MLOps Productionalization
OVERVIEW
Machine Learning (ML) has empowered a broad range of industries
including life sciences, manufacturing, consumer goods, financial
services, and telecom. Despite its ubiquity, many enterprises face
myriad challenges and shortcomings in developing, deploying and
managing their Machine Learning applications and find it difficult to
shift from experimentation to production grade AI. The need to adopt
DevOps practices in ML to increase automation whilst addressing
business and regulatory requirements is becoming a foundational
need that’s converging into a set of common practices, tools and
governance functions under Machine Learning Operations or MLOps.
MLOps encompasses a combination of practices and processes that
aim to make seamless and efficient development, deployment, scaling
and maintenance of Machine Learning models. It emphasizes on
fostering common practices, automating repetitive tasks and
establishing governance layer to expedite implementation of ML
at an enterprise level.
Only 14.6% of firms report By 2024, 75% of organizations
that they have deployed AI will shift from piloting to
capabilities into widespread operationalizing AI- Gartner
production- Forbes
Enterprise challenges in scaling MLOps
Inadequate governance and maintenance of models lead to
incorrect predictions.
-
Lack of collaboration among siloed teams.
Suboptimal processes cause long lead time
to take ML models into production
without meeting availability and
$
reliability requirements. Skills shortage & high cost of continuity.
!
Long cycle times before proving valuable.
Introducing HCL’s MLOps Framework
A cloud native approach to productionizing AI experiments to enterprise scale
Built from the experience of dozens of Security
experiments and production deployments, whitelist
Manage ML libraries and
HCL’s MLops framework fuses the growing infrastructure frameworks for
capabilities of cloud hyperscale, specialized & environments use in model
AI tools and tribal knowledge gained within using development.
the enterprise across its AI maturity curve. Infra-as-code
with approved
libraires & tools.
The framework defines and orchestrates Model
the AI lifecycle across the dimensions of monitoring &
Track & version reporting.
infrastructure, model development,
production, monitoring and business control models
Create & manage and associated
feedback. Much like AI models, the continuous content for
framework is built to learn with each new integration and reference &
experiment, adding new knowledge through deployment reuse.
recipes, modules, metadata and metrics. (CI/CD)
pipelines.
Incrementally, the framework improves
Automated
standardization, reusability and efficiency model
of AI development across the enterprise. Profile models
for retraining to
The blend of knowledge and software Discover existing consumption enable faster
engineering practices enables the science enterprise vetted readiness. adoption and
of AI to mature while automation brings recipes, features, course
models for correction.
agility and reliability. reusability.
Framework Building Blocks
MLOps Framework – Key Features
Model Development
Catalog Management - A catalog that lists the models, including when they were
tested, trained, and deployed. Also includes validation datasets and accuracy
measurements of stored models.
1
ML Pipelines Orchestration - Develop multiple pipelines for preprocessing, feature
engineering, training, validation, and testing to update ML models automatically. 2
Feature Store – Create a store for curating features for machine learning pipelines. 3
Model Asset Versioning - Ability to roll out, roll back with flexibility by leveraging
continuous integration and continuous testing/deployment tools. 4
Data & Model Bias - Assess success of different modelling
approaches against goodness-of-fit criteria. 5
A/B Testing - Evaluating the results of different experiments.
6
Model Operation & Governance
1 ML Model CI/CD - Automating model management from deployment to production.
Operationalize ML Models at Scale - Enabling data scientists to quickly deploy ML models into
2 production by containerization.
Model Drift Measurement - Measure data/concept drifts and generate automated alerts
3 generated for business users.
4 Model Re-Training - Automated model training to generate accurate predictions.
5 Feedback from Business - Enable feedback loop from business stakeholders.
Model Health Monitoring & Alerting - Identify deviated predictions beyond the acceptable
6 threshold levels.
Reporting & Dashboarding - Publish exceedance report, weekly report on validation and model
7 inventory monthly report.
8 Ethical AI - Enable model interpretability and model fairness.
Menu of Offerings
MLOps Framework &
CoE Services Environments
Platform Setup, Architecture Design, Cloud Agnostic Framework
Training, Platform Support. (Azure, AWS, GCP), Integrated
Framework with Data Science tools
like Data Robot and Dataiku.
1
2
Functionalities
Deploy ML Models, CI/CD Automated
Pipelines, Model Serving
and Monitoring , Model Governance,
Model Explainability.
Model & Asset Library 4
Data Science Libraries, Notebooks,
Algorithms, Data Sets and Recipes.
Business Benefits
Maximize business Maintenance of Provide ethical,
impact from the right coding responsible, and
proactive model and evaluation explainable decision
health monitoring. standards. making support
and capability.
Highly adaptive Augmentation of Feature store helps
and responsive to ML/AI to mitigate to reduce time on
variability in the regulatory data collection and
incoming data. concerns of bias data preparation.
detection & testing.
Mitigate data Ensure quality Accelerate the
security and of the data and model validation
privacy. models. process.
BI-105244747974-EN00GL
