Hyperlocal Drift Detection with Goko
Demonstrates Goko, an algorithm and tool for detecting fine-grained, localized concept drift in production ML models.
Summary
A machine learning researcher presents Goko in the AI Village, an open-source algorithmic tool designed for hyperlocal real-time drift detection across streaming data in production machine learning systems [00:04, 07:34]. The talk addresses the challenge of concept and covariate drift, where shifting input distributions quietly degrade model accuracy over time without triggering traditional error flags.
The speaker introduces Goko's core architecture, which leverages cover tree data structures and k-nearest neighbors to construct computationally efficient topological representations of reference data distributions [12:38, 12:45]. Benchmarking against the EMBER open-source malware dataset, the presentation demonstrates how Goko conducts rapid, streaming inference to detect adversarial perturbations and distribution anomalies in real time with minimal computational overhead [17:40, 23:01].
(AI summary — Gemini Flash)
Related talks
AI Village·24 min
Nahid Farhady
Introduces the Spectrum framework for deploying and operationalizing machine learning models in threat monitoring pipelines.
AI Village·21 min
Nahid Farhady
Examines privacy-preserving machine learning techniques, homomorphic encryption, and differential privacy safeguards.
AI Village·22 min
Hyrum Anderson · Zoltan Balaz
Reviews results and top evasion strategies from the AI Village machine learning security and evasion competition.