In [1]:

import math
from pathlib import Path

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from wordcloud import WordCloud

In [2]:

permutation_csv = Path("results/rsa_permutation_results.csv")
null_dist_dir = Path("results/null_distributions")
out_dir = Path("./")
out_dir.mkdir(parents=True, exist_ok=True)

df = pd.read_csv(permutation_csv)

monkey = "monkeyF"
roi = "IT"

subset = df[(df["monkey"] == monkey) & (df["roi"] == roi)].sort_values("noise_level")
noise_levels = subset["noise_level"].tolist()

noise_levels = [0.01, 0.1, 0.25, 0.35, 0.5, 0.65, 0.75, 0.90, 0.99]

n_plots = len(noise_levels)
ncols = 3
nrows = 3

fig, axes = plt.subplots(
    nrows=nrows, ncols=ncols, figsize=(4.5 * ncols, 3.5 * nrows), squeeze=False
)
axes = axes.flatten()

subject_name = monkey.replace("monkey", "")
fig.suptitle(
    f"Permutation Null Distributions: Macaque {subject_name} | {roi}",
    fontsize=18,
    fontweight="bold",
    y=1.02,
)

for idx_plot, noise in enumerate(noise_levels):
    ax = axes[idx_plot]
    row_data = subset[subset["noise_level"] == noise].iloc[0]

    true_score = float(row_data["true_alignment_score"])
    p_value = float(row_data["p_value"])

    npy_filename = f"null_dist_{monkey}_{roi}_noise_{noise:.2f}.npy"
    npy_path = null_dist_dir / npy_filename

    if not npy_path.exists():
        ax.text(0.5, 0.5, f"Missing:\n{npy_filename}", ha="center", va="center")
        ax.set_title(f"Noise: {noise:.2f}")
        ax.set_xticks([])
        ax.set_yticks([])
        continue

    null_dist = np.load(npy_path)

    ax.hist(
        null_dist,
        bins=30,
        density=False,
    )

    ax.axvline(
        x=true_score,
        color="red",
        linewidth=2.5,
        linestyle="--",
        label=f"True Score ({true_score:.3f})",
    )

    ax.set_title(f"Stable Diffusion Noise: {noise:.2f}", fontsize=12)
    ax.set_xlabel("RSA Score (rho-a)", fontsize=10)
    ax.set_ylabel("Frequency", fontsize=10)
    ax.spines["top"].set_visible(False)
    ax.spines["right"].set_visible(False)

    p_text = "p < 0.001" if p_value < 0.001 else f"p = {p_value:.3f}"
    ax.text(
        0.05,
        0.90,
        p_text,
        transform=ax.transAxes,
        fontsize=11,
        fontweight="bold",
        color="#333333",
        bbox=dict(facecolor="white", alpha=0.6, edgecolor="none"),
    )
    ax.legend(loc="upper right", fontsize=9)

for empty_idx in range(n_plots, len(axes)):
    fig.delaxes(axes[empty_idx])

fig.tight_layout()
plt.show()

Figure 1: The null distribution for the permutation test, visualizing 9 of the 21 noise levels compared with the IT region of Macaque F. The red dashed line represents the RSA score between the biological brain and the diffusion model. The figure shows statistically significant evidence (\(p < 0.01\)) of a non-zero relationship between the two representations across all noise levels

In [3]:

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from pathlib import Path

# Load data
permutation_csv = Path("results/rsa_permutation_results.csv")
scores_df = pd.read_csv(permutation_csv)

monkeys = ["monkeyF", "monkeyN"]
ROI_ORDER = ["V1", "V4", "IT"]

# Pre-calculate noise levels from the first monkey to set up figure size
temp_data = scores_df[scores_df["monkey"] == "monkeyF"].pivot_table(
    index="roi", columns="noise_level", values="true_alignment_score", aggfunc="mean"
)
noise_levels = sorted(temp_data.columns)

fmt_string = "{:.3f}"
cbar_label = "RSA Score (Spearman rho-a)"

# Figure setup (2 rows, 1 column)
fig, axes = plt.subplots(nrows=2, ncols=1, figsize=(0.8 * len(noise_levels) + 3, 7))

for idx, monkey in enumerate(monkeys):
    ax = axes[idx]
    monkey_data = scores_df[scores_df["monkey"] == monkey]

    # Create the pivot table for the heatmap
    heat = monkey_data.pivot_table(
        index="roi", columns="noise_level", values="true_alignment_score", aggfunc="mean"
    )
    heat = heat.reindex(index=[r for r in ROI_ORDER if r in heat.index])
    heat = heat[noise_levels]
    
    masked_data = np.ma.masked_invalid(heat.to_numpy(dtype=float))
    im = ax.imshow(masked_data, aspect="auto", cmap="magma")

    ax.set_xticks(range(len(noise_levels)))
    # Only show x-axis labels on the bottom plot to keep it clean, or both if preferred
    ax.set_xticklabels([f"{n:.2f}" for n in noise_levels], rotation=45)
    
    ax.set_yticks(range(len(heat.index)))
    ax.set_yticklabels(heat.index)

    if idx == 1:
        ax.set_xlabel("Normalized Noise Level", fontsize=12)
    ax.set_ylabel("Brain Region (ROI)", fontsize=12)

    subject_name = monkey.replace("monkey", "")
    ax.set_title(f"Macaque {subject_name}", fontsize=14, pad=10)

    fig.colorbar(im, ax=ax, label=cbar_label)

    # Calculate min/max just for text contrast
    heat_min = heat.min().min()
    heat_max = heat.max().max()

    # Add text annotations
    for i, area in enumerate(heat.index):
        for j, noise in enumerate(noise_levels):
            value = heat.loc[area, noise]
            if pd.isna(value):
                continue

            norm_val = (value - heat_min) / (heat_max - heat_min) if heat_max != heat_min else 0.5
            text_color = "black" if norm_val > 0.7 else "white"

            ax.text(j, i, fmt_string.format(value), ha="center", va="center", color=text_color, fontsize=8)

plt.tight_layout()
plt.show()

Figure 2: Representational Alignment Heatmaps for Macaque F and Macaque N. The plots display the raw RSA scores (Spearman’s rho-a) across all three visual regions (V1, V4, IT) and 21 normalized noise levels. Higher scores (lighter colors) indicate stronger alignment between the diffusion model and the biological brain.

In [4]:

import matplotlib.pyplot as plt
from matplotlib.ticker import FormatStrFormatter
import pandas as pd
from pathlib import Path

# Load data
bootstrap_csv = Path("results/rsa_bootstrap_ci_results.csv")
ceiling_csv = Path("results/monkey_rsa_comparison.csv")

boot_df = pd.read_csv(bootstrap_csv)
ceiling_df = pd.read_csv(ceiling_csv)

def get_noise_ceiling(ceiling_df, roi):
    subset = ceiling_df[ceiling_df["ROI"] == roi]
    if subset.empty:
        return None
    return subset["rsa_score"].mean()

ROI_COLORS = {"V1": "#4C72B0", "V4": "#DD8452", "IT": "#8172B2"}
ROI_ORDER = ["V1", "V4", "IT"]
monkeys = ["monkeyF", "monkeyN"]

fig, axes = plt.subplots(nrows=2, ncols=1, figsize=(9, 10))

for idx, monkey in enumerate(monkeys):
    ax = axes[idx]
    monkey_data = boot_df[boot_df["monkey"] == monkey]

    for roi in ROI_ORDER:
        roi_data = monkey_data[monkey_data["roi"] == roi].sort_values("noise_degree")
        if roi_data.empty:
            continue

        noise_levels = roi_data["noise_degree"].to_numpy(dtype=float)
        boot_means = roi_data["boot_mean"].to_numpy(dtype=float)
        ci_low = roi_data["ci_low"].to_numpy(dtype=float)
        ci_high = roi_data["ci_high"].to_numpy(dtype=float)

        ceiling = get_noise_ceiling(ceiling_df, roi)
        legend_label = f"{roi} (Ceiling: {ceiling:.2f})" if ceiling else f"{roi}"

        ax.plot(
            noise_levels, boot_means, marker="o", markersize=6, linewidth=2,
            color=ROI_COLORS.get(roi, "#333333"), label=legend_label
        )
        ax.fill_between(
            noise_levels, ci_low, ci_high, color=ROI_COLORS.get(roi, "#333333"),
            alpha=0.15, linewidth=0
        )

    subject_name = monkey.replace("monkey", "")
    ax.set_title(f"Macaque {subject_name}", fontsize=14, pad=10)
    
    if idx == 1:
        ax.set_xlabel("Normalized Noise Level", fontsize=14)
    ax.set_ylabel("RSA Score (Spearman rho-a)", fontsize=14)

    ax.spines["top"].set_visible(False)
    ax.spines["right"].set_visible(False)

    ax.xaxis.set_major_formatter(FormatStrFormatter("%.2f"))
    ax.tick_params(axis="both", labelsize=12)
    ax.grid(linestyle="--", alpha=0.7)

    ci_level = int(monkey_data["ci"].iloc[0]) if "ci" in monkey_data.columns else 95
    left_handles, left_labels = ax.get_legend_handles_labels()
    legend = ax.legend(
        left_handles, left_labels, loc="upper right", fontsize=11,
        title=f"Shaded bands denote {ci_level}% Bootstrap CI"
    )
    legend.get_title().set_fontsize(11)

plt.tight_layout()
plt.show()

Figure 3: Representational Alignment Curves for Macaque F and Macaque N. The plots display the mean RSA scores (Spearman’s rho-a) across different noise level with 95% subsampling bootstrap confidence intervals. The Macaque-to-Macaque alignment for each region is indicated in the legend.

In [5]:

semantic_csv = Path("results/semantic_ordering.csv")
semantic_df = pd.read_csv(semantic_csv)

text = " ".join(semantic_df['clean_category'].astype(str))
wordcloud = WordCloud(width=1000, height=500, background_color='white', max_words=200, random_state=42).generate(text)

fig, ax = plt.subplots(figsize=(10, 5))
ax.imshow(wordcloud, interpolation='bilinear')
ax.axis('off')

plt.tight_layout(pad=0)
plt.show()

Figure 4: Word cloud displaying the categories of images from the THINGS dataset.