OpML '20
as the organisers put it
The 2020 USENIX Conference on Operational Machine Learning (OpML '20) will take place as a series of virtual events, free of charge and open to everyone, during the weeks of July 27 and August 3.
who it's for
mlops engineers, machine learning engineers, data engineers, platform engineers, site reliability engineers, research engineers
topics
model lifecycle management; model registries and deployment; distributed model inference; GPU data science stack; feature provenance and time travel; online experimentation and tuning; explainable model techniques; autoML for time series; edge and on-device inference; privacy preserving techniques; ML outage root causes; kubernetes ML workload management
| title | author | length |
|---|---|---|
| DLSpec: A Deep Learning Task Exchange Specification | Abdul Dakkak, Cheng Li, Jinjun Xiong, Wen-mei Hwu | 15m |
Claims a reproducible DL task format that is model-, dataset-, software-, and hardware-agnostic, defines tasks as four manifest files (hardware, software… | ||
| End-to-End Data Science on GPUs | John Zedlewski | 20m |
RAPIDS implements an end-to-end GPU data science stack with cuDF for GPU DataFrame ETL, DLPack/CUDA array interface for zero-copy handoff to… | ||
| Operational Challenges in Machine Learning Model Life Cycle | Chandra Mohan Meena, Sarwesh Suman, Vijay Agneeswaran | 19m |
Walmart Labs builds an Analyzer that instruments user events, stores them in an event repository, tags signals with a signal generator, and models behavior… | ||
| Runway - Model Lifecycle Management | Eugen Cepoi, Liping Peng | 19m |
Runway is Netflix’s model registry and deployment gate that manages thousands of personalization models, parsing model files, tracking lineage, and enforcing… | ||
| Detecting Feature Eligibility Illusions in Enterprise AI Autopilots | Fabio Casati, Veeru Metha, Gopal Sarda, Sagar Davasam, Kannan Govindarajan | 24m |
Describes a concrete method to detect three failure modes—inverse causality, observability illusion, and end-state bias—that make ML models look accurate on… | ||
| Time Travel and Provenance for Machine Learning Pipelines | Alexandru A. Ormenisan, Moritz Meister, Fabio Buso, Robin Andersson, Seif Haridi, Jim Dowling | 20m |
Claims that intrusive code hooks are replaced by standardized change‑data‑capture from the distributed file system, feature store, resource manager and apps… | ||
| An Experimentation and Analytics Framework for Large-Scale AI Operations Platforms | Thomas Rausch, Waldemar Hummer, Vinod Muthusamy | 19m |
A trace-driven experimentation and analytics framework models IBM production AI ops traces, builds a SimPy-based discrete-event simulator and generates… | ||
| Challenges Towards Production-Ready Explainable Machine Learning | Lisa Veiber, Kevin Allix, Yusuf Arslan, Tegawendé F. Bissyandé, Jacques Klein | 17m |
Fintech deployment uses a preexisting random forest and needs both global reports for regulators and local justifications for clients and analysts, so the… | ||
| RIANN: Real-time Incremental Learning with Approximate Nearest Neighbor on Mobile Devices | Jiawen Liu, Zhen Xie, Dimitrios Nikolopoulos, Dong Li | 16m |
RIANN is an on-device system that builds graph-based ANN indexes with per-point, dynamic construction parameters and uses a trained statistical prediction… | ||
| Rise of the Machines: Removing the Human-in-the-Loop | Viral Gupta, Yunbo Ouyang | 21m |
System automatically tunes linear-combination ranking hyperparameters (example: in-app threshold x_th combining p_click and p_visit) as a constrained black-box… | ||
| Cluster Serving: Distributed Model Inference using Big Data Streaming | Jiaming Song, Dongjie Shi, Qiyuan Gong, Lei Xia, Jinquan Dai | 18m |
Cluster Serving is an automated, distributed inference system built on Analytics Zoo that wires models into Apache Spark Streaming and Apache Flink pipelines… | ||
| Scalable AutoML for Time Series Forecasting | Shengsheng Huang, Jason Dai | 22m |
An AutoML pipeline for time series runs on Ray backend with a real-spark integration to launch distributed trials, use Ray Tune for hyperparameter search, and… | ||
| FlexServe: Deployment of PyTorch Models as Flexible REST Endpoints | Edward Verenich, Alvaro Velasquez, M. G. Sarwar Murshed, Faraz Hussain | 17m |
FlexServe is a lightweight Python library that deploys PyTorch models as REST endpoints, avoiding ONNX/TensorFlow conversion by running multi-model ensemble… | ||
| Managing ML Models @ Scale | Srivathsan Canchi, Tobias Wenzel | 26m |
Intuit’s ML Platform claims 200% increase in model publishing velocity by combining GitOps, Argo Workflows, Kubernetes and AWS SageMaker with a feature store… | ||
| Edge Inference on Unknown Models | Georgiana Copil, Iulian Radu, Akash Maharaj | 21m |
Adobe Target enables running customer-provided ML models on Java edge servers by converting models to ONNX runtime and keeping inference inside the existing… | ||
| More Data Science, Less Engineering | Savin Goyal | 27m |
Netflix uses an open source Metaflow framework to let data scientists write idiomatic Python DAGs with decorators that declare steps, automatic state… | ||
| Auto Content Moderation in C2C e-Commerce | Shunya Ueta, Suganprabu Nagaraja, Mizuki Sango | 20m |
Mercari, inc. deploys an ML pipeline for C2C item moderation since late 2017 that cuts operation cost and catches policy-violating listings rule-based checks… | ||
| Inside NVIDIA’s AI Infrastructure for Self-driving Cars | Clement Farabet, Nicolas Koumchatzky | 26m |
Project MagLev is NVIDIA’s end-to-end AV AI platform that ingests TBs/hour from fleet vehicles, stores multi-PB datasets, and orchestrates training across… | ||
| Challenges and Experiences with MLOps for Performance Diagnostics in Hybrid-Cloud | Amitabha Banerjee, Chien-Chia Chen, Chien-Chun Hung, Xiaobo Huang, Yifang Wang, Razvan Chevesaran | 25m |
ML-based performance diagnostics for VMware hybrid-cloud builds a cloud-hosted pipeline that ingests anonymized telemetry via Apache Spark, trains unsupervised… | ||
| Automating Operations with ML | Todd Underwood, Steven Ross | 25m |
Survey finds most ML-for-ops ideas fail in practice because labeled data is scarce, systems are dynamic, and engineering costs outweigh gains; anomaly… | ||
| How ML Breaks: A Decade of Outages for One Large ML Pipeline | Daniel Papasian, Todd Underwood | 19m |
Google analyzes 96 post-mortems from a 15-year continuous pipeline running thousands of concurrent models with about 100 billion parameters, finds failures… | ||
| SPOK - Managing ML/Big Data Spark Workloads at scale on Kubernetes | Nagaraj Janardhana, Mikhail Arov | 22m |
Built an elastic Spark-on-Kubernetes platform that moves workloads from EMR/YARN to Kubernetes using the Spark Operator, Argo CD/workflows, Jupyter on K8s and… | ||
| Addressing Some of Challenges When Optimizing Long-to-Train-Models | Tobias Andreasen | 26m |
SigOpt provides an API that runs Bayesian optimization with acquisition functions, encodes previous runs as prior probability density functions, and supports… | ||
| "SECRETS ARE LIES, SHARING IS CARING, PRIVACY IS THEFT." - A Dive into Privacy Preserving | Nahid Farhady Ghalaty, Vincent Pham | 21m |
Describes concrete attacks and defenses rather than slogans: shows linkage, reconstruction, model inversion, membership inference attacks including a demo that… | ||
| ML Artifacts Ownership Enforcement | Yan Yan | 19m |
Centralized inventory enforces ownership for every ML artifact by registering models, features, pipelines and code in an entity framework backed by a graph… | ||
| nothing matches | ||