Project overview
Arm Cascade Lab is a reproducible Cloud AI experiment for an Arm64 CPU inference service. It asks a concrete operational question: can a compact model handle obvious requests while a full model protects uncertain ones?
What it does
The tool trains two classifiers on a real handwritten-digit dataset: a low-cost probabilistic logistic classifier and a higher-cost RBF support-vector classifier. A configurable confidence gate accepts only high-confidence cheap predictions and escalates the remaining requests to the full model. The CLI writes a JSON artifact containing accuracy, escalation rate, and prediction-only latency.
Arm optimization evidence
On the local Arm64 machine, the 0.93 confidence gate achieved 99.21% cascade accuracy on 629 held-out samples, compared with 99.05% for the full model alone. It sent 44.67% of samples to the full model and measured 29.42 microseconds per cascade prediction versus 69.11 microseconds for the full model: a 2.35x prediction-latency improvement.
How it was built
The project uses Python, NumPy, and scikit-learn with a deterministic train/test split and fixed model seed. It includes tests, linting, an Apache-2.0 license, and a sample machine-readable report.
Limitations
Timing excludes request transport, serialization, queues, concurrency, p95/p99 latency, and model training. Developers should rerun it on their target Arm cloud instance before deployment.
Log in or sign up for Devpost to join the conversation.