OSCR

A single computational objective can produce specialization of streams in visual cortex.

Code ↔ Paper

16 matches 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 16 matches · 1 of them tie a paragraph to a whole file, not to given lines: a weak match, whose lines are not tinted
  1. [1] § Results › Stream-hypothesized task models do not yield stream-specific spatial or functional correspondence ↔ notebooks/demo_plot_combined_source_data.ipynb, lines 333–399 · score 0.88 · MB v3 categorization, ConvNeXT, SlowFast, MB v2, MB v1, ResNet
  2. [2] § Results › Computational framework for stream-organization hypotheses ↔ notebooks/demo_plot_combined_source_data.ipynb, lines 333–399 · score 0.87 · DepthAnythingv2, ConvNeXT, DEKR Pose, MB v3, MB v2, MB v1
  3. [3] § Methods › Representational overlap across MB v1 task models ↔ scripts/cka_mb18_overlap.py, lines 1–45 · score 0.87 · backbone.feature_provider.feature_provid, slow.layer4.1, linear CKA, overlap, detection, stimuli
  4. [4] § Methods › Representational overlap across MB v1 task models ↔ scripts/fitting_one_to_one_unit2voxel.py, lines 30–112 · score 0.77 · backbone.feature_provider.feature_provid, slow.layer4.1, detection, v1, layers, models
  5. [5] § Methods › Model unit selectivity and receptive field properties ↔ spacetorch/datasets/floc.py, lines 79–200 · score 0.69 · category selective, corridors, houses, instruments, limbs, adult
  6. [6] § Results › Stream-hypothesized task models do not yield stream-specific spatial or functional correspondence ↔ matlab/F04_A.m, lines 86–103 · score 0.67 · ConvNeXT, SlowFast, ResNet, CNN, SSD, Faster
  7. [7] § Methods › Training: Neural network architectures and training tasks › Multiple behaviors models ↔ spacestream/utils/get_utils.py, lines 206–287 · score 0.66 · SlowFast model, detection model, Kinetics, vision, Faster, backbone
  8. [8] § Methods › Testing: Evaluating theories › Linear regression ↔ spacestream/utils/mapping_utils.py, lines 17–84 · score 0.62 · linear regression, voxel responses, ridge, splits, mapping, layer
  9. [9] § Results › Stream-hypothesized task models do not yield stream-specific spatial or functional correspondence ↔ notebooks/F04_A_processing.ipynb, lines 129–147 · score 0.59 · v3 depth, v3 pose, v3 categorization, MB, dorsal, lateral
  10. [10] § Methods › Training: Neural network architectures and training tasks › Spatial constraints models › Loss functions ↔ spacestream/models/slow_fast.py, lines 10–76 · score 0.59 · cross entropy loss, loss function, network, layer, model
  11. [11] § Methods › Testing: Evaluating theories › Brain data › Category selectivity by stream in human cortex ↔ matlab/transform_selectivity_to_fsaverage.m, the whole file · a weak match · score 0.59 · fsaverage space, contrast maps, selectivity, bodies, fLoc, NSD
  12. [12] § Results › Stream-hypothesized task models do not yield stream-specific spatial or functional correspondence ↔ notebooks/F03_B_processing.ipynb, lines 155–164 · score 0.55 · MB v2, MB v1, multiple behavior, collapsed, hypothesized, models
  13. [13] § Methods › Training: Neural network architectures and training tasks › Multiple behaviors models ↔ spacestream/utils/get_utils.py, lines 206–287 · score 0.55 · OpenAI, encoder, vision, Pretraining, ConvNeXT, DETR
  14. [14] § Methods › Training: Neural network architectures and training tasks › Spatial constraints models › Loss functions ↔ spacetorch/losses/cross_entropy_spatial_correlation_loss.py, lines 11–58 · score 0.54 · cross entropy loss, batch, layer
  15. [15] § Methods › Training: Neural network architectures and training tasks › Spatial constraints models › Initialization of model unit position ↔ spacetorch/datasets/__init__.py, lines 1–35 · score 0.54 · retinal waves, sine grating
  16. [16] § Methods › Testing: Evaluating theories › Linear regression ↔ spacestream/analyses/base_fitter.py, lines 102–209 · score 0.52 · lower dimensional space, regression, splits, layer, train, stream

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

Jupyter notebook · 991 lines · 41 KB · Apache-2.0 · 2 matches

  1. # %% [markdown]
  2. # # Demo: Plot Combined Manuscript Source Data
  3. #
  4. # This notebook reads only `data/spacestream_manuscript_source_data.xlsx`. Each worksheet is one manuscript figure panel; panels with multiple source tables are parsed from stacked sections within that worksheet.
  6. # %%
  7. from pathlib import Path
  8. import os
  9. import zipfile
  10. import xml.etree.ElementTree as ET
  12. import matplotlib.pyplot as plt
  13. from matplotlib.lines import Line2D
  14. import numpy as np
  15. import pandas as pd
  16. import seaborn as sns
  18. ROOT = Path.cwd()
  19. if ROOT.name != "manuscript_figure_source_data":
  20. ROOT = Path("/oak/stanford/groups/kalanit/biac2/kgs/projects/Dawn/manuscript_figure_source_data")
  22. WORKBOOK = ROOT / "data" / "spacestream_manuscript_source_data.xlsx"
  23. OUT = ROOT / "outputs" / "combined_demo_plots"
  24. OUT.mkdir(parents=True, exist_ok=True)
  26. os.environ.setdefault("MPLCONFIGDIR", str(ROOT / ".matplotlib"))
  27. os.environ.setdefault("XDG_CACHE_HOME", str(ROOT / ".cache"))
  28. (ROOT / ".matplotlib").mkdir(exist_ok=True)
  29. (ROOT / ".cache").mkdir(exist_ok=True)
  31. plt.rcParams.update({
  32. "figure.dpi": 120,
  33. "savefig.dpi": 200,
  34. "axes.spines.top": False,
  35. "axes.spines.right": False,
  36. "font.size": 10,
  37. "font.family": "sans-serif",
  38. "font.sans-serif": ["Helvetica", "Arial", "DejaVu Sans"],
  39. "pdf.use14corefonts": True,
  40. "ps.useafm": True,
  41. "svg.fonttype": "none",
  42. "pdf.fonttype": 42,
  43. "ps.fonttype": 42,
  44. })
  45. sns.set_theme(style="ticks", rc={
  46. "font.family": "sans-serif",
  47. "font.sans-serif": ["Helvetica", "Arial", "DejaVu Sans"],
  48. })
  50. CORE_ROI_NAMES = ["Ventral", "Lateral", "Dorsal"]
  51. ROI_ORDER = ["Dorsal", "Lateral", "Ventral"]
  52. ROI_COLORS = ["#377E2C", "#1A1AAC", "#8C1A4C"]
  53. STREAM_PALETTE = {"Ventral": ROI_COLORS[2], "Lateral": ROI_COLORS[1], "Dorsal": ROI_COLORS[0]}
  54. STREAM_COLORS = {
  55. "Dorsal": "#006600",
  56. "Parietal": "#006600",
  57. "Lateral": "#4d7fff",
  58. "Ventral": "#990000",
  59. }
  60. STREAM_ORDER = ["Dorsal", "Lateral", "Ventral"]
  61. LEGACY_STREAM_ORDER = ["Parietal", "Lateral", "Ventral"]
  62. CONTRAST_ORDER = ["Places", "Bodies", "Faces"]
  63. TDANN_COLORS = {"simCLR": "#720298", "self_supervised": "#720298", "supervised": "#CB6D4A"}
  66. def savefig(fig, stem):
  67. png = OUT / f"{stem}.png"
  68. pdf = OUT / f"{stem}.pdf"
  69. fig.tight_layout()
  70. fig.savefig(png)
  71. fig.savefig(pdf)
  72. plt.close(fig)
  73. print(png.relative_to(ROOT))
  74. print(pdf.relative_to(ROOT))
  77. def despine(ax):
  78. ax.spines["top"].set_visible(False)
  79. ax.spines["right"].set_visible(False)
  82. def clean_legend_to_streams(ax, stream_order):
  83. handles, labels = ax.get_legend_handles_labels()
  84. stream_handles = {}
  85. for handle, label in zip(handles, labels):
  86. if label in stream_order and label not in stream_handles:
  87. stream_handles[label] = handle
  88. if stream_handles:
  89. ax.legend(stream_handles.values(), stream_handles.keys(), frameon=False, title="Stream")
  90. else:
  91. ax.legend([], [], frameon=False)
  95. # %% [markdown]
  96. # ## Load Combined Panel Workbook
  98. # %%
  99. NS_MAIN = "{http://schemas.openxmlformats.org/spreadsheetml/2006/main}"
  100. NS_REL = "{http://schemas.openxmlformats.org/package/2006/relationships}"
  101. NS_DOC_REL = "{http://schemas.openxmlformats.org/officeDocument/2006/relationships}"
  104. def _read_shared_strings(zf):
  105. if "xl/sharedStrings.xml" not in zf.namelist():
  106. return []
  107. root = ET.fromstring(zf.read("xl/sharedStrings.xml"))
  108. return ["".join(node.text or "" for node in item.findall(f".//{NS_MAIN}t")) for item in root.findall(f"{NS_MAIN}si")]
  111. def _read_sheet_rows(zf, sheet_path, shared_strings):
  112. root = ET.fromstring(zf.read(sheet_path))
  113. rows = []
  114. for row in root.findall(f".//{NS_MAIN}row"):
  115. values = []
  116. expected_col = 1
  117. for cell in row.findall(f"{NS_MAIN}c"):
  118. ref = cell.attrib.get("r", "A1")
  119. col_idx = 0
  120. for ch in "".join(ch for ch in ref if ch.isalpha()):
  121. col_idx = col_idx * 26 + (ord(ch.upper()) - 64)
  122. while expected_col < col_idx:
  123. values.append("")
  124. expected_col += 1
  125. cell_type = cell.attrib.get("t")
  126. if cell_type == "inlineStr":
  127. text_node = cell.find(f"{NS_MAIN}is/{NS_MAIN}t")
  128. values.append(text_node.text if text_node is not None and text_node.text is not None else "")
  129. elif cell_type == "s":
  130. value_node = cell.find(f"{NS_MAIN}v")
  131. idx = int(value_node.text) if value_node is not None and value_node.text is not None else -1
  132. values.append(shared_strings[idx] if 0 <= idx < len(shared_strings) else "")
  133. else:
  134. value_node = cell.find(f"{NS_MAIN}v")
  135. values.append(value_node.text if value_node is not None and value_node.text is not None else "")
  136. expected_col += 1
  137. while values and values[-1] == "":
  138. values.pop()
  139. rows.append(values)
  140. return rows
  143. def _numericize(df):
  144. out = df.copy()
  145. for col in out.columns:
  146. try:
  147. out[col] = pd.to_numeric(out[col])
  148. except (TypeError, ValueError):
  149. pass
  150. return out
  153. def load_source_workbook(path):
  154. with zipfile.ZipFile(path) as zf:
  155. shared_strings = _read_shared_strings(zf)
  156. workbook_xml = ET.fromstring(zf.read("xl/workbook.xml"))
  157. rels_xml = ET.fromstring(zf.read("xl/_rels/workbook.xml.rels"))
  158. rel_targets = {rel.attrib["Id"]: rel.attrib["Target"] for rel in rels_xml.findall(f"{NS_REL}Relationship")}
  159. panel_rows = {}
  160. for sheet in workbook_xml.findall(f".//{NS_MAIN}sheet"):
  161. name = sheet.attrib["name"]
  162. rid = sheet.attrib[f"{NS_DOC_REL}id"]
  163. sheet_path = "xl/" + rel_targets[rid].lstrip("/")
  164. panel_rows[name] = _read_sheet_rows(zf, sheet_path, shared_strings)
  165. return panel_rows
  168. def extract_panel_tables(panel_rows):
  169. panel_tables = {}
  170. tables = {}
  171. csvs = {}
  172. metadata_keys = {"figure_panel", "source_workbook", "source_table", "staging_csv", "rows"}
  173. for panel, rows in panel_rows.items():
  174. panel_tables[panel] = {}
  175. i = 0
  176. while i < len(rows):
  177. row = rows[i]
  178. if row and row[0] == "source_table":
  179. table_name = row[1]
  180. staging_csv = ""
  181. i += 1
  182. while i < len(rows):
  183. row = rows[i]
  184. if row and row[0] == "staging_csv":
  185. staging_csv = row[1] if len(row) > 1 else ""
  186. if row and row[0] not in metadata_keys:
  187. break
  188. i += 1
  189. if i >= len(rows):
  190. break
  191. header = rows[i]
  192. data_rows = []
  193. i += 1
  194. while i < len(rows):
  195. row = rows[i]
  196. if row and row[0] == "source_table":
  197. break
  198. if row:
  199. data_rows.append(row)
  200. i += 1
  201. width = len(header)
  202. padded = [r + [""] * max(width - len(r), 0) for r in data_rows]
  203. df = _numericize(pd.DataFrame([r[:width] for r in padded], columns=header))
  204. panel_tables[panel][table_name] = df
  205. tables[table_name] = df
  206. if staging_csv:
  207. csvs[Path(staging_csv).stem] = df
  208. continue
  209. i += 1
  210. return panel_tables, tables, csvs
  212. panel_rows = load_source_workbook(WORKBOOK)
  213. panel_tables, sheets, csvs = extract_panel_tables(panel_rows)
  214. print(f"Loaded {len(panel_tables)} panel sheets from {WORKBOOK}")
  215. print(f"Extracted {len(sheets)} source tables")
  216. for panel, table_map in panel_tables.items():
  217. print(panel, sorted(table_map))
  221. # %% [markdown]
  222. # ## Fig 1A: Mega Matrix And MDS
  224. # %%
  225. coords = csvs["fig1a_mds_coordinates_random_state0"].copy()
  226. mega_matrix = csvs["fig1a_mega_matrix"].to_numpy(dtype=float)
  228. roi_colors = {
  229. "Early": "#a6a6a6",
  230. "Midventral": "#f4bdd8",
  231. "Midlateral": "#ccdaff",
  232. "Midparietal": "#b3ffc6",
  233. "Ventral": "#DC267F",
  234. "Lateral": "#4d7fff",
  235. "Parietal": "#006600",
  236. }
  237. markers = {"lh": "v", "rh": "o"}
  239. fig, ax = plt.subplots(figsize=(10, 10))
  240. for (roi, hemi), group in coords.groupby(["roi_raw", "hemi"]):
  241. ax.scatter(
  242. group["mds_dim1"],
  243. group["mds_dim2"],
  244. s=35 + 20 * group["subject"].astype(int),
  245. c=roi_colors[roi],
  246. marker=markers[hemi],
  247. edgecolors="white",
  248. linewidths=0.5,
  249. alpha=0.9,
  250. label=f"{roi}, {hemi}",
  251. )
  252. ax.set_xlabel("MDS dimension 1")
  253. ax.set_ylabel("MDS dimension 2")
  254. ax.set_title("Fig 1A MDS from saved mega matrix")
  255. ax.spines[["top", "right"]].set_visible(False)
  256. savefig(fig, "fig1a_mds_source_style_demo")
  258. fig, ax = plt.subplots(figsize=(8, 8))
  259. im = ax.imshow(mega_matrix, cmap="magma", vmin=0, vmax=1)
  260. ax.set_title("Fig 1A source mega matrix")
  261. ax.set_xlabel("Matrix index")
  262. ax.set_ylabel("Matrix index")
  263. fig.colorbar(im, ax=ax, fraction=0.046, pad=0.04)
  264. savefig(fig, "fig1a_mega_matrix_demo")
  267. # %% [markdown]
  268. # ## Fig 1B: Human fLoc Selectivity
  270. # %%
  271. def plot_floc_bar_strip(subject_df, hemi_df, stem, title, hue_order=None):
  272. sns.set_theme(style="ticks", context="talk")
  273. hue_order = hue_order or ["Dorsal", "Lateral", "Ventral"]
  274. fig, ax = plt.subplots(figsize=(6, 6), facecolor="white")
  275. ax.set_facecolor("white")
  277. sns.barplot(
  278. data=subject_df,
  279. x="contrast",
  280. y="frac_t_above",
  281. hue="stream",
  282. order=CONTRAST_ORDER,
  283. hue_order=hue_order,
  284. palette=STREAM_PALETTE,
  285. ci="sd",
  286. capsize=0.08,
  287. errwidth=1.5,
  288. edgecolor="white",
  289. linewidth=4,
  290. ax=ax,
  291. )
  293. markers = {"lh": "o", "rh": "^"}
  294. for hemi, marker in markers.items():
  295. sns.stripplot(
  296. data=hemi_df[hemi_df["hemi"] == hemi],
  297. x="contrast",
  298. y="frac_t_above",
  299. hue="stream",
  300. order=CONTRAST_ORDER,
  301. hue_order=hue_order,
  302. dodge=True,
  303. jitter=0.15,
  304. marker=marker,
  305. size=4.5,
  306. edgecolor="white",
  307. linewidth=0.5,
  308. palette=STREAM_PALETTE,
  309. ax=ax,
  310. )
  312. clean_legend_to_streams(ax, hue_order)
  313. ax.spines["right"].set_visible(False)
  314. ax.spines["top"].set_visible(False)
  315. ax.set_xlabel("Contrast")
  316. ax.set_ylabel("Fraction t > 3")
  317. ax.set_title(title)
  318. ax.grid(False)
  319. ax.tick_params(colors="black", labelcolor="black")
  320. ax.spines["left"].set_color("black")
  321. ax.spines["bottom"].set_color("black")
  322. savefig(fig, stem)
  324. plot_floc_bar_strip(
  325. csvs["fig1b_human_floc_selectivity_subject_bars"],
  326. csvs["fig1b_human_floc_selectivity_hemi_dots"],
  327. "fig1b_human_floc_selectivity_source_style_demo",
  328. "Fig 1B human fLoc selectivity",
  329. hue_order=["Dorsal", "Lateral", "Ventral"],
  330. )
  333. # %% [markdown]
  334. # ## Figs 3B/3C/4A/4B: Model Correspondence Panels
  336. # %%
  337. FIG3_MODEL_ORDER = [
  338. "MB_RN50", "MB_RN50_v2", "MB_RN50_v3", "MB_RN18",
  339. "TDANN_Supervised_0.0", "TDANN_SimCLR_0.0", "TDANN_Supervised", "TDANN_SimCLR",
  340. ]
  341. FIG3_MODEL_LABELS = ["MB v1", "MB v2", "MB v3", "MB v1", "Cat", "SimCLR", "TDANN Cat", "TDANN SimCLR"]
  342. FIG3_IN_BAR_LABELS = ["50", "50", "50", "18", "18", "18", "18", "18"]
  343. FIG3_XVALS = [1, 2, 3, 4.5, 5.5, 6.5, 8, 9]
  345. FIG4_MODEL_ORDER = [
  346. "MB_RN50_detection", "MB_RN50_action", "MB_RN50_categorization",
  347. "MB_RN50_v2_detection", "MB_RN50_v2_clip", "MB_RN50_v2_categorization",
  348. "MB_RN50_v3_depth", "MB_RN50_v3_pose", "MB_RN50_v3_categorization",
  349. "MB_RN18_detection", "MB_RN18_action", "MB_RN18_categorization",
  350. ]
  351. FIG4_MODEL_LABELS = ["SSD", "SlowFast", "ResNet", "DETR", "CLIP", "ConvNeXT-T", "DepthAnythingv2", "DEKR-Pose", "ResNet", "Faster R-CNN", "SlowFast", "ResNet"]
  352. FIG4_IN_BAR_LABELS = ["50", "50", "50", "50", "50", "50", "50", "50", "50", "18", "18", "18"]
  353. FIG4_XVALS = [1, 2, 3, 5, 6, 7, 9, 10, 11, 13, 14, 15]
  355. MATLAB_BAR_COLORS_FIG3 = ["#808080", "#808080", "#808080", "#808080", "#cc3300", "#660091", "#cc3300", "#660091"]
  356. MATLAB_BAR_ALPHAS_FIG3 = [0.35, 0.35, 0.35, 0.5, 0.5, 0.5, 0.75, 1.0]
  357. MATLAB_BAR_COLORS_FIG4 = ["#80b380", "#8080b3", "#cc8066", "#80b380", "#6666e6", "#cc8066", "#66cc66", "#4d4de6", "#cc8066", "#80b380", "#8080b3", "#cc8066"]
  360. def plot_matlab_like_three_roi(points, noise, model_order, model_labels, in_bar_labels, xvals, colors, alphas, stem, title, ylabel):
  361. noise_summary = noise.groupby("ROI")["result"].agg(["mean", "std"])
  362. ymax = float((noise_summary["mean"] + noise_summary["std"]).max() * 1.1)
  363. fig, axes = plt.subplots(1, 3, figsize=(16, 6), sharey=True)
  365. for ax, roi in zip(axes, ROI_ORDER):
  366. roi_points = points[points["ROIS"] == roi]
  367. band = noise_summary.loc[roi]
  368. ax.axhspan(band["mean"] - band["std"], band["mean"] + band["std"], color="0.9", alpha=1.0, zorder=0)
  369. means = roi_points.groupby("model_type")["result"].mean()
  370. for i, model in enumerate(model_order):
  371. mean = means.loc[model] if model in means.index else np.nan
  372. ax.bar(xvals[i], mean, width=0.95, color=colors[i], alpha=alphas[i], edgecolor="none", zorder=2)
  373. if np.isfinite(mean):
  374. ax.text(xvals[i], ymax * 0.04, in_bar_labels[i], rotation=90, color="0.82", fontsize=13, ha="center", va="bottom")
  375. model_rows = roi_points[roi_points["model_type"] == model]
  376. lh = model_rows[model_rows["hemi"] == "lh"]["result"]
  377. rh = model_rows[model_rows["hemi"] == "rh"]["result"]
  378. edge = {"Dorsal": "#003300", "Lateral": "#000033", "Ventral": "#330018"}[roi]
  379. ax.scatter(np.full(len(lh), xvals[i]), lh, marker="^", facecolors="none", edgecolors=edge, linewidths=1.4, zorder=3)
  380. ax.scatter(np.full(len(rh), xvals[i]), rh, marker="o", facecolors="none", edgecolors=edge, linewidths=1.4, zorder=3)
  381. ax.set_xticks(xvals)
  382. ax.set_xticklabels(model_labels, rotation=90, fontsize=10)
  383. ax.set_xlim(0, max(xvals) + 1.5)
  384. ax.set_ylim(0, ymax)
  385. ax.set_title(roi, fontsize=18)
  386. ax.spines["right"].set_visible(False)
  387. ax.spines["top"].set_visible(False)
  388. if ax is axes[0]:
  389. ax.set_ylabel(ylabel, fontsize=14)
  390. else:
  391. ax.tick_params(labelleft=False)
  392. ax.spines["left"].set_visible(False)
  393. fig.suptitle(title, fontsize=18)
  394. savefig(fig, stem)
  396. plot_matlab_like_three_roi(csvs["fig3b_spatial_correspondence_points"], csvs["fig3b_spatial_correspondence_noise_ceiling"], FIG3_MODEL_ORDER, FIG3_MODEL_LABELS, FIG3_IN_BAR_LABELS, FIG3_XVALS, MATLAB_BAR_COLORS_FIG3, MATLAB_BAR_ALPHAS_FIG3, "fig3b_spatial_correspondence_source_style_demo", "Fig 3B spatial correspondence", "Distance similarity [r]")
  397. plot_matlab_like_three_roi(csvs["fig3c_functional_correspondence_points"], csvs["fig3c_functional_correspondence_noise_ceiling"], FIG3_MODEL_ORDER, FIG3_MODEL_LABELS, FIG3_IN_BAR_LABELS, FIG3_XVALS, MATLAB_BAR_COLORS_FIG3, MATLAB_BAR_ALPHAS_FIG3, "fig3c_functional_correspondence_source_style_demo", "Fig 3C functional correspondence", "Functional similarity [r]")
  398. plot_matlab_like_three_roi(csvs["fig4a_mb_task_stream_assignment_points"], csvs["fig4a_mb_task_stream_assignment_noise_ceiling"], FIG4_MODEL_ORDER, FIG4_MODEL_LABELS, FIG4_IN_BAR_LABELS, FIG4_XVALS, MATLAB_BAR_COLORS_FIG4, [0.5] * 12, "fig4a_mb_task_stream_assignment_source_style_demo", "Fig 4A MB task-stream assignment", "Task-stream assignment [%]")
  399. plot_matlab_like_three_roi(csvs["fig4b_mb_task_functional_correspondence_points"], csvs["fig4b_mb_task_functional_correspondence_noise_ceiling"], FIG4_MODEL_ORDER, FIG4_MODEL_LABELS, FIG4_IN_BAR_LABELS, FIG4_XVALS, MATLAB_BAR_COLORS_FIG4, [0.5] * 12, "fig4b_mb_task_functional_correspondence_source_style_demo", "Fig 4B MB task functional correspondence", "Functional similarity [r]")
  402. # %% [markdown]
  403. # ## Fig 5B: Spatial Weight Curves
  405. # %%
  406. def plot_fig5b_source_style(points, bands, stem_prefix, title_prefix, ylim_top):
  407. sns.set_theme(style="ticks")
  408. line_kwargs = {"marker": ".", "markersize": 18, "lw": 1, "markeredgecolor": "w", "markeredgewidth": 1.5}
  409. version_order = [v for v in ["self_supervised", "supervised"] if v in set(points["version"])]
  410. label_map = {"self_supervised": "Self-supervised", "supervised": "Supervised"}
  412. for streamx, stream in enumerate(CORE_ROI_NAMES):
  413. fig, ax = plt.subplots(figsize=(3, 5))
  414. df_stream = points[points["stream"] == stream].rename(columns={"spatial_weight": "Spatial_Weight", "corr": "Corr", "version": "Version"})
  415. sns.lineplot(
  416. data=df_stream,
  417. x="Spatial_Weight",
  418. y="Corr",
  419. hue="Version",
  420. hue_order=version_order,
  421. palette=[TDANN_COLORS[v] for v in version_order],
  422. errorbar="se",
  423. ax=ax,
  424. **line_kwargs,
  425. )
  426. ax.set_xscale("symlog", linthresh=0.09)
  427. ax.set_xlim([-0.01, 50])
  428. ax.set_xticks([], minor=True)
  429. ax.set_xticks([0, 0.1, 0.25, 0.5, 1.25, 2.5, 25])
  430. ax.set_xticklabels([0, 0.1, "", "", 1.25, "", 25])
  431. band = bands[bands["stream"] == stream].iloc[0]
  432. ax.axhspan(band["mean"] - band["std"], band["mean"] + band["std"], xmin=0, xmax=1, color="lightgray", alpha=0.75)
  433. if streamx == 2:
  434. h, _ = ax.get_legend_handles_labels()
  435. ax.legend(h, [label_map[v] for v in version_order], frameon=False)
  436. else:
  437. ax.legend([], [], frameon=False)
  438. ax.set_xlabel("Spatial Weight", fontsize=14)
  439. ax.set_ylabel("Correlation", fontsize=14)
  440. ax.set_yticks([])
  441. ax.spines["left"].set_visible(False)
  442. ax.spines["right"].set_visible(False)
  443. ax.spines["top"].set_visible(False)
  444. ax.set_ylim(bottom=0.0, top=ylim_top)
  445. ax.set_title(f"{title_prefix}: {stream}")
  446. savefig(fig, f"{stem_prefix}_{stream.lower()}_source_style_demo")
  448. plot_fig5b_source_style(csvs["fig5b_sc_spatial_correspondence_by_weight_points"], csvs["fig5b_sc_spatial_correspondence_brain_band_mean_sd"], "fig5b_top_spatial_correspondence", "Fig 5B top", 0.3)
  449. plot_fig5b_source_style(csvs["fig5b_sc_functional_correspondence_by_weight_points"], csvs["fig5b_sc_functional_correspondence_brain_band_mean_sd"], "fig5b_bottom_functional_correspondence", "Fig 5B bottom", 0.5)
  452. # %% [markdown]
  453. # ## Fig 5C: Effective Dimensionality vs Functional Similarity
  455. # %%
  456. points = csvs["fig5c_effective_dimensionality_vs_functional_similarity_points"].copy()
  457. points = points.rename(columns={"spatial_weight": "Spatial Weight", "type": "Type", "combined": "Combined", "stream": "Stream", "ed": "ED", "corr": "Corr"})
  458. bands = csvs["fig5c_brain_reference_bands_mean_sd"]
  460. for ridx, roi in enumerate(CORE_ROI_NAMES):
  461. fig, ax = plt.subplots(figsize=(7, 9))
  462. sns.set_theme(style="ticks")
  463. df_roi = points[points["Stream"] == roi]
  464. band = bands[bands["stream"] == roi].iloc[0]
  465. ax.axhspan(band["mean_corr"] - band["std_corr"], band["mean_corr"] + band["std_corr"], xmin=0, xmax=1, color="lightgray", alpha=0.75)
  466. ax.axvspan(band["mean_ed"] - band["std_ed"], band["mean_ed"] + band["std_ed"], color="lightgray", alpha=0.75)
  468. sup_open = df_roi[(df_roi["Type"] == "supervised") & ~(df_roi["Spatial Weight"].astype(str).isin(["2.5"]))]
  469. sim_open = df_roi[(df_roi["Type"] == "simCLR") & ~(df_roi["Spatial Weight"].astype(str).isin(["0.25", "0.5"]))]
  470. sup_filled = df_roi[(df_roi["Type"] == "supervised") & (df_roi["Spatial Weight"].astype(str).isin(["2.5"]))]
  471. sim_filled = df_roi[(df_roi["Type"] == "simCLR") & (df_roi["Spatial Weight"].astype(str).isin(["0.25", "0.5"]))]
  473. ax.scatter(sup_open["ED"], sup_open["Corr"], s=70, facecolors="white", edgecolors="#CB6D4A", linewidths=2, alpha=0.3, zorder=10, label="supervised")
  474. ax.scatter(sim_open["ED"], sim_open["Corr"], s=70, facecolors="white", edgecolors="#720298", linewidths=2, alpha=0.3, zorder=10, label="simCLR")
  475. ax.scatter(sup_filled["ED"], sup_filled["Corr"], s=130, color="#CB6D4A", alpha=0.65, zorder=11)
  476. ax.scatter(sim_filled["ED"], sim_filled["Corr"], s=90, color="#720298", alpha=0.65, zorder=11)
  477. ax.set_xscale("log")
  478. ax.set_ylim(0, 0.43)
  479. ax.set_xlabel("Effective Dimensionality")
  480. ax.set_ylabel("Correlation")
  481. ax.set_title(f"Fig 5C {roi}")
  482. ax.legend(frameon=False)
  483. ax.spines["right"].set_visible(False)
  484. ax.spines["top"].set_visible(False)
  485. savefig(fig, f"fig5c_ed_vs_functional_similarity_{roi.lower()}_source_style_demo")
  488. # %% [markdown]
  489. # ## Figs 6A/6B: Transfer Panels
  491. # %%
  492. def plot_transfer_source_style(points, value_col, stem, title, ylim=None):
  493. sns.set_theme(style="ticks")
  494. fig, ax = plt.subplots(figsize=(3, 6))
  495. raw_streams = {"Ventral": "Ventral", "Dorsal": "Parietal"}
  496. wide = points.pivot_table(index="line_index", columns="stream_display", values=value_col, aggfunc="first")
  497. for _, row in wide.dropna(subset=["Dorsal", "Ventral"]).iterrows():
  498. ax.plot([0.025, 0.975], [row["Dorsal"], row["Ventral"]], c="k", alpha=0.15)
  500. for stream_display, xpos, color in [("Ventral", 1, "#8C1A4C"), ("Dorsal", 0, "#377E2C")]:
  501. for hemi, marker in [("lh", "^"), ("rh", "o")]:
  502. vals = points[(points["stream_display"] == stream_display) & (points["hemi"] == hemi)][value_col]
  503. ax.scatter(
  504. np.full(len(vals), xpos), vals, color=color, s=70, alpha=0.95,
  505. edgecolors="w", marker=marker, linewidths=0.5,
  506. )
  507. ax.set_xticks([0, 1])
  508. ax.set_xticklabels(["Dorsal", "Ventral"], fontsize=16)
  509. ax.set_xlim([-0.2, 1.2])
  510. if ylim:
  511. ax.set_ylim(ylim)
  512. ax.spines["right"].set_visible(False)
  513. ax.spines["top"].set_visible(False)
  514. ax.set_ylabel(value_col.replace("_", " "))
  515. ax.set_title(title)
  516. savefig(fig, stem)
  518. plot_transfer_source_style(csvs["fig6a_object_position_transfer_points"], "accuracy_proportion", "fig6a_object_position_transfer_source_style_demo", "Fig 6A object position", ylim=(0.425, 0.535))
  519. plot_transfer_source_style(csvs["fig6b_imagenet_category_transfer_points"], "max_accuracy_percent", "fig6b_imagenet_category_transfer_source_style_demo", "Fig 6B ImageNet category")
  522. # %% [markdown]
  523. # ## Fig 6C: Model fLoc Selectivity
  525. # %%
  526. plot_floc_bar_strip(
  527. csvs["fig6c_model_floc_selectivity_subject_bars"],
  528. csvs["fig6c_model_floc_selectivity_hemi_dots"],
  529. "fig6c_model_floc_selectivity_source_style_demo",
  530. "Fig 6C model fLoc selectivity",
  531. hue_order=["Dorsal", "Lateral", "Ventral"],
  532. )
  535. # %% [markdown]
  536. # ## Fig 6D: Face-Selective RF Eccentricity
  538. # %%
  539. df = csvs["fig6d_face_selective_rf_eccentricity_plot_points"].copy()
  540. sns.set_theme(style="ticks")
  541. fig, ax = plt.subplots(figsize=(4, 12))
  542. order = ["Lateral", "Ventral"]
  543. palette = [ROI_COLORS[1], ROI_COLORS[2]]
  545. sns.stripplot(
  546. x="stream", y="eccen", hue="stream", jitter=0.1, linewidth=0.75, edgecolor="w",
  547. palette=palette, data=df[(df["hemi"] == "rh") & (df["stream"].isin(order))],
  548. order=order, dodge=True, size=9, ax=ax,
  549. )
  550. sns.stripplot(
  551. x="stream", y="eccen", hue="stream", jitter=0.2, linewidth=1, edgecolor="w", marker="^",
  552. palette=palette, data=df[(df["hemi"] == "lh") & (df["stream"].isin(order))],
  553. order=order, dodge=True, size=9, ax=ax,
  554. )
  555. sns.violinplot(
  556. x="stream", y="eccen", hue="stream", fill=True, linewidth=3, inner="box",
  557. saturation=0.9, palette=palette, data=df[df["stream"].isin(order)], order=order,
  558. dodge=False, ax=ax,
  559. )
  560. for collection in ax.collections:
  561. collection.set_alpha(0.85)
  562. ax.spines["right"].set_visible(False)
  563. ax.spines["top"].set_visible(False)
  564. ax.legend([], [], frameon=False)
  565. ax.set_xlabel("")
  566. ax.set_ylabel("RF eccentricity")
  567. ax.set_title("Fig 6D face-selective RF eccentricity")
  568. savefig(fig, "fig6d_face_selective_rf_eccentricity_source_style_demo")
  571. # %% [markdown]
  572. # ## S2A/S2B: Stream Correlations And Cross-ROI Fits
  574. # %%
  575. def plot_s2a(points, summary):
  576. data = points.copy()
  577. summary = summary.copy()
  578. pair_order = ["Ventral/Lateral", "Ventral/Parietal", "Lateral/Parietal"]
  579. comparison_order = ["Within", "Between"]
  580. colors = {"Within": "#bdbdbd", "Between": "#525252"}
  581. x = np.arange(len(pair_order))
  582. width = 0.34
  583. offsets = {"Within": -width / 2, "Between": width / 2}
  585. fig, ax = plt.subplots(figsize=(4.6, 3.4), facecolor="white")
  586. for comparison in comparison_order:
  587. vals = []
  588. errs = []
  589. for pair in pair_order:
  590. row = summary[(summary["stream_pair"] == pair) & (summary["comparison"] == comparison)].iloc[0]
  591. vals.append(float(row["mean_fisher_z"]))
  592. errs.append(float(row["sem_fisher_z"]))
  593. xpos = x + offsets[comparison]
  594. ax.bar(xpos, vals, yerr=errs, width=width, color=colors[comparison], edgecolor="black", linewidth=0.8, capsize=2, label=comparison)
  595. for idx, pair in enumerate(pair_order):
  596. subj_vals = data[(data["stream_pair"] == pair) & (data["comparison"] == comparison)]["fisher_z"].to_numpy(float)
  597. jitter = np.linspace(-0.06, 0.06, len(subj_vals)) if len(subj_vals) else []
  598. ax.scatter(np.full(len(subj_vals), xpos[idx]) + jitter, subj_vals, s=18, color="lightgray", edgecolor="black", linewidth=0.35, zorder=3)
  599. ax.set_xticks(x)
  600. ax.set_xticklabels(pair_order, rotation=25, ha="right")
  601. ax.set_ylabel("RSM correlation (Fisher z)")
  602. ax.legend(frameon=False)
  603. despine(ax)
  604. fig.tight_layout()
  605. savefig(fig, "s2a_within_between_stream_correlations_demo")
  608. def plot_s2b(points, summary):
  609. data = points.copy()
  610. summary = summary.copy()
  611. roi_order = ["Ventral", "Lateral", "Parietal"]
  612. model_order = ["same_roi", "other_roi1", "other_roi2"]
  613. model_labels = {"same_roi": "Same ROI", "other_roi1": "Other ROI 1", "other_roi2": "Other ROI 2"}
  614. colors = {"same_roi": "#f0f0f0", "other_roi1": "#969696", "other_roi2": "#252525"}
  615. x = np.arange(len(roi_order))
  616. width = 0.22
  617. offsets = {"same_roi": -width, "other_roi1": 0, "other_roi2": width}
  619. fig, ax = plt.subplots(figsize=(4.5, 4.4), facecolor="white")
  620. for model in model_order:
  621. vals = []
  622. errs = []
  623. for roi in roi_order:
  624. row = summary[(summary["roi"] == roi) & (summary["Model"] == model)].iloc[0]
  625. vals.append(float(row["mean_corrected"]))
  626. errs.append(float(row["sem_corrected"]))
  627. xpos = x + offsets[model]
  628. ax.bar(xpos, vals, yerr=errs, width=width, color=colors[model], edgecolor="black", linewidth=0.8, capsize=2, label=model_labels[model])
  629. for idx, roi in enumerate(roi_order):
  630. subj_vals = data[(data["roi"] == roi) & (data["Model"] == model)]["corrected"].to_numpy(float)
  631. jitter = np.linspace(-0.035, 0.035, len(subj_vals)) if len(subj_vals) else []
  632. ax.scatter(np.full(len(subj_vals), xpos[idx]) + jitter, subj_vals, s=18, color="lightgray", edgecolor="blue", linewidth=0.35, zorder=3)
  633. ax.axhline(1, color="black", linestyle="--", linewidth=1.0)
  634. ax.set_ylim(0, 1.05)
  635. ax.set_xticks(x)
  636. ax.set_xticklabels(roi_order, rotation=25, ha="right")
  637. ax.set_ylabel("Corrected R-squared")
  638. ax.legend(frameon=False, loc="upper right")
  639. despine(ax)
  640. fig.tight_layout()
  641. savefig(fig, "s2b_cross_roi_subject_to_subject_demo")
  643. plot_s2a(sheets["s2a_points"], sheets["s2a_summary"])
  644. plot_s2b(sheets["s2b_points"], sheets["s2b_summary"])
  648. # %% [markdown]
  649. # ## S3C: Voxel-To-Voxel Accuracy, Checkpoint 0
  651. # %%
  652. df = sheets["s3c_points"].copy()
  653. df["display_roi"] = df["roi"].replace({"Parietal": "Dorsal"})
  654. order = ["Dorsal", "Lateral", "Ventral"]
  656. fig, ax = plt.subplots(figsize=(3.2, 4.2), facecolor="white")
  657. x = np.arange(len(order))
  658. for idx, roi in enumerate(order):
  659. vals = df.loc[df["display_roi"] == roi, "accuracy_percent"].to_numpy(float)
  660. ax.bar(idx, np.nanmean(vals), yerr=np.nanstd(vals, ddof=1), width=0.62,
  661. color=STREAM_COLORS[roi], alpha=0.75, edgecolor="white", linewidth=0.8)
  662. jitter = np.linspace(-0.12, 0.12, len(vals)) if len(vals) else []
  663. ax.scatter(np.full(len(vals), idx) + jitter, vals, s=26, color="lightgray", edgecolor="black", linewidth=0.4, zorder=3)
  664. ax.axhline(33.3, color="black", linestyle="--", linewidth=1.0)
  665. ax.set_xticks(x)
  666. ax.set_xticklabels(order, rotation=30, ha="right")
  667. ax.set_ylabel("Voxel assignment correspondence (%)")
  668. ax.set_ylim(0, 100)
  669. despine(ax)
  670. fig.tight_layout()
  671. savefig(fig, "s3c_voxel2voxel_accuracy_checkpoint0_demo")
  674. # %% [markdown]
  675. # ## S5A/S5B: TDANN sw0.25 vs V1-Control
  677. # %%
  678. def plot_s5_paired(points, stem, ylabel):
  679. data = points.copy()
  680. data["ROI"] = data["ROI"].replace({"Parietal": "Dorsal"})
  681. subj = data.groupby(["condition", "subject", "ROI"], as_index=False)["result"].mean()
  682. cond_order = ["TDANN_sw0.25", "TDANN_sw0.25_v1_control"]
  683. cond_labels = ["TDANN\nsw0.25", "V1-control\nsw0.25"]
  684. fig, axes = plt.subplots(1, 3, figsize=(7.2, 3.3), sharey=True, facecolor="white")
  685. for ax, roi in zip(axes, STREAM_ORDER):
  686. wide = subj[subj["ROI"] == roi].pivot(index="subject", columns="condition", values="result").dropna()
  687. vals = [wide[c].to_numpy(float) for c in cond_order]
  688. means = [np.nanmean(v) for v in vals]
  689. sems = [np.nanstd(v, ddof=1) / np.sqrt(np.sum(np.isfinite(v))) for v in vals]
  690. ax.bar([0, 1], means, yerr=sems, width=0.62, color=["#d7d7d7", "#9e9e9e"],
  691. edgecolor="#595959", linewidth=0.9, capsize=2, zorder=1)
  692. for _, row in wide.iterrows():
  693. ax.plot([0, 1], [row[cond_order[0]], row[cond_order[1]]], color="black", alpha=0.22, linewidth=0.7, zorder=2)
  694. ax.scatter([0, 1], [row[cond_order[0]], row[cond_order[1]]], color=STREAM_COLORS[roi],
  695. edgecolor="white", linewidth=0.4, s=24, zorder=3)
  696. ax.set_title(roi)
  697. ax.set_xticks([0, 1])
  698. ax.set_xticklabels(cond_labels, rotation=25, ha="right")
  699. despine(ax)
  700. axes[0].set_ylabel(ylabel)
  701. fig.tight_layout()
  702. savefig(fig, stem)
  704. plot_s5_paired(sheets["s5_spat_points"], "s5a_tdann_v1_control_spatial_demo", "Spatial similarity [r]")
  705. plot_s5_paired(sheets["s5_func_points"], "s5b_tdann_v1_control_functional_demo", "Noise-corrected correlation")
  708. # %% [markdown]
  709. # ## S6A/S6B: ViT-Control Correspondence Panels
  711. # %%
  712. S6A_ORDER = [
  713. "MB_RN50_v2_detection",
  714. "MB_RN50_v2_vit_control_detection",
  715. "MB_RN50_v2_clip",
  716. "MB_RN50_v2_vit_control_clip",
  717. "MB_RN50_v2_categorization",
  718. "MB_RN50_v2_vit_control_categorization",
  719. ]
  720. S6B_ORDER = [
  721. "MB_RN50_v2_detection",
  722. "MB_RN50_v2_detection_vit_control",
  723. "MB_RN50_v2_clip",
  724. "MB_RN50_v2_clip_vit_control",
  725. "MB_RN50_v2_categorization",
  726. "MB_RN50_v2_categorization_vit_control",
  727. ]
  728. MODEL_LABELS = {
  729. "MB_RN50_v2_detection": "Det.",
  730. "MB_RN50_v2_vit_control_detection": "Det.\nV1",
  731. "MB_RN50_v2_detection_vit_control": "Det.\nV1",
  732. "MB_RN50_v2_clip": "CLIP",
  733. "MB_RN50_v2_vit_control_clip": "CLIP\nV1",
  734. "MB_RN50_v2_clip_vit_control": "CLIP\nV1",
  735. "MB_RN50_v2_categorization": "Cat.",
  736. "MB_RN50_v2_vit_control_categorization": "Cat.\nV1",
  737. "MB_RN50_v2_categorization_vit_control": "Cat.\nV1",
  738. }
  739. MODEL_COLORS = ["#80b380", "#80b380", "#6666cc", "#6666cc", "#cc8066", "#cc8066"]
  740. MODEL_ALPHA = [1.0, 0.48, 1.0, 0.48, 1.0, 0.48]
  743. def plot_s6(points, noise, model_order, stem, ylabel, chance=None, ylim=None):
  744. data = points.copy()
  745. data["ROIS"] = data["ROIS"].replace({"Parietal": "Dorsal"})
  746. noise = noise.copy()
  747. noise["ROI"] = noise["ROI"].replace({"Parietal": "Dorsal"})
  748. fig, axes = plt.subplots(1, 3, figsize=(9.0, 3.2), sharey=True, facecolor="white")
  749. x = np.array([0, 1, 2.15, 3.15, 4.3, 5.3])
  750. for ax, roi in zip(axes, STREAM_ORDER):
  751. sub = data[data["ROIS"] == roi]
  752. for idx, model in enumerate(model_order):
  753. vals = sub.loc[sub["model_type"] == model, "result"].to_numpy(float)
  754. ax.bar(x[idx], np.nanmean(vals), yerr=np.nanstd(vals, ddof=1), width=0.82,
  755. color=MODEL_COLORS[idx], alpha=MODEL_ALPHA[idx], edgecolor="none", capsize=1.8, zorder=1)
  756. lh = sub[(sub["model_type"] == model) & (sub["hemi"] == "lh")]["result"].to_numpy(float)
  757. rh = sub[(sub["model_type"] == model) & (sub["hemi"] == "rh")]["result"].to_numpy(float)
  758. ax.scatter(np.full(len(lh), x[idx]) - 0.08, lh, marker="^", s=11, color="none", edgecolor="black", linewidth=0.35, zorder=3)
  759. ax.scatter(np.full(len(rh), x[idx]) + 0.08, rh, marker="o", s=11, color="none", edgecolor="black", linewidth=0.35, zorder=3)
  760. nvals = noise.loc[noise["ROI"] == roi, "result"].to_numpy(float)
  761. if len(nvals):
  762. ax.axhspan(np.nanmean(nvals) - np.nanstd(nvals, ddof=1), np.nanmean(nvals) + np.nanstd(nvals, ddof=1),
  763. color="lightgray", alpha=0.35, zorder=0)
  764. if chance is not None:
  765. ax.axhline(chance, color="black", linestyle=":", linewidth=1.0)
  766. ax.set_title(roi)
  767. ax.set_xticks(x)
  768. ax.set_xticklabels([MODEL_LABELS[m] for m in model_order], rotation=0)
  769. if ylim:
  770. ax.set_ylim(*ylim)
  771. despine(ax)
  772. axes[0].set_ylabel(ylabel)
  773. fig.tight_layout()
  774. savefig(fig, stem)
  776. plot_s6(sheets["s6a_points"], sheets["s6a_noise"], S6A_ORDER, "s6a_vit_control_stream_assignment_demo", "Stream assignment (%)", chance=33, ylim=(0, 100))
  777. plot_s6(sheets["s6b_points"], sheets["s6b_noise"], S6B_ORDER, "s6b_vit_control_functional_correspondence_demo", "Functional correspondence [r]", ylim=(0, 0.36))
  780. # %% [markdown]
  781. # ## S10: TDANN CKA Across Spatial Weights
  783. # %%
  784. summary = sheets["s10_summary"].copy()
  785. layer_order = [
  786. "base_model.maxpool", "base_model.layer1.0", "base_model.layer1.1",
  787. "base_model.layer2.0", "base_model.layer2.1", "base_model.layer3.0", "base_model.layer3.1",
  788. "base_model.layer4.0", "base_model.layer4.1",
  789. ]
  790. pretty = [l.replace("base_model.", "").replace("layer", "layer ") for l in layer_order]
  791. ceiling = summary[summary["comparison"] == "ceiling pooled (sw0.0 + sw0.25)"].set_index("layer").reindex(layer_order)
  792. cross = summary[summary["comparison"] == "cross-weight (sw0.0 vs sw0.25)"].set_index("layer").reindex(layer_order)
  793. x = np.arange(len(layer_order))
  795. fig, ax = plt.subplots(figsize=(5.7, 2.8), facecolor="white")
  796. c_mean = ceiling["mean_cka"].to_numpy(float)
  797. c_sem = ceiling["sem_cka"].to_numpy(float)
  798. x_mean = cross["mean_cka"].to_numpy(float)
  799. ax.plot(x, c_mean, color="#2F6DAE", lw=1.4, label="Noise ceiling")
  800. ax.fill_between(x, c_mean - c_sem, c_mean + c_sem, color="#2F6DAE", alpha=0.18, linewidth=0)
  801. ax.plot(x, x_mean, color="#C55A11", marker="o", ms=4, lw=2.0, label="Cross-weight")
  802. ax.set_xticks(x)
  803. ax.set_xticklabels(pretty, rotation=35, ha="right")
  804. ax.set_ylabel("Linear CKA")
  805. ax.set_xlabel("TDANN layer")
  806. ax.set_ylim(0, 1)
  807. ax.grid(axis="y", color="#d9d9d9", linewidth=0.7, alpha=0.7)
  808. ax.legend(frameon=False, loc="lower left")
  809. despine(ax)
  810. fig.tight_layout()
  811. savefig(fig, "s10_cka_tdann_spatial_weight_demo")
  814. # %% [markdown]
  815. # ## S12A: Functional Correspondence Across Spatial Weights
  817. # %%
  818. def plot_s12a(points, bands):
  819. data = points.copy()
  820. data["stream"] = data["stream"].replace({"Parietal": "Dorsal"})
  821. bands = bands.copy()
  822. bands["stream"] = bands["stream"].replace({"Parietal": "Dorsal"})
  823. data["spatial_weight"] = pd.to_numeric(data["spatial_weight"])
  824. versions = [v for v in ["self_supervised", "supervised"] if v in set(data["version"])]
  825. palette = {"self_supervised": "#720298", "supervised": "#B59410"}
  826. labels = {"self_supervised": "SimCLR", "supervised": "Supervised"}
  827. fig, axes = plt.subplots(1, 3, figsize=(8.4, 3.0), sharey=True, facecolor="white")
  828. for ax, stream in zip(axes, STREAM_ORDER):
  829. sub = data[data["stream"] == stream]
  830. for version in versions:
  831. vsub = sub[sub["version"] == version]
  832. grouped = vsub.groupby("spatial_weight")["corr"].agg(["mean", "std", "count"]).reset_index().sort_values("spatial_weight")
  833. grouped["sem"] = grouped["std"] / np.sqrt(grouped["count"])
  834. ax.plot(grouped["spatial_weight"], grouped["mean"], marker=".", markersize=9, lw=1.5,
  835. color=palette.get(version, "black"), label=labels.get(version, version))
  836. ax.fill_between(grouped["spatial_weight"].to_numpy(float),
  837. (grouped["mean"] - grouped["sem"]).to_numpy(float),
  838. (grouped["mean"] + grouped["sem"]).to_numpy(float),
  839. color=palette.get(version, "black"), alpha=0.12, linewidth=0)
  840. band = bands[bands["stream"] == stream]
  841. if len(band):
  842. mean = float(band["mean"].iloc[0])
  843. sd = float(band["std"].iloc[0])
  844. ax.axhspan(mean - sd, mean + sd, color="lightgray", alpha=0.45, zorder=0)
  845. ax.set_xscale("symlog", linthresh=0.09)
  846. ax.set_xlim(-0.01, 50)
  847. ax.set_xticks([0, 0.1, 0.25, 0.5, 1.25, 2.5, 25])
  848. ax.set_xticklabels(["0", "0.1", "", "", "1.25", "", "25"])
  849. ax.set_title(stream)
  850. ax.set_xlabel("Spatial weight")
  851. despine(ax)
  852. axes[0].set_ylabel("Functional correspondence [r]")
  853. axes[-1].legend(frameon=False, loc="upper right")
  854. fig.tight_layout()
  855. savefig(fig, "s12a_functional_correspondence_by_weight_demo")
  857. plot_s12a(sheets["s12a_points"], sheets["s12a_brain_band"])
  860. # %% [markdown]
  861. # ## S11: HVM Transfer, Selected Tasks
  863. # %%
  864. def plot_s11(seedavg):
  865. data = seedavg.copy()
  866. data["spatial_weight"] = pd.to_numeric(data["spatial_weight"])
  867. task_order = ["categorization", "combined_position", "rotation_xz", "size"]
  868. task_titles = {
  869. "categorization": "Categorization",
  870. "combined_position": "Position",
  871. "rotation_xz": "Rotation XZ",
  872. "size": "Size",
  873. }
  874. ylabels = {
  875. "categorization": "ImageNet accuracy (%)",
  876. "combined_position": "Position accuracy",
  877. "rotation_xz": "Rotation correlation [r]",
  878. "size": "Size correlation [r]",
  879. }
  880. fig, axes = plt.subplots(1, 4, figsize=(11.5, 3.0), facecolor="white")
  881. for ax, task in zip(axes, task_order):
  882. sub = data[data["task"] == task]
  883. for stream in STREAM_ORDER:
  884. grouped = (
  885. sub[sub["stream"] == stream]
  886. .groupby("spatial_weight")["result"]
  887. .agg(["mean", "std", "count"])
  888. .reset_index()
  889. .sort_values("spatial_weight")
  890. )
  891. grouped["sem"] = grouped["std"] / np.sqrt(grouped["count"])
  892. ax.plot(
  893. grouped["spatial_weight"],
  894. grouped["mean"],
  895. marker=".",
  896. markersize=9,
  897. lw=1.6,
  898. color=STREAM_COLORS[stream],
  899. label=stream,
  900. )
  901. ax.fill_between(
  902. grouped["spatial_weight"].to_numpy(float),
  903. (grouped["mean"] - grouped["sem"]).to_numpy(float),
  904. (grouped["mean"] + grouped["sem"]).to_numpy(float),
  905. color=STREAM_COLORS[stream],
  906. alpha=0.12,
  907. linewidth=0,
  908. )
  909. ax.set_xscale("symlog", linthresh=0.09)
  910. ax.set_xlim(-0.01, 50)
  911. ax.set_xticks([0, 0.1, 0.25, 0.5, 1.25, 2.5, 25])
  912. ax.set_xticklabels(["0", "0.1", "", "", "1.25", "", "25"])
  913. ax.set_title(task_titles[task])
  914. ax.set_xlabel("Spatial weight")
  915. ax.set_ylabel(ylabels[task])
  916. despine(ax)
  917. axes[-1].legend(frameon=False, loc="best")
  918. fig.tight_layout()
  919. savefig(fig, "s11_hvm_transfer_selected_tasks_demo")
  921. plot_s11(sheets["s11_seedavg"])
  926. # %% [markdown]
  927. # ## S12B: Supervised vs SimCLR Across Spatial Weights
  929. # %%
  930. def plot_s12b(points, bands):
  931. data = points.copy()
  932. data["spatial_weight"] = pd.to_numeric(data["spatial_weight"])
  933. bands = bands.copy()
  934. palette = {"simCLR": "#720298", "supervised": "#B59410"}
  935. labels = {"simCLR": "SimCLR", "supervised": "Supervised"}
  936. versions = [v for v in ["simCLR", "supervised"] if v in set(data["version"])]
  937. fig, axes = plt.subplots(1, 3, figsize=(8.4, 3.0), sharey=True, facecolor="white")
  938. for ax, stream in zip(axes, STREAM_ORDER):
  939. sub = data[data["stream"] == stream]
  940. for version in versions:
  941. grouped = (
  942. sub[sub["version"] == version]
  943. .groupby("spatial_weight")["means"]
  944. .agg(["mean", "std", "count"])
  945. .reset_index()
  946. .sort_values("spatial_weight")
  947. )
  948. grouped["sem"] = grouped["std"] / np.sqrt(grouped["count"])
  949. ax.plot(
  950. grouped["spatial_weight"],
  951. grouped["mean"],
  952. marker=".",
  953. markersize=9,
  954. lw=1.5,
  955. color=palette[version],
  956. label=labels[version],
  957. )
  958. ax.fill_between(
  959. grouped["spatial_weight"].to_numpy(float),
  960. (grouped["mean"] - grouped["sem"]).to_numpy(float),
  961. (grouped["mean"] + grouped["sem"]).to_numpy(float),
  962. color=palette[version],
  963. alpha=0.12,
  964. linewidth=0,
  965. )
  966. band = bands[bands["stream"] == stream]
  967. if len(band):
  968. mean = float(band["mean"].iloc[0])
  969. sd = float(band["std"].iloc[0])
  970. ax.axhspan(mean - sd, mean + sd, color="lightgray", alpha=0.45, zorder=0)
  971. ax.set_xscale("symlog", linthresh=0.09)
  972. ax.set_xlim(-0.01, 50)
  973. ax.set_xticks([0, 0.1, 0.25, 0.5, 1.25, 2.5, 25])
  974. ax.set_xticklabels(["0", "0.1", "", "", "1.25", "", "25"])
  975. ax.set_title(stream)
  976. ax.set_xlabel("Spatial weight")
  977. despine(ax)
  978. axes[0].set_ylabel("Noise-corrected correlation")
  979. axes[-1].legend(frameon=False, loc="upper right")
  980. fig.tight_layout()
  981. savefig(fig, "s12b_supervised_vs_simclr_by_weight_demo")
  983. plot_s12b(sheets["s12b_points"], sheets["s12b_brain_band"])
  987. # %%
  988. print(f"Wrote combined demo plots to {OUT}")
  989. print("PNG files:", len(list(OUT.glob("*.png"))))
  990. print("PDF files:", len(list(OUT.glob("*.pdf"))))

demo_plot_combined_source_data.ipynb at commit d85997e, under Apache-2.0 · at the source

Overview

Authors: Dawn Finzi1,2, Eshed Margalit3, Kendrick Kay4, Daniel L K Yamins1,2,5, Kalanit Grill-Spector1,5
ORCID iDs: Dawn Finzi, Eshed Margalit, Kendrick Kay, Kalanit Grill-Spector
  1. Department of Psychology, Stanford University, Stanford, CA USA
  2. Department of Computer Science, Stanford University, Stanford, CA USA
  3. Neurosciences Graduate Program, Stanford University, Stanford, CA USA
  4. Center for Magnetic Resonance Research, Department of Radiology, University of Minnesota, Minneapolis, MN USA
  5. Wu Tsai Neurosciences Institute, Stanford University, Stanford, CA USA
Institutions: Stanford University (United States); University of Minnesota (United States)
Journal: Nature communications, volume 17, issue 1, article 9762
Dates: received 19 April 2026; accepted 20 July 2026; published online 13 August 2026
Type: Research article · Language: English
License: CC BY-NC-ND
Identifiers: DOI 10.1038/s41467-026-76098-y · PMID 42733086 · PMCID PMC13572372 · OpenAlex W7202347126
Open access: gold, a free copy (OpenAlex)
Status: code verified
Categories: human (organism)
Methods: Connectivity, Statistics, Smoothing, state filtering, decompositions, Machine learning, Preprocessing, fMRI & imaging
Keywords: Computational neuroscience, Visual system, Computer science
MeSH: Neural Networks, Computer*, Visual Cortex*, Visual Pathways*, Brain Mapping, Computer Simulation, Humans, Magnetic Resonance Imaging, Models, Neurological, Visual Perception (* major topic)
Topic: Visual perception and processing mechanisms (Cognitive Neuroscience, Neuroscience), according to OpenAlex
Funding: NEI NIH HHS (R01 EY023915); National Science Foundation (NSF) (IIS-1822929); U.S. Department of Health & Human Services | National Institutes of Health (NIH) (R01EY023915)
Citations: not cited yet (Europe PMC); 119 references in the paper

Abstract

The abstract is not reproduced here: the paper's license (CC BY-NC-ND) does not allow it. Read it in the paper, at the publisher or on Europe PMC.

Repositories

Its files are read in the Code ↔ Paper reader above, with 16 matches between paragraphs and lines of code.

dawnfinzi/spacestream

License: Apache-2.0
State: the link answers, verified on 27 September 2026
Evidence: files inventoried
Commit: d85997e327886f119441cc6a8355584373878d29, 26 September 2026
Languages: Python (49), Jupyter (12), MATLAB (9)
Size: 73 files, 70 scripts
Software Heritage: not archived
Found in: “Code availability”
Holds: README, license file, environment (setup.py), 12 notebooks
Not found: CITATION.cff, tests, continuous integration, documentation
Tools: NumPy (49 files), PyTorch (23 files), h5py (21 files), SciPy (17 files), pandas (14 files), NiBabel (12 files), Matplotlib (11 files), scikit-learn (9 files), seaborn (7 files), Pillow (3 files), statsmodels (3 files), scikit-image (2 files), Hugging Face Transformers (2 files), xarray (2 files), FreeSurfer (1 file), Statistics and Machine Learning Toolbox (1 file)
Availability: 1 check, the latest on 27 September 2026: the link answers
  • 27 September 2026: the link answers
72 files
At the source: github.com/dawnfinzi/spacestream

neuroailab/TDANN

License: none: the authors keep all their rights
State: the link answers, verified on 27 September 2026
Evidence: files inventoried
Commit: 80c585df69dcb831d1d6802e0776a8cf25d8cfef, 18 October 2024
Languages: Python (92), Jupyter (2)
Size: 479 files, 94 scripts
Software Heritage: not archived
Found in: “Code availability”
Holds: README, environment (requirements.txt, setup.cfg, setup.py), 2 notebooks
Not found: license file, CITATION.cff, tests, continuous integration, documentation
Tools: NumPy (47 files), PyTorch (46 files), SciPy (15 files), Matplotlib (13 files), pandas (7 files), scikit-image (7 files), xarray (5 files), Pillow (4 files), scikit-learn (4 files), h5py (3 files)
Availability: 1 check, the latest on 27 September 2026: the link answers
  • 27 September 2026: the link answers
95 files
At the source: github.com/neuroailab/TDANN

Zenodo 20753397

License: Apache-2.0
State: the link answers, verified on 27 September 2026
Evidence: files inventoried
Size: 1 file
Software Heritage: not checked
Found in: the references
Not found: README, license file, CITATION.cff, environment file, tests, continuous integration, documentation
Tools: NumPy (49 files), PyTorch (23 files), h5py (21 files), SciPy (17 files), pandas (14 files), NiBabel (12 files), Matplotlib (11 files), scikit-learn (9 files), seaborn (7 files), Pillow (3 files), statsmodels (3 files), scikit-image (2 files), Hugging Face Transformers (2 files), xarray (2 files), FreeSurfer (1 file), Statistics and Machine Learning Toolbox (1 file)
Availability: 1 check, the latest on 27 September 2026: the link answers (HTTP 200)
  • 27 September 2026: the link answers (HTTP 200)
72 files
At the source:

Code availability statement

The paper has a code availability statement. Its license (CC BY-NC-ND) does not allow reproducing it here; in short, from what the harvester recognized in it:

Read it in the paper: doi.org/10.1038/s41467-026-76098-y.

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:

  • 3 repositories of the authors' code, each at its verified commit, with its license and how the link was found in the paper;
  • 234 scripts, each with its path and the digest of its content;
  • 16 matches 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 statement

The paper has a data availability statement. Its license (CC BY-NC-ND) does not allow reproducing it here; in short, from what the harvester recognized in it:

Read it in the paper: doi.org/10.1038/s41467-026-76098-y.

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, 5 authors, 3 keywords, 9 MeSH terms, 3 funders, 88 references.

Cite

This paper

Finzi, D., Margalit, E., Kay, K., Yamins, D. L. K., & Grill-Spector, K. (2026). A single computational objective can produce specialization of streams in visual cortex. Nature communications, 17(1), 9762. https://doi.org/10.1038/s41467-026-76098-y

BibTeX

@article{finzi2026single,
author = {Finzi, Dawn and Margalit, Eshed and Kay, Kendrick and Yamins, Daniel L K and Grill-Spector, Kalanit},
title = {{A single computational objective can produce specialization of streams in visual cortex}},
journal = {Nature communications},
year = {2026},
month = aug,
volume = {17},
number = {1},
pages = {9762},
publisher = {Nature Publishing Group},
issn = {2041-1723},
doi = {10.1038/s41467-026-76098-y},
url = {https://doi.org/10.1038/s41467-026-76098-y},
pmid = {42733086},
pmcid = {PMC13572372}
}

RIS

TY - JOUR
AU - Finzi, Dawn
AU - Margalit, Eshed
AU - Kay, Kendrick
AU - Yamins, Daniel L K
AU - Grill-Spector, Kalanit
TI - A single computational objective can produce specialization of streams in visual cortex
T2 - Nature communications
J2 - Nat Commun
PY - 2026
DA - 2026/08/13
VL - 17
IS - 1
SP - 9762
SN - 2041-1723
PB - Nature Publishing Group
DO - 10.1038/s41467-026-76098-y
UR - https://doi.org/10.1038/s41467-026-76098-y
LA - en
ER -

CSL-JSON

{
"id": "10.1038/s41467-026-76098-y",
"type": "article-journal",
"title": "A single computational objective can produce specialization of streams in visual cortex",
"container-title": "Nature communications",
"author": [
{
"family": "Finzi",
"given": "Dawn"
},
{
"family": "Margalit",
"given": "Eshed"
},
{
"family": "Kay",
"given": "Kendrick"
},
{
"family": "Yamins",
"given": "Daniel L K"
},
{
"family": "Grill-Spector",
"given": "Kalanit"
}
],
"container-title-short": "Nat Commun",
"volume": "17",
"issue": "1",
"page": "9762",
"DOI": "10.1038/s41467-026-76098-y",
"PMID": "42733086",
"PMCID": "PMC13572372",
"ISSN": "2041-1723",
"publisher": "Nature Publishing Group",
"URL": "https://doi.org/10.1038/s41467-026-76098-y",
"language": "en",
"issued": {
"date-parts": [
[
2026,
8,
13
]
]
}
}

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/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: Hugging Face Transformers, h5py, Pillow, 9 other tools, 14 references
[2] doi:10.1523/jneurosci.0038-26.2026 [code]
Multidimensional Feature Tuning in Category Selective Areas of Human Visual Cortex.
Journal: The Journal of neuroscience : the official journal of the Society for Neuroscience
In common: h5py, Pillow, NiBabel, 8 other tools, 12 references
[3] doi:10.7554/elife.107933 [code]
Modality-agnostic decoding of vision and language from fMRI.
Journal: eLife
In common: Hugging Face Transformers, FreeSurfer, scikit-image, 11 other tools, 5 references
[4] doi:10.1162/imag.a.1309 [code]
Probing the content of semantic representations in body-selective regions.
Journal: Imaging neuroscience (Cambridge, Mass.)
In common: h5py, Pillow, NiBabel, 8 other tools, 8 references
[5] doi:10.1038/s41467-026-73153-6 [code]
Latent neural architecture organising shared aesthetic evaluations of visual artworks.
Journal: Nature communications
In common: h5py, Pillow, PyTorch, 7 other tools, 8 references
[6] doi:10.1038/s42003-026-10843-3 [code]
Self-supervised learning yields representational signatures of category-selective cortex.
Journal: Communications biology
In common: 13 references
[7] doi:10.1038/s41467-026-73938-9 [code]
Higher visual areas act like domain-general filters with strong selectivity and functional specialization.
Journal: Nature communications
In common: scikit-image, Pillow, NiBabel, 5 other tools, 8 references
[8] doi:10.1038/s41597-026-07248-6 [code]
A large-scale fMRI dataset for vision-language semantic association.
Journal: Scientific data
In common: FreeSurfer, scikit-image, h5py, 8 other tools, 5 references
[9] doi:10.7554/elife.105953 [code]
Top-down feedback in deep neural networks leads to functional differences during audiovisual integration.
Journal: eLife
In common: h5py, Pillow, PyTorch, 6 other tools, 7 references
[10] doi:10.1038/s41597-026-06869-1 [code]
Individual Brain Charting: fifth release of high-resolution fMRI data for cognitive mapping.
Journal: Scientific data
In common: FreeSurfer, scikit-image, Pillow, 7 other tools, 2 references, author Kalanit Grill-Spector

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.