = self.nonlin_input(inp)
  65. if state is None: # at t=0, state is initialized to zeros
  66. state = torch.zeros([batch_size, self.out_channels, self.out_shape, self.out_shape], device=device)
  68. if timestep == 0:
  69. s = torch.zeros([batch_size, self.out_channels, self.out_shape, self.out_shape], device=device)
  71. # Generate Gaussian noise (mean=0, std=1) and scale it
  72. gaussian_noise = torch.randn(inp.size(), device=device) # Gaussian noise ~ N(0, 1)
  73. scaled_random_tensor = gaussian_noise * self.noise_scale * alpha
  74. # Update adaptive state `s`
  75. if timestep > 0:
  76. s = self.α * s + (1 - self.α) * output
  79. if output is None:
  80. output = torch.zeros_like(inp)
  81. x_adjusted = inp -(self.β * s)
  82. #x_adjusted = inp + output
  83. x = self.conv1(x_adjusted)
  84. x = self.norm1(x)
  86. # Update state
  87. if timestep == 0:
  88. state = x + scaled_random_tensor
  89. else:
  90. state = x + (1 - self.γ) * state + scaled_random_tensor
  92. state_before_relu = state.clone()
  93. self._last_pre_relu_state = state_before_relu
  94. self._last_adaptive_state = s
  95. output = self.output(self.nonlin1(state))
  97. return output, state, s, state_before_relu
  100. class CORnet_RT(nn.Module):
  101. def __init__(self, times=601):
  102. super().__init__()
  103. self.times = times
  104. self.V1 = CORblock_RT(3, 64, kernel_size=7, stride=4, out_shape=56, α=0.96, β=0., γ=0.6, noise_scale=0.7)
  105. self.V2 = CORblock_RT(64, 128, stride=2, out_shape=28, α=0.96, β=0., γ=0.5, noise_scale=0.7)
  106. self.V4 = CORblock_RT(128, 256, stride=2, out_shape=14, α=0.96, β=0., γ=0.4, noise_scale=0.7)
  107. self.IT = CORblock_RT(256, 512, stride=2, out_shape=7, α=0.96, β=0., γ=0.3, noise_scale=0.7)
  112. self.decoder = nn.Sequential(OrderedDict([
  113. ('avgpool', nn.AdaptiveAvgPool2d(1)),
  114. ('flatten', Flatten()),
  115. ('linear', nn.Linear(512, 1000))
  116. ]))
  118. def forward(self, inp, alpha=0.00001):
  119. # Get the device of the model
  120. device = next(self.V1.conv_input.parameters()).device
  122. outputs = {'inp': inp.to(device)}
  123. states = {}
  124. adaptive_states = {}
  125. state_sums = {}
  126. state_counts = {}
  127. blocks = ['inp', 'V1', 'V2', 'V4', 'IT']
  129. # Initialize the outputs, states, and adaptive states for t=0
  130. for block in blocks[1:]:
  131. if block == 'V1': # at t=0 input to V1 is the image
  132. new_output, new_state, new_s, state_before_relu = getattr(self, block)(
  133. inp[:, 0].to(device), batch_size=len(outputs['inp']))
  134. else: # at t=0 there is no input yet to V2 and up
  135. new_output, new_state, new_s, state_before_relu = getattr(self, block)(
  136. inp=None, batch_size=len(outputs['inp']))
  138. # Initialize output, state, and adaptive state tensors
  139. outputs[block] = torch.zeros((new_output.shape[0], self.times, *new_output.shape[1:]), device=device)
  140. outputs[block][:, 0, :, :, :] = new_output
  142. states[block] = torch.zeros((new_state.shape[0], self.times, *new_state.shape[1:]), device=device)
  143. states[block][:, 0] = state_before_relu
  145. adaptive_states[block] = torch.zeros((new_output.shape[0], self.times, *new_output.shape[1:]), device=device)
  146. adaptive_states[block][:, 0] = new_s
  149. # Run the forward pass for all time steps
  150. for t in range(1, self.times):
  151. for block in blocks[1:]:
  152. prev_block = blocks[blocks.index(block) - 1]
  153. prev_output = outputs[prev_block][:, t if block == 'V1' else t - 1]
  154. current_prev_output = outputs[block][:, t - 1]
  155. prev_state = states[block][:, t - 1]
  156. prev_s = adaptive_states[block][:, t - 1]
  158. # Run the forward pass of the block
  159. new_output, new_state, new_s, state_before_relu = getattr(self, block)(
  160. inp=prev_output, state=prev_state, s=prev_s, timestep=t, output=current_prev_output,
  161. batch_size=len(outputs['inp']), alpha=alpha
  162. )
  164. # Update the dictionaries
  165. outputs[block][:, t, :, :, :] = new_output
  166. states[block][:, t] = state_before_relu
  167. adaptive_states[block][:, t] = new_s
  169. out = self.decoder(outputs['IT'][:, -1])
  170. return out

cornet_rt.py at commit dfded3d, no license · at the source

Overview

Authors: Sara Varetti1, Sebastian Goldt2, Eugenio Piasini1
ORCID iDs: Eugenio Piasini
  1. Cognitive Neuroscience, International School for Advanced Studies (SISSA), Trieste, Italy
  2. Theoretical and Scientific Data Science, International School for Advanced Studies (SISSA), Trieste, Italy
Institutions: Scuola Internazionale Superiore di Studi Avanzati (Italy)
Journal: PLoS computational biology, volume 22, issue 8, article e1014653
Dates: received 3 December 2025; accepted 3 August 2026; published online 27 August 2026
Type: Research article · Language: English
License: CC BY
Identifiers: DOI 10.1371/journal.pcbi.1014653 · PMID 42658886 · PMCID PMC13541256 · OpenAlex W7127105578
Open access: gold, a free copy (OpenAlex)
Preprint: doi:10.64898/2026.01.30.702784
Status: code verified
Categories: computational modeling (no new data) (modality), human (organism), systems (subfield)
Methods: Spectral & time-frequency, Connectivity, Machine learning
MeSH: Models, Neurological*, Visual Cortex*, Visual Pathways*, Visual Perception*, Animals, Computational Biology, Computer Simulation, Humans, Learning, Recurrent Neural Networks (* major topic)
Topic: Visual perception and processing mechanisms (Cognitive Neuroscience, Neuroscience), according to OpenAlex
Funding: European Union – NextGenerationEU – PNRRM4C2-I.1.1 / PRIN (2022XE8X9E); European Research Council (101166056 (Project “beyond2”), 101166056); European Union – NextGenerationEU; Italian government (EP)
Citations: not cited yet (Europe PMC); 53 references in the paper

Abstract

In vision neuroscience, the temporal dynamics of the sensory stream and of its neural representations are thought to be deeply linked to the function of the hierarchy of cortical areas that deal with object recognition, known as the visual ventral stream. Neural representations that are invariant under identity-preserving object transformations, and therefore allow for efficient learning of object identity, are theorized to emerge from a self-supervised learning process that attempts to extract “temporally stable” features from the sensory input. Conversely, invariance increases along the hierarchy, putatively implying progressively slower neural codes in higher-level areas. Recent neurophysiological evidence shows that indeed, as one moves along this cortical hierarchy, neural representations of dynamic stimuli become slower, and additionally the temporal scales of the within-trial fluctuations of these representations (called “intrinsic timescales”) increase starkly. However, while these timescale hierarchies have been reproduced in biologically grounded recurrent models, their network determinants have remained largely unexplored in image-computable models of the ventral stream. Here we investigate the temporal structure of the neural codes in a noisy, recurrent and adaptive model of the ventral visual stream. We show that, surprisingly, the organization of the representation timescales is set by the broad architectural features of the network, regardless of training, while the ordering of the intrinsic timescales across layers is sensitive to the details of the functions implemented by each layer. Our work underscores the importance of the temporal structure of the neural code as a probe for the link between structure and function in models of the vertebrate visual system.

Reproduced under the paper's license (CC BY), from the paper cited above.

Repository

Its files are read in the Code ↔ Paper reader above, with 1 match between paragraphs and lines of code.

Saravaretti/hierarchical-cnn-rnn

License: none: the authors keep all their rights
State: the link answers, verified on 27 September 2026
Evidence: files inventoried
Commit: dfded3def434921d4c8b8b400ec636c894badadf, 30 May 2026
Languages: Python (8)
Size: 10 files, 8 scripts
Software Heritage: not archived
Found in: “Data Availability”
Holds: README
Not found: license file, CITATION.cff, environment file, tests, continuous integration, documentation
Tools: PyTorch (6 files), NumPy (4 files), SciPy (2 files), h5py (1 file), Pillow (1 file)
Availability: 1 check, the latest on 27 September 2026: the link answers
  • 27 September 2026: the link answers
9 files
At the source: github.com/Saravaretti/hierarchical-cnn-rnn

The paper's code and data availability statement is in the Data section.

Tracing map

Proposed by the machine: these links were found in the paper and verified at the source, without human review. The map will receive a Zenodo DOI once one of the paper's authors has validated it with their ORCID.

What the map holds:

  • 1 repository of the authors' code, each at its verified commit, with its license and how the link was found in the paper;
  • 8 scripts, each with its path and the digest of its content;
  • 1 match between paragraphs of the paper and lines of the code (method lexical-v1);
  • neither the text of the paper nor the code itself.

Its JSON (tracing-map.json) is deposited on Zenodo with its DOI once the map is validated.

Data

Datasets cited

Data Availability

All stimuli used in this study are publicly available at: https://osf.io/7gteq/. The code used to generate the results and figures in this paper is publicly available at https://github.com/Saravaretti/hierarchical-cnn-rnn.

Reproduced under the paper's license (CC BY), from the paper cited above.

Versions

The history of this record: each version stored by the harvester or made by a correction of its authors or of the maintainers of its code, and what changed in its facts. The texts of the paper (its abstract, its availability statements) are not part of it; versions that changed only those are not listed.

Version 1, 27 September 2026: the first record

Recorded: type, language, journal, volume, issue, pages, dates, 3 authors, 10 MeSH terms, 4 funders, 47 references.

Cite

This paper

Varetti, S., Goldt, S., & Piasini, E. (2026). Noisy models of the ventral stream reveal the impact of recurrence and learned representations on information processing timescales. PLoS computational biology, 22(8), e1014653. https://doi.org/10.1371/journal.pcbi.1014653

BibTeX

@article{varetti2026noisy,
author = {Varetti, Sara and Goldt, Sebastian and Piasini, Eugenio},
title = {{Noisy models of the ventral stream reveal the impact of recurrence and learned representations on information processing timescales}},
journal = {PLoS computational biology},
year = {2026},
month = aug,
volume = {22},
number = {8},
pages = {e1014653},
publisher = {PLOS},
issn = {1553-734X},
doi = {10.1371/journal.pcbi.1014653},
url = {https://doi.org/10.1371/journal.pcbi.1014653},
pmid = {42658886},
pmcid = {PMC13541256}
}

RIS

TY - JOUR
AU - Varetti, Sara
AU - Goldt, Sebastian
AU - Piasini, Eugenio
TI - Noisy models of the ventral stream reveal the impact of recurrence and learned representations on information processing timescales
T2 - PLoS computational biology
J2 - PLoS Comput Biol
PY - 2026
DA - 2026/08/27
VL - 22
IS - 8
SP - e1014653
SN - 1553-734X
PB - PLOS
DO - 10.1371/journal.pcbi.1014653
UR - https://doi.org/10.1371/journal.pcbi.1014653
LA - en
ER -

CSL-JSON

{
"id": "10.1371/journal.pcbi.1014653",
"type": "article-journal",
"title": "Noisy models of the ventral stream reveal the impact of recurrence and learned representations on information processing timescales",
"container-title": "PLoS computational biology",
"author": [
{
"family": "Varetti",
"given": "Sara"
},
{
"family": "Goldt",
"given": "Sebastian"
},
{
"family": "Piasini",
"given": "Eugenio"
}
],
"container-title-short": "PLoS Comput Biol",
"volume": "22",
"issue": "8",
"page": "e1014653",
"DOI": "10.1371/journal.pcbi.1014653",
"PMID": "42658886",
"PMCID": "PMC13541256",
"ISSN": "1553-734X",
"publisher": "PLOS",
"URL": "https://doi.org/10.1371/journal.pcbi.1014653",
"language": "en",
"issued": {
"date-parts": [
[
2026,
8,
27
]
]
}
}

The tracing map gets a citation of its own once an author has validated it and it has a DOI.

Similar papers

The papers with a page that share the most with this one: the tools found in their code, their categories, datasets, cited references and authors, the rarest counting most.

[1] doi:10.1038/s41467-026-74816-0 [code]
Hierarchical optimization predicts plasticity in the macaque inferior temporal cortex following object training.
Journal: Nature communications
In common: h5py, PyTorch, SciPy, 1 other tool, 5 references
[2] doi:10.1038/s42003-026-10169-0 [code]
Shared representations in brains and models reveal a two-route cortical organization during scene perception.
Journal: Communications biology
In common: h5py, Pillow, PyTorch, 2 other tools, 3 references
[3] doi:10.1038/s42003-026-10094-2 [code]
E/I imbalance and internal noise cause weak neural representations and face recognition challenges in ASD.
Journal: Communications biology
In common: Pillow, PyTorch, SciPy, 1 other tool, computational modeling (no new data), 3 references
[4] doi:10.1162/imag.a.1207 [code]
Investigating the temporal dynamics and modeling of mid-level feature representations in humans.
Journal: Imaging neuroscience (Cambridge, Mass.)
In common: Pillow, PyTorch, SciPy, 1 other tool, 3 references
[5] doi:10.1371/journal.pbio.3003915 [code]
Noise-invariant representations of sound emerge along the canonical cortical hierarchy.
Journal: PLoS biology
In common: h5py, Pillow, PyTorch, 2 other tools, 2 references
[6] doi:10.1038/s41597-026-07248-6 [code]
A large-scale fMRI dataset for vision-language semantic association.
Journal: Scientific data
In common: h5py, Pillow, PyTorch, 2 other tools, 2 references
[7] doi:10.1038/s41467-026-76106-1 [code]
Facial expression discrimination emerges from partially overlapping neural subspaces of detection and identity.
Journal: Nature communications
In common: h5py, SciPy, NumPy, systems, 3 references
[8] doi:10.1016/j.patter.2026.101619 [code]
Sampling bias corrections for discrete and Gaussian partial information decompositions.
Journal: Patterns (New York, N.Y.)
In common: SciPy, NumPy, 4 references
[9] doi:10.1038/s41467-026-74466-2 [code]
Neuromorphic hierarchical modular reservoirs.
Journal: Nature communications
In common: SciPy, NumPy, 4 references
[10] doi:10.1093/cercor/bhag077 [code]
The longitudinal development of intrinsic timescales in infancy and their relation to alpha brain rhythm.
Journal: Cerebral cortex (New York, N.Y. : 1991)
In common: SciPy, NumPy, 4 references

Contribute

The authors of this paper can claim it, correct its record and validate its tracing map, and the maintainers of its code (its owner, or a public member of its organization) correct what it says of their repository; anyone signed in can ask for its removal. Every request goes to OSCR's own machine, which answers it; your account page follows them.

Sign in with ORCID to claim this paper as one of its authors, correct its record or validate its tracing map: when the paper's metadata lists your ORCID iD, you are recognized at once. Maintainers of its code: sign in with GitHub, then claim the repository on your account page.

Request its removal

To ask OSCR to remove this record, the copies of its authors' scripts or its tracing map, use the removal request page: signed in, you say who you are, what to remove and why, then review and confirm the request. Published rules decide every request (how).

Discussion, reproductions, activity

Discussion: questions and error reports about this paper and its code, from signed-in readers and its authors. It opens with sign-in.

Reproductions: reports from readers who ran the authors' code: what they reproduced, with which environment, commit and data. It opens with sign-in.

Activity: what happens around this paper: new versions of its record, its map's validation, discussions and reproductions. It opens with sign-in.