Results & Data Analysis
Simulated data reproducing the paper's main findings with interactive visualizations
These visualizations use computationally generated data that reproduce the qualitative patterns reported in the paper. The original experimental data (neural recordings from 1,081 neurons across 5 brain regions) is available at FigShare and analysis code at Zenodo.
Behavioural Performance (Fig. 1e,f)
Both monkeys performed all three tasks well. Performance improved over the first ~75 trials of S1 and C1 blocks as monkeys discovered which task was in effect. C2 performance was high immediately after switch.
Key Behavioural Findings
| Task | Monkey Si | Monkey Ch | Combined | Significance |
|---|---|---|---|---|
| S1 | 81% | 77% | 79% | P < 0.001 |
| C1 | 83% | 78% | 80.5% | P < 0.001 |
| C2 | 92% | 92% | 92% | P < 0.001 |
Cross-Task Classifier Generalization (Fig. 2)
Classifiers trained on one task can decode the same variable in another task, demonstrating shared representational subspaces.
Cross-task decoding accuracy was significantly above chance for both colour (between C1 and C2) and motor response (between S1 and C1), confirming shared representational subspaces in LPFC. Motor responses generalized across all regions, while colour generalization was strongest in LPFC.
Sensory-Motor Transformation (Fig. 3)
Information flows sequentially from shared sensory subspaces to shared motor subspaces. Colour encoding precedes motor response encoding by ~63ms. Trial-by-trial correlation confirms the transformation.
Task Discovery Dynamics (Fig. 4)
Monkeys discover which task is in effect through feedback. Task belief, colour subspace engagement, and shape subspace engagement all evolve systematically during the block.
Task belief increases as monkeys discover the C1 task. This belief predicts the engagement of shared colour subspaces (positive correlation) and inversely predicts shape subspace engagement (negative correlation). Motor response decoding remains stable throughout, consistent with both S1 and C1 using axis 1.
Compression & Gain Modulation (Fig. 5)
Task-relevant dimensions are amplified while irrelevant dimensions are compressed. The Compression Index (CPI) tracks the relative strength of colour vs. shape representations.
Regional Differences
Neural recordings from five brain regions revealed different patterns of shared vs. task-specific representations.
Recording Summary
| Region | Neurons | Colour Shared? | Motor Shared? | Key Role |
|---|---|---|---|---|
| LPFC | 480 | Strong | Strong | Task control, shared subspaces |
| FEF | 149 | Weak/Delayed | Strong | Motor preparation |
| PAR | 64 | Weak/Delayed | Strong | Spatial attention |
| aIT | 239 | Weak/Delayed | Strong | Veridical stimulus |
| STR | 149 | Not significant | Strong | Reward/learning |