create_example_summary_table.py

"""
Create Summary Table of Example Pairs (No Images Needed)

This creates a simple table/figure showing the best example pairs
without needing the actual images.
"""

import pandas as pd
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches

print("="*80)
print("CREATING EXAMPLE SUMMARY TABLE")
print("="*80)

# Load exemplar pairs
exemplar_file = 'hierarchy_analysis/exemplar_image_pairs.csv'
print(f"\nLoading: {exemplar_file}")
exemplars = pd.read_csv(exemplar_file, index_col=0)

print(f"Loaded {len(exemplars)} exemplar pairs")

# Create summary figure
fig = plt.figure(figsize=(16, 12))

# Title
fig.suptitle('Exemplar Image Pairs Demonstrating Key Findings\n' +
             'Use these pairs to show concrete examples to your professor',
             fontsize=16, fontweight='bold', y=0.98)

# SECTION 1: Dissociation examples
dissociation = exemplars[exemplars['example_type'] == 'dissociation'].head(5)

ax1 = plt.subplot(3, 1, 1)
ax1.axis('off')

y_pos = 0.95
ax1.text(0.5, y_pos, 'FINDING 1: Semantic Dissociation', ha='center',
         fontsize=14, fontweight='bold', transform=ax1.transAxes)
y_pos -= 0.08
ax1.text(0.5, y_pos, 'Language similarity HIGH, Vision similarity LOW → Language captures meaning, not just pixels',
         ha='center', fontsize=11, style='italic', transform=ax1.transAxes)

y_pos -= 0.12
for idx, (pair_idx, row) in enumerate(dissociation.iterrows()):
    # Header
    ax1.text(0.02, y_pos, f"Example {idx+1}:", fontweight='bold',
            fontsize=10, transform=ax1.transAxes)

    # Image names
    y_pos -= 0.05
    ax1.text(0.02, y_pos, f"  Images: {row['image_1'][:30]}... vs {row['image_2'][:30]}...",
            fontsize=9, family='monospace', transform=ax1.transAxes)

    # Scores
    y_pos -= 0.04
    lang_val = row['language_avg']
    vis_val = row['vision_avg']
    diff = lang_val - vis_val
    human = row['human_judgement']

    score_text = (f"  Language: {lang_val:+.2f}  |  Vision: {vis_val:+.2f}  |  "
                 f"Diff: {diff:+.2f}  |  Human: {human:.2f}")

    # Color code based on values
    if lang_val > 0.5:
        lang_color = '#27ae60'  # Green for high
    else:
        lang_color = 'black'

    if vis_val < -0.5:
        vis_color = '#e74c3c'  # Red for low
    else:
        vis_color = 'black'

    ax1.text(0.02, y_pos, score_text, fontsize=9, transform=ax1.transAxes)

    y_pos -= 0.07
    if y_pos < 0.05:
        break

# SECTION 2: Language-Human Alignment
alignment = exemplars[exemplars['example_type'] == 'language_human_alignment'].head(5)

ax2 = plt.subplot(3, 1, 2)
ax2.axis('off')

y_pos = 0.95
ax2.text(0.5, y_pos, 'FINDING 2: Language-Human Alignment', ha='center',
         fontsize=14, fontweight='bold', transform=ax2.transAxes)
y_pos -= 0.08
ax2.text(0.5, y_pos, 'Language models predict human judgments better than vision models',
         ha='center', fontsize=11, style='italic', transform=ax2.transAxes)

y_pos -= 0.12
for idx, (pair_idx, row) in enumerate(alignment.iterrows()):
    # Header
    ax2.text(0.02, y_pos, f"Example {idx+1}:", fontweight='bold',
            fontsize=10, transform=ax2.transAxes)

    # Image names
    y_pos -= 0.05
    ax2.text(0.02, y_pos, f"  Images: {row['image_1'][:30]}... vs {row['image_2'][:30]}...",
            fontsize=9, family='monospace', transform=ax2.transAxes)

    # Alignment scores
    y_pos -= 0.04
    human = row['human_judgement']
    lang = row['language_avg']
    vis = row['vision_avg']
    lang_dist = abs(human - lang)
    vis_dist = abs(human - vis)
    closer = vis_dist - lang_dist

    score_text = (f"  Human: {human:.2f}  |  Language: {lang:.2f} (dist: {lang_dist:.2f})  |  "
                 f"Vision: {vis:.2f} (dist: {vis_dist:.2f})  →  Language {closer:.2f} closer")

    ax2.text(0.02, y_pos, score_text, fontsize=9, transform=ax2.transAxes)

    y_pos -= 0.07
    if y_pos < 0.05:
        break

# SECTION 3: Brain Semantic Preference
brain_pref = exemplars[exemplars['example_type'] == 'brain_semantic_preference'].head(5)

ax3 = plt.subplot(3, 1, 3)
ax3.axis('off')

y_pos = 0.95
ax3.text(0.5, y_pos, 'FINDING 3: Brain Prioritizes Semantics', ha='center',
         fontsize=14, fontweight='bold', transform=ax3.transAxes)
y_pos -= 0.08
ax3.text(0.5, y_pos, 'Late semantic brain regions respond more than early visual regions',
         ha='center', fontsize=11, style='italic', transform=ax3.transAxes)

y_pos -= 0.12
for idx, (pair_idx, row) in enumerate(brain_pref.iterrows()):
    # Header
    ax3.text(0.02, y_pos, f"Example {idx+1}:", fontweight='bold',
            fontsize=10, transform=ax3.transAxes)

    # Image names
    y_pos -= 0.05
    ax3.text(0.02, y_pos, f"  Images: {row['image_1'][:30]}... vs {row['image_2'][:30]}...",
            fontsize=9, family='monospace', transform=ax3.transAxes)

    # Brain responses
    y_pos -= 0.04
    early = row['hierarchy_early_visual_avg']
    late = row['hierarchy_late_semantic_avg']
    diff = late - early

    score_text = (f"  Early Visual: {early:+.3f}  |  Late Semantic: {late:+.3f}  |  "
                 f"Diff: {diff:+.3f}  |  Human: {row['human_judgement']:.2f}")

    ax3.text(0.02, y_pos, score_text, fontsize=9, transform=ax3.transAxes)

    y_pos -= 0.07
    if y_pos < 0.05:
        break

plt.tight_layout(rect=[0, 0, 1, 0.96])
output_file = 'hierarchy_analysis/exemplar_pairs_summary.png'
plt.savefig(output_file, dpi=300, bbox_inches='tight', facecolor='white')
print(f"\nSaved: {output_file}")
plt.close()

# Also save as a text file for easy reference
print("\nCreating text summary...")
with open('hierarchy_analysis/exemplar_pairs_summary.txt', 'w') as f:
    f.write("="*80 + "\n")
    f.write("EXEMPLAR IMAGE PAIRS - SUMMARY FOR PROFESSOR\n")
    f.write("="*80 + "\n\n")

    f.write("FINDING 1: SEMANTIC DISSOCIATION\n")
    f.write("-" * 80 + "\n")
    f.write("Language models show high similarity, vision models show low similarity\n")
    f.write("Proves: Language models capture semantic meaning, not just visual features\n\n")

    for idx, (pair_idx, row) in enumerate(dissociation.iterrows()):
        f.write(f"Example {idx+1}:\n")
        f.write(f"  Pair Index: {pair_idx}\n")
        f.write(f"  Image 1: {row['image_1']}\n")
        f.write(f"  Image 2: {row['image_2']}\n")
        f.write(f"  Language Similarity: {row['language_avg']:+.3f} (HIGH)\n")
        f.write(f"  Vision Similarity:   {row['vision_avg']:+.3f} (LOW)\n")
        f.write(f"  Difference: {row['language_avg'] - row['vision_avg']:+.3f}\n")
        f.write(f"  Human Judgment: {row['human_judgement']:.3f}\n")
        f.write(f"  Early Brain: {row['hierarchy_early_visual_avg']:+.3f}\n")
        f.write(f"  Late Brain:  {row['hierarchy_late_semantic_avg']:+.3f}\n")
        f.write("\n")

    f.write("\n" + "="*80 + "\n")
    f.write("FINDING 2: LANGUAGE-HUMAN ALIGNMENT\n")
    f.write("-" * 80 + "\n")
    f.write("Language models align with human judgments better than vision models\n")
    f.write("Proves: Humans judge similarity based on semantics, not just visual features\n\n")

    for idx, (pair_idx, row) in enumerate(alignment.iterrows()):
        f.write(f"Example {idx+1}:\n")
        f.write(f"  Pair Index: {pair_idx}\n")
        f.write(f"  Image 1: {row['image_1']}\n")
        f.write(f"  Image 2: {row['image_2']}\n")
        f.write(f"  Human Judgment: {row['human_judgement']:.3f}\n")
        f.write(f"  Language Model: {row['language_avg']:.3f} (distance: {abs(row['human_judgement'] - row['language_avg']):.3f})\n")
        f.write(f"  Vision Model:   {row['vision_avg']:.3f} (distance: {abs(row['human_judgement'] - row['vision_avg']):.3f})\n")
        f.write(f"  Language is {abs(row['human_judgement'] - row['vision_avg']) - abs(row['human_judgement'] - row['language_avg']):.3f} closer to human\n")
        f.write("\n")

    f.write("\n" + "="*80 + "\n")
    f.write("FINDING 3: BRAIN PRIORITIZES SEMANTICS\n")
    f.write("-" * 80 + "\n")
    f.write("Late semantic brain regions respond more than early visual regions\n")
    f.write("Proves: Brain uses semantic content for similarity judgments\n\n")

    for idx, (pair_idx, row) in enumerate(brain_pref.iterrows()):
        f.write(f"Example {idx+1}:\n")
        f.write(f"  Pair Index: {pair_idx}\n")
        f.write(f"  Image 1: {row['image_1']}\n")
        f.write(f"  Image 2: {row['image_2']}\n")
        f.write(f"  Early Visual (V1-hV4): {row['hierarchy_early_visual_avg']:+.3f} (LOW)\n")
        f.write(f"  Late Semantic (FFA, VWFA, etc): {row['hierarchy_late_semantic_avg']:+.3f} (HIGH)\n")
        f.write(f"  Difference: {row['hierarchy_late_semantic_avg'] - row['hierarchy_early_visual_avg']:+.3f}\n")
        f.write(f"  Human Judgment: {row['human_judgement']:.3f}\n")
        f.write("\n")

print("Saved: hierarchy_analysis/exemplar_pairs_summary.txt")

print("\n" + "="*80)
print("DONE!")
print("="*80)

print("""
Created:
  1. exemplar_pairs_summary.png - Visual summary (no images needed)
  2. exemplar_pairs_summary.txt - Text file with all details

These show the BEST example pairs for each finding.

For your professor meeting:
  1. Show the bar charts (main statistical findings)
  2. Show exemplar_pairs_summary.png (specific examples)
  3. Walk through 1-2 examples in detail:
     - "Look at this pair: [image names]"
     - "Language says similar (0.86), vision says different (-1.34)"
     - "Humans agree with language (2.0)"
     - "Brain semantic regions respond (0.22), visual regions don't (-0.22)"
     - "This proves our point!"

If you have the actual images, run:
  python create_image_examples_figure.py
to create figures WITH the images.

The pair indices are saved so you can look them up in your GUI!
""")