Machine Learning Operations
Adoption of AI across all industries both private and public is growing rapidly, MLOps or Machine Learning Operations is one of the hot topics for 2020.
So what is exactly MLOps and how is it going to meet the growing need for AI adoption? To understand that, we need to go over what is Machine Learning and what is IT Operations.
Machine Learning
There is no single definition of ML. I’ve come across many. Just a couple of them to note:
“a tool for turning information into knowledge”,
“ enabling computers to tackle tasks that have, until now, only been carried out by people”,
“algorithms that find and apply patterns in data”
One another definition, a picture for a thousand words to show how ML fits in Artificial Intelligence and Deep learning.
I really like this picture as it shows how ML fits into AI territory and how it relates to DL. Right conclusion from the above is that ML is just a part of Artificial Intelligence and really important point you can’t miss is that there is no equal sign between the two. AI in general can be split into two buckets “Strong Artificial Intelligence” and “Weak Artificial Intelligence”. Strong AI is referred to be human level intelligence while Weak AI is everything else designed to solve one particular task i.e. face recognition, house prices, document translations and so one. As you probably guest it there are plenty of algorithms and methods in the bucket labelled ‘Weak AI’ while the bucket labelled ‘Strong AI’ is still empty. I have to make one remark here about ‘Weak AI’ label, term coined long time ago, which if you smile when you read ‘Weak’ to refer to all the great achievements ML brings to the modern world you are absolutely right it’s not that Weak any more, however if you compare it to the definition of Strong AI it is still nothing close to it.
Machine Learning to be able to solve tasks needs 3 things, 1. Data, 2. Good algorithm and evaluation techniques to be able to produce 3. Accurate Model. In this very simplify description complemented with below diagram I’m providing a definition that will help us in the further discussion on MLOps.
Operations
Let’s look at IT Operations definition and how it is being structured by IT companies.
The IT Operations is a team of people or more broadly used in the definitions ‘an entity’ that makes sure company has right processes and procedures in place to deliver a product, infrastructure or a service to the customers with control and maintenance on continues basis. IT Ops team focus on delivering stable service within agreed levels of service. To be able to do so service monitoring and control needs to be put in place.
To give you an example if you provide any kind of online service whether this is an online game, social type of website or store, depending on the size of your service you need a process to allow users to report an issue with the product (Service Desk), someone should take a look on the issue and decide what to do with it(Incident Management), you would than need to solve that issue (Support team). After the issue is resolved for the client, ideally someone should review this issue and make sure this will never happen again (Problem Management/Service Improvement) and when you already have the fix, implement it (Change Management). Those are examples of reactive behaviour while IT Operations focus on proactivity too, for example it might be improvements from Monitoring activities, outcome of brain storming sessions an so one. This and many more is what IT Operations look after.
Why the demand for MLOps
Number of applications for machine learning models is growing rapidly. Let me shared some data released by Google on theirs adoption of Deep Learning which says it all. Similar thing is happening all over the IT from startups to large enterprises.
Not mentioning that according to Forbes, in next five years AI marked will grow almost 8 times !
Machine Learning + Operations = ML Ops
Intersect of the two above definitions IT Operations and Machine Learning is what we call MLOps. It applies to Machine Learning models the things IT Operations perform for any other IT service. To be very specific it contains
- Production life cycle management for Machine Learning applications
- Govern production models
- Monitoring for Machine Learning
Microsoft explain key MLOps phases as:
a) Build and train reproducible models which aims to achieve reproducible pipelines of model generation
b) Package and deploy models — which is a technique for encapsulating models into a container image and deploy it.
c) Automate workflows, monitor and manage — use GitHub to introduce version control to the models and manage new roll outs.
d) Apply governance and control — audits who publish the model, when and why the changes are being made.
As you can see by now MLOps is nothing else but applying well know IT Operations techniques to new and raising demand of deploying and managing Machine Learning models in production environment. Don’t get me wrong, this is not brainless process, IT Operations still require some tuning to be apply to Machine Learning however it is an improvement to the process rather then revolution. The same way Operations Management adjusted to Big Data demand and others in the past. Part of MLOps that requires more innovations then processes is ‘The Tooling Department’ where there is not so much out of the box solutions that helps drive MLOps.
Hope this article provided you an idea on what ML Ops is and that it’s not that different from IT Operations in general that is well know process from quite some time.
Feel free to connect me or ask any questions via Linkedin
