CIS 731 | Project : Week 5: Sensor Pruning Study Begins
April 19, 2026
What I worked on
- Continued hyperparameter tuning from W4 and reached a stable Gen 3 config:
- EER has improved from initial runs but remains above the Gen 0’s baseline.
- Treating this as the working baseline for the sensor pruning study going forward.
- Designed the sensor pruning study structure:
- Single-sensor removal, as it says removing 1 sensor at a time and measure the EER impact.
- Anatomical grouping: upper body only, lower body only, extremities only.
- Ran an initial single-sensor removal experiments:
- Some sensors show minimal EER impact when removed.
- While others cause a sharp EER increase, carry identity-relevant information.
What I learned
- Not all sensors contribute equally, early results suggest a small subset of sensors is responsible for most of the identity-relevant signal.
- EER degradation is not linear with sensor removal, removing certain sensors has disproportionately large impact compared to others.
Challenges
- Gen 3 EER is not yet at or below the Gen 0’s baseline, which makes interpreting pruning results harder, it is difficult to define “acceptable degradation” relative to a baseline that is itself not fully optimised.
- Running one experiment per sensor configuration is time-consuming, and results so far are incomplete.
Next week
- Complete single-sensor removal experiments across all sensors.
- Begin anatomical group experiments.
- Extract Transformer attention weights and compare against pruning results.