← Research Lab / Systems
FIG. AAA2 // FANTASMA RESEARCH NOTE

Detecting Concept Drift in Production ML Without Ground Truth Labels

Oleh Dr. Selin Karaçay
9 min baca
2026-05-11

Detecting Concept Drift Without Ground Truth

Every ML team learns the same hard lesson eventually: models degrade in production. The world changes. Distributions shift. And by the time you have labelled data to prove it, the damage is done.

The standard drift detection playbook — monitor prediction distributions, wait for label lag to catch up, retrigger training — adds 12–48 hours of latency to the detection loop. For high-stakes production systems, this is too slow.

The Core Insight

If we cannot wait for labels, we must detect drift from inputs alone. The challenge: distinguishing covariate shift (the input distribution changed) from natural variation (inputs vary but the decision boundary hasn't moved).

We frame this as a two-sample test problem. At training time, we learn a low-dimensional representation of the input manifold using a contrastive encoder. At inference time, we run a streaming Maximum Mean Discrepancy (MMD) test between a sliding window of live inputs and the training distribution embedding.

Detection Performance

On our internal sensor fusion benchmark (injected drift at known timestamps):

Median time-to-detection: 2.8 hours
False positive rate: < 0.3%/day
Compute overhead: < 0.1% of inference cost

For reference, waiting for label lag in this system averages 19 hours.

Publikasi Riset Terkait

Architecture 11 min

Attention Is Not Enough: Toward Structured World Models at the Edge

Transformer attention scales quadratically with sequence length. For always-on edge inference, this is a thermodynamic impossibility. We propose a hybrid SSM-attention architecture that achieves 94% of full-attention quality at 8% of the compute.

HCI 8 min

Calibration-Free Gaze Tracking at 60fps on Commodity Hardware

Classical gaze tracking requires a 90-second calibration ritual that most users abandon. We trained a universal gaze model on 4.2M synthetic eye-renders that generalizes to unseen users with 0.4° accuracy — no calibration, no personal data.