You have trained a model that gives you high accuracy on validation and test data. You were very happy that your POC was a success. High on spirit and motivation, you deployed the model you just trained on production. But now you observe, the accuracy is not acceptable on actual production data and the accuracy of your model is decaying day by day.
Have you ever faced such issues?
What is causing such issues?
We have answered the "WHAT" of the issue in Part I of this series. We recommend reading it before going further.
Let's talk about what can be done to overcome all the challenges and deploy a scalable AI solution on production. Let's look at some of the pointers that you need to keep in mind.
The most important step before starting the implementation is to define the scope. This step involves identifying and understanding the problem we are trying to solve here. The scoping process can be broadly divided into 5 steps:
Step1: Brainstorm the business problem
Try to understand the problem to be solved and identify if an AI solution is actually needed or a simple software engineering solution can be applied here.
Step2: Brainstorm AI solution
Finalize the approach, algorithm, and data strategy to be applied to solve the problem.
Step3: Feasibility Check
Identify the feasibility of the solution proposed. If the solution we are proposing, does not match with the business requirements or infrastructure, going any further with it will gain us no merit.
Step4: Prepare a Roadmap
Chalk out the future steps which will be taken to make the solution scalable. This gives us a futuristic view of what is in the store and helps us to plan accordingly.
Step5: Money Money Money
Everything comes down to the dollar value. It's very important to finalize a budget for the solution. This not only helps us to keep the expenditures in check but also helps us in pricing our product/service accordingly.
The most widely used approach while building AI solutions is to "build fast and fail fast".
A typical approach to build an AI solution majorly involves the following steps:
- Train your model on the available dataset
- Monitor the performance of the model on test data
- If model accuracy is not up to the mark, go back to modify the model architecture and then repeat the whole process.
There is nothing wrong with this approach, in fact, this is the most efficient approach for the use-cases where there is abundant data available.
But this approach acts as a bottleneck for some use-cases where data is not available in the required quantity. One example can be the healthcare industry.
In such scenarios quality of data - that clearly illustrates the concepts we need the AI to learn, becomes more important than the quantity of data.
The transition from Big Data to Good Data.
So instead of going back and changing the model architecture, training the model on good quality data can gain us more accurate results.
Bridge The Gap
Often there is a gap observed between the ML engineering team and domain experts. And to ensure the success of our solution on production, bridging this gap is very vital.
We discussed how taking a data-centric approach streamlines the AI development process. To ensure we have only the relevant data, knowledge of domain experts is needed. This knowledge cannot be gain overnight.
This makes it very important to have a collaborative effort between business experts and the ML engineering team.
Once the AI solution is deployed on production, we need a mechanism in place that can help us monitor the accuracy of the model and helps us improve the accuracy with time.
To achieve this in the AI project that you intend to take to production, be sure to plan the deployment process and provide MLOps tools to support it.
But wait... What exactly is MLOps?
MLOps is a perfect blend of a Machine Learning Engineer and a DevOps Engineer
MLOps is an End-to-End pipeline for any AI-based product. It takes care of Data Ingestion, Model Architecture, Training, Tracking, Evaluation, Deployment, and Feedback.
There are multiple tools available in the market which may help you with the MLOps problem. But before you pick up any tool, it is important to identify the problem these tools are solving and does it match your requirements.
Read more about MLOps here. https://chroniclesofai.com/know-how-of-mlops/
AI solutions have a lot of scope in the future. AI-based products have already touched major parts of our lives and this is only going to increase in the coming future.
Just like electricity has transformed the way we operate in the modern world, AI has the potential to become something like that.
But the developer community needs to understand the need of the hour and adapt a modern approach to AI. The Data-Centric approach clubbed with MLOps is going to be a game-changer.
What are your thoughts on this? Feel free to write to us and convey what you think about the future of AI.
Hope this article was able to spark some ideas in your mind. Stay tuned for more such content. STAY SAFE.