TWIMLcon: AI Platforms 2019
2019-10-01..2019-10-02
San Francisco, United States
[TWIMLcon: AI Platforms 2019, TWIMLcon: AI Platforms 2021]
| title | author | length |
|---|---|---|
| Opening Address | Sam Charrington | 06m |
Narrative traces ML’s shift from AlexNet 2012 to production: TensorFlow rises then PyTorch gains ground, cloud infra lowers barriers, and enterprises stop with… | ||
| Overcoming the Barriers to Deep Learning in Production | Andrew Ng, Sam Charrington | 33m |
Overview of concrete tactics and failures: commercial engagements build vision-guided smart agricultural machinery that increases yield without new land; early… | ||
| Encoding Company Culture in Applied AI Systems | Deepak Agarwal, Sam Charrington | 32m |
Deepak Agarwal, vice president of artificial intelligence at LinkedIn, will join Sam Charrington on the TWIMLcon stage for a live podcast interview to explore… | ||
| ML Platform Journey | Jessica Egoyibo Mong | 27m |
SurveyMonkey describes moving from ad hoc notebooks to a standardized ML platform that serves 100 million predictions/day, cuts model deployment time from… | ||
| Model as a Service for Real-Time Decisioning | Sumit Daryani, John Swift | 25m |
Capital One builds a Model-as-a-Service stack that lets data scientists deploy models without engineering handoffs, using Seldon Core on Kubernetes to serve… | ||
| X1 Recommendations Platform: Started from the bottom.. | Leemay Nassery | 19m |
Comcast rebuilds X1 recommendations by replacing brittle HDFS + daily copies and Lambda handlers with a fleet of Flink jobs consuming Kafka/Kinesis and… | ||
| Model Versioning: Why, When, and How | Manasi Vartak | 24m |
Argues that production ML needs true model versioning and describes ModelDB and Verta workflows that track the model artifact plus code, data, config and… | ||
| Exploring Tradeoffs in Experiment Management as part of AI Platforms | Scott Clark | 25m |
Framework for deciding what to build versus buy for Experiment Management, applied to deployments across firms totaling $500B market cap and trading clients… | ||
| Ethical Human-in-the-Loop ML Pipelines | Robert Munro, Radha Basu | 32m |
Argues that ethical human-in-the-loop ML is practical and profitable by giving annotators fair pay, regular payment, transparent ownership, and by building… | ||
| Machine Learning in Production: What to do after you’ve trained the models? | Ameen Kazerouni | 29m |
Search is treated as a production case study: millions of unique queries times millions of customers times thousands of attributes makes realtime scoring… | ||
| Operationalizing Machine Learning | Pallav Agrawal | 30m |
Levi's goes from outsourced experiments to over a dozen models deployed and ROI positive in about two years, replacing a third‑party recommender that refreshed… | ||
| Why and How to Build Explainability into your ML Workflow | Krishna Gade | 24m |
Companies report nearly half lack trust in AI and regulators invoke Article 22 (GDPR), so explainability becomes a compliance and risk tool. Practical methods… | ||
| Chaos and Pain in Machine Learning, and the ‘DevOps for ML Manifesto' | Nick Ball | 24m |
Argues that ad hoc deployment causes “chaos and pain,” citing the oft‑quoted 80% of AI projects fail and that models can start spewing garbage in seconds… | ||
| Removing Unfair Bias in the Machine Learning Pipeline | Trisha Mahoney | 23m |
Demonstrates a concrete pipeline: uses AI Fairness 360 (open source Python) with over 77 fairness metrics and 10 mitigation algorithms to detect and reduce… | ||
| Semblance: Feature Generation In Real-Time and Batch Without Time-Travel | Tianshi Zhu | 20m |
Semblance implements an FRP-style API with events and features so feature values derive only from past events, eliminating no time travel mistakes; a single… | ||
| Culture & Organization for Effective ML at Scale | Eric Colson, Pardis Noorzad, Jennifer Prendki, Maribel Lopez | 27m |
Companies fail when ML is treated as a one-off instead of a running product: models need lifecycle management, monitoring and regular retraining rather than a… | ||
| Scaling ML in the Traditional Enterprise | Amr Awadallah, Pallav Agrawal, Jürgen Weichenberger, Joshua Bloom | 34m |
Traditional enterprises face a talent gap, operator-versus-innovator deadlock and real security risks from ad hoc data copies, so teams move to staging… | ||
| Team Teardown: Organizing and Supporting ML | Atul Kale, Peng Ye, Evan Cox, Sam Charrington | 32m |
Airbnb runs an internal ML platform called Bighead (launched October prior year) that teams use for notebooking, feature engineering, model exploration and… | ||
| Productionizing Machine Learning Models at Scale with Kubernetes | Luigi Patruno | 28m |
Builds a Kubernetes-based pipeline at 2U that runs roughly a thousand training and inference processes per day and spikes to tens of thousands in a week, using… | ||
| Operationalizing ML at Scale | Hussein Mehanna | 34m |
What should we takeaway from how web giants and autonomous vehicles are redefining scale and impact for ML platforms? Hussein Mehanna, Head of AI/ML at Cruise… | ||
| MagLev: A Production-grade AI Platform Running on GPU-enabled Kubernetes Clusters | Clement Farabet | 28m |
MagLev is NVIDIA’s production platform that ingests multiple terabytes per hour and manages petabyte-scale datasets to run training, replay and simulation at… | ||
| Use-Case Driven ML Platforms | Franziska Bell | 31m |
A platform team of 100 builds higher-level ML services on top of Michelangelo and reports concrete wins: push-button forecasting integrated into the BI… | ||
| Foundations for ML at Scale: Where to Start in Your Organization | Pete Skomoroch | 24m |
Start projects that directly move business metrics and prioritize by projected ROI, treating ~10–30% of work as venture bets for high-risk innovation. Invest… | ||
| Embeddings | Abhishek Tayal | 28m |
Twitter builds an embeddings pipeline that produces regularly retrained user and content vectors, stores them in the in‑house Strato feature store, and serves… | ||
| Why is Production ML so Hard? | Justin Norman | 21m |
Claims that production ML fails because the model is only the small part of a larger system, and organizations trip over three concrete issues: organizational… | ||
| Increasing the Impact of AI Through Better Software | Seth Weidman | 22m |
PyTorch explains how researchers move from prototype to production by compiling models to an intermediate representation and running them in high‑performance… | ||
| Closing Address | Sam Charrington | 13m |
An experiment runs cloud transcription services on podcast audio and finds cloud transcription defaults are useless: services need tuned modes, fail on… | ||
| Team Teardown: Evolving an ML Platform | Jing Huang, Shubhi Jain, Jolene Zhou, Sam Charrington | 32m |
Centralized ML team of about 7–8 engineers and 5 data scientists builds a serving platform, model management and a forthcoming feature store; onboarding moves… | ||
| Operationalizing Responsible AI | Rachel Thomas, Guillaume Saint-Jacques, Parinaz Sobhani, Khari Johnson | 31m |
Teams describe concrete tooling and outcomes: use TensorFlow Privacy, differential privacy and federated learning for data protection, and interpretability… | ||
| nothing matches | ||