SynchroStream-ML
DSS for MLOps: Neural Perturbation Score to prevent catastrophic forgetting
PythonPyTorchFastAPIReactNumPyPandas
Research-grade Decision Support System that analyzes gradient conflicts between incoming data streams and a model's existing knowledge. Computes the Neural Perturbation Score (NPS) — a novel metric quantifying data-to-model conflict — and recommends optimal ingestion strategies. Under review at Elsevier.
- › NPS-driven ingestion strategy selection (Linear/Interleaved/Parallel)
- › Dynamic buffer resizing based on NPS (22% forgetting reduction)
- › Layer-wise perturbation heatmaps for interpretability
- › FastAPI backend + React dashboard with live Nivo heatmaps