CIS 731 | Project : Week 1: Re-Onboarding & Semester Planning

March 26, 2026

What I worked on

Reestablished the core research question: how many IMU sensors are necessary for reliable gait-based identity verification? The underlying hypothesis is that a minimal sensor subset exists, are the ones that preserves authentication performance while meaningfully reducing hardware burden. Answering this has direct implications for real-world deployment where a full 16-sensor body area network is impractical.
Reviewed last semester’s documented results and logs to understand what worked and what didn’t before writing any new code:
1. Gen 0 (6 independent CNNs, TripletLoss) -> 6.66% EER; best result, achieved with a simpler architecture.
2. Gen 1 (shared-weight Siamese CNN, BatchHard triplet loss) -> 14.88% EER; architectural complexity hurt.
Finalised the 3phase implementation plan as submitted in the project proposal.

Increased architectural complexity does not guarantee better performance:
- Gen 1’s shared-weight CNN performed significantly worse than Gen 0’s independent CNNs
- because it lost sensor-specific information by treating all sensors identically.
This failure directly shaped the Gen 3 design decision, where per-sensor representations must be fully extracted by the CNN before the Transformer performs any cross-sensor fusion.
Last semester established that gait authentication works; this semester’s real contribution is answering which sensors are actually driving it and refinging the algorithm.

Re-reading last semester’s codebase after a break made it clear that the existing Gen 3 architecture has known unresolved bugs that need to be addressed before any new work can begin.

Full codebase audit, identify and resolve all known pipeline bugs, and confirm Gen 0 and Gen 1 baselines reproduce correctly before touching any new model code.