How to deploy AI in an Organization?

I was motivated to write this post after seeing an article on AI. A researcher at MIT observed that the share price of Berkshire Hathaway went up every time an Anne Hathaway movie was released. See the post here https://www.dhirubhai.net/posts/vaibhavjain87_licreatoraccelerator-activity-6951025077446418432-VhXt?

This story reaffirms the need for subject matter experts in end to end AI/ML deployment.?It is all about the ‘context’ that AI/ML will not be able to understand. At least not in the short term.

So how do we deploy AI in an organization? Here are my recommended steps. While my focus here is Engineering and Manufacturing, the steps wouldn’t be too different for other industries:

1.??????When an organization is at the inception of its AI journey, there isn’t much choice but to hire a few Data Scientists. Preferably with some background in Engineering or Manufacturing or the industry you are in. I am using the term Data Scientist with a broad brush to include Data Engineers, Architects, Machine Learning Engineers etc.

2.??????The job of the Data Scientists is not to start revamping the processes in the organization. This is a recipe for disaster. The subject matter experts (SME) have acquired their knowledge with education followed by years of on-the-job training and experience. We should pick a few SMEs who are willing to learn about AI and ML.

3.??????The newly hired Data Scientists should start training the SMEs on AI and ML. Get them familiar with the terminology. Follow it with training on some basic ML algorithms and continue with Deep learning at a very high level.

4.??????Once the SMEs have had some basic AI/ML training, start picking up some pilot projects. Don’t worry too much about business impact. Focus on projects which are easy and can be accomplished in weeks to a few months. The idea is to create excitement.

5.??????Successful pilot projects will create the necessary momentum. Now, pick some more critical projects. Involve the SMEs in data cleanup, dataset selection for training/testing, feature engineering, model development, interpreting misclassifications, Hyperparameter tuning and troubleshooting in production etc.

6.??????Larger projects will rarely work at the first go. So make sure all stake holders (Management and SMEs) understand that this is expected and we need to iterate: Define problem -> Collect Data -> Feature Engineering -> Create ML Model -> Test Accuracy -> Jump to Collect Data and repeat.

7.??????Shorter the iteration cycle, faster will be the success. An important element here is to avoid the temptation that large data and complex model means better results. Instead, focus on improving input data quality. Make sure there are no misclassifications in the training data. Simpler the model, less is the computation cost (time, money needed to train ML algorithms and deploy them).

8.??????Once you have tasted success in some critical projects, you have collected critical mass to define the AI strategy for your organization. This is counter-intuitive since usually you start with strategy. So why are we doing strategy later? That’s because, at the inception of the AI journey, we don’t know what we don’t know. Further, AI is also evolving. So it is hard to shoot a moving target.

For further reading, I highly recommend the AI transformation playbook by Andrew Ng. PhD. He has highly influenced me and many others in this exciting AI journey. https://landing.ai/resources/ai-transformation-playbook/

Bottom line: AI/ML will not replace subject matter experts (SME). Involving SMEs in AI/ML deployment will improve efficiency and effectiveness. Excluding them will cause errors and chaos. if you have any questions, feel free to reach out to me on linkedin.com/in/milind-sawant-phd-6ab68114. Good luck!

#ai #ml #engineering #researchanddevelopment #manufacturing #digitaltransformation