Join this virtual event and interactive discussion with industry experts about operationalizing machine learning models in the enterprise. Learn how HPE (with solutions from BlueData) is helping address the challenges of deploying and managing machine learning models at scale.
Q6: How is the ML model lifecycle different than DevOps? What aspects of DevOps can be applied to ML models? What aspects are unique to ML Ops? https://www.crowdchat.net/s/35vit
While DevOps is more about marrying the developers and IT operations folks, ML Ops builds on DevOps and incorporates some additional personas, such as data engineers and data scientists.
MLops is building on top of the DevOPS concepts. However, MLops focusses on the ML lifecycle which is a pipeline that relies on continuous feedback and improvement.
For ML Ops specifically, it's heavily centered on providing intelligence derived from data to a specific business objective, goal, or use case, and enabling developers to consume models and utilize the insights for rapid innovation.
Training a model and reiterating through the cycle until you achieve the acceptance criteria of the model. Devops doesn’t have this requirement or iterative step.
ML Lifecycle is very iterative with a heavy dependency on data. MLOps leverages DevOps processes, however the iterative nature of the ML Lifecycle and the dependency on data should be taken into account. Changes in underlying data will drive retraining decisions
Operationalization for the Machine Learning Lifecycle
Operationalization for the Machine Learning Lifecycle The HPE ML Ops solution supports every stage of machine learning lifecycle — data preparation, model build, model training, model deployment, collaboration, ...
The ML lifecycle is a very iterative and continuous process (e.g. retraining the models to ensure accuracy and reduce model drift or model decay) - some of the requirements are quite different than traditional DevOps.
Devops concepts such as reproducibility and consistency between stages is very much the same in ML Ops. It just goes beyond deployment into monitoring and retraining and redeployment.
But it is possible to apply CI/CD best practices and bring the speed and agility of DevOps to ML Ops. Last week HPE introduced our new ML Ops software solution focused on addressing these challenges and bringing a DevOps approach to the ML lifecycle: https://bit.ly/2lKOnBO
https://bit.ly/2lKOnBO
HPE accelerates Artificial Intelligence innovation with enterprise-grade solution for managing entire machine learning lifecycle
Bringing DevOps to Machine Learning | HPE ML Ops Nanda Vijaydev, from HPE's BlueData team, discusses HPE Machine Learning Ops. HPE ML Ops brings a DevOps approach to machine learning and allows enterprises ...
Bringing DevOps to Machine Learning | HPE ML Ops Nanda Vijaydev, from HPE's BlueData team, discusses HPE Machine Learning Ops. HPE ML Ops brings a DevOps approach to machine learning and allows enterprises ...
Gartner estimates that 80-85% of enterprises are running into the 'last mile' problem with ML model deployment and operations. And by 2021, at least 50% of ML projects won’t be fully deployed due to lack of operationalization.
One challenge I have seen is proving value on too many initiatives. We have had success with BlueData on proving one workflow or service and scaling adoption incrementally
primarily it is lack the right tools for the operationalization and large-scale implementation of ML models. In addition, there is a lack of standards around coding, sharing, and collaboration.
Ensuring the model was t built in a silo and that it actually solves the problem is was built for. Many business struggle to understand what was built and how to use it.
@patrick_osborne This is a complex and growing field. No longer controlled by specific vendors. Start small, prove value, and scale is applicable at every level
The dictionary defines operationalize as “to put into use” or “to make operational”. There’s a good Wikipedia entry about ML Ops at: https://en.wikipedia.org/wiki/MLOps
https://en.wikipedia.org/wiki/MLOps
MLOps - Wikipedia
MLOps - Wikipedia MLOps (a compound of “machine learning” and “operations”) is a practice for collaboration and communication between data scientists and operations professionals to help manage production ML (or deep learning) lifecycle.[1] Similar to the DevOps or Da...
Operationalization is enabling models to be deployed at scale and be seamlessly consumed by modern applications. And it's not just for better and faster insights. It's also about being able to take those insights and effectively apply them to the business.
As data science teams in the enterprise move beyond initial experimentation to production deployments for ML / DL, they need to operationalize their models and the end-to-end lifecycle
Operationalization in the ML context is to put the model in a DevOps like workflow with performance monitoring, versioning, etc. For operational success with machine learning, processes and tools need to support the iterative nature of the ML lifecycle
Establishing effective deployment processes, accurately scoring and serving models to the right areas of the business, measuring success and failures with the right performance metrics, monitoring ongoing model performance, and identifying when to retrain a model
It’s not just about deployment, it’s about the complete ML lifecycle. Operationalization requires a wide range of capabilities (including collaboration, reproducibility, monitoring, etc.) for the development, training, deployment, and management of operational models at scale
models don’t serve any purpose unless they can be put to use by the business. Operationalization is the ability to take a model and integrate into some sort of workflow that can feed it data and accept responses. Most companies struggle at this point due to the complexity of it
Being able to deploy ML models quickly, efficiently and in a routine way is very important - deploying and managing each ML model as "one-offs" an anti-pattern.