## Line Charts: Brain Alignment Across Model Sizes ### Overview Three line charts compare brain alignment (Pearson's r) between "Language Network" (green circles) and "V1" (purple crosses) across three model sizes: 14M, 70M, and 160M. Each chart tracks alignment as tokens increase from 0 to 2868M, with a vertical reference line at 512M tokens. --- ### Components/Axes - **X-axis**: "Number of Tokens" (0, 2M, 4M, ..., 2868M) - **Y-axis**: "Brain Alignment (Pearson's r)" (-0.025 to 0.150) - **Legend**: Located at bottom center, with: - Green circles: Language Network - Purple crosses: V1 - **Vertical Line**: At 512M tokens in all charts --- ### Detailed Analysis #### 14M Model - **Language Network**: - Starts at ~0.055 (0 tokens) - Peaks at ~0.125 (2868M tokens) - Steady upward trend with minor fluctuations - **V1**: - Starts at ~0.01 (0 tokens) - Peaks at ~0.03 (2868M tokens) - Fluctuates between 0.01 and 0.03 #### 70M Model - **Language Network**: - Starts at ~0.05 (0 tokens) - Peaks at ~0.12 (2868M tokens) - Consistent upward slope with slight dips - **V1**: - Starts at ~0.015 (0 tokens) - Peaks at ~0.035 (2868M tokens) - More variability than 14M model #### 160M Model - **Language Network**: - Starts at ~0.05 (0 tokens) - Peaks at ~0.12 (2868M tokens) - Stable increase with minor noise - **V1**: - Starts at ~0.02 (0 tokens) - Peaks at ~0.04 (2868M tokens) - Smoother trend than smaller models --- ### Key Observations 1. **Model Size Correlation**: Larger models (160M > 70M > 14M) show consistently higher brain alignment for Language Network. 2. **Token Count Impact**: Alignment improves for both metrics as token count increases, with sharper gains after 512M tokens. 3. **V1 Variability**: V1 shows more fluctuation in smaller models (14M) but stabilizes in larger models (160M). 4. **Shaded Regions**: Confidence intervals widen with token count, indicating increased measurement uncertainty at higher token volumes. --- ### Interpretation - **Language Network Dominance**: The green line (Language Network) consistently outperforms V1 across all model sizes, suggesting it better captures brain-related patterns. - **Scaling Benefits**: Larger models (160M) achieve higher alignment with fewer tokens compared to smaller models, indicating improved efficiency. - **V1 as Baseline**: V1's lower alignment values and higher variability suggest it represents a less optimized or smaller-scale baseline. - **512M Threshold**: The vertical line at 512M tokens may mark a critical point where model performance stabilizes or diverges significantly. The data implies that model size and token processing capacity directly influence brain alignment, with larger models achieving stronger, more stable correlations.