Research with
clear boundaries.
ATLAS studies the relationship between model capability and computational cost. Each direction distinguishes objective, proposal, current state, and evidence.
Model Quantization
Reducing numerical precision while preserving useful signal.
Study how lower-precision representations can reduce memory and compute requirements without treating quality loss as an acceptable default.
A single precision choice can hide layer-level sensitivity. Uniform compression may save memory while creating avoidable degradation in the most informative parts of a model.
Build a measurement-first quantization layer that separates representation choice, calibration, and verification. The system is designed to make every trade-off observable rather than assume a universal bit-width.
x → Q(x; s, z), with scale s and optional zero-point z. The error term ε = x − Q(x) remains an explicit object for downstream analysis.
Calibration → sensitivity map → candidate representation → reconstruction check
Research direction defined; validated benchmark data not yet published.
No verified public result is attached to this project yet.
Compare layer-wise, channel-wise, and activation-aware policies under a shared evaluation protocol.
Measure the gap
before closing it.
Separate what is implemented from what is proposed.
Keep quality, memory, and latency as distinct signals.
Publish only what another researcher can inspect.