Musings about AI

Conversations with Gyan...

Gyan is back from US after a successful stint in the MNC. Meeting him after a long time was the delightful experience. In between our chitchats about the Corona, Economy, Education system etc. we started discussing Data Science. From Data Science to AI to AI Models to AI models in production. Some Excerpts from our discussion…

AI Models & Deployment

Gyan: With surmounting interest in data science and the fast growing Data Scientist community, AI as a technology has come a long way crossing the chasm from Innovators and early adopters to the Early Majority. Along with all the hype that’s there today around AI, there is still the unaddressed issue of less than 12% models reaching the production stage

Me: That’s astonishing, 12% is too low. Why is it so

Gyan: Data Scientists are creating models day in and day out but there are millions of models that are still waiting to see the light of the day in production. 

*Gyan drew this on a piece of paper..I am trying to recreate it..

Me: What do you think are the challenges? 

Gyan: I was reading the Gartner report the other day. Gartner identified time to value as one of their biggest challenges, reporting that it takes an average of 52 business days for the team to build a predictive model, and longer to deploy into production

Me: Don’t you think that the deployment should need fewer days then building a model

Gyan: Yes it should but it depends

(In the meanwhile Tea arrived…Taking his fist sip he continued)

I think the larger issue here is operationalizing AI. I believe there are significant challenges in Operationalizing AI

Me: Like?

Gyan: Some of them are:

Analytics challenged leadership: This one serves as the major hurdle in operationalizing AI. The senior leaders in the Organizations are not that conversant in AI and hence lose out to the AI Integrated Organizations. Leadership support is the most important factor in operationalizing AI.

Our Organization started losing to the new start-ups in this space, that’s when our CEO decided to adopt AI. It took us 3 years to regain the market share which we lost due to new entrants. Our CEO made it mandatory for the entire leadership to pursue courses in Data Science, that was just the beginning, we started having weekly knowledge sharing sessions, Innovation lab workshops etc.

Me: That’s really great to know. I believe it’s really important for the change to be initiated and adopted by the senior leadership

Gyan: Exactly! It’s easier said than done. It takes lot of focus and perseverance to bring change

Me: What else do you see as the reasons?

Gyan: Well…(Enjoying the last sip of Tea)

Me: Do you think Data Quality is also an issue

Gyan: It’s not the issue my friend, it’s one of the major issue.

You remember Gupta Sir (Our beloved Computer Science Teacher in School – We have had lot of great memories with him – Let’s keep it for some other day)

Me: (Smiling) Yes – how can I forget him

Gyan: He used to say Garbage in Garbage out. Same is the story here

Data Quality: Gartner predicted that through 2022, 85 percent of AI projects will deliver erroneous outcomes due to bias in data, algorithms or the teams responsible for managing them.

Testing and Validation is practiced in the controlled environment and hence the data used is of good quality while when it comes to deploying the model in production, it has to work on real world data which in most of the cases lacks quality which results in low Model accuracy

Another reason is Legacy Infrastructure: Another big challenge in operationalizing AI is the issue of legacy infra of the large organizations which makes it impossible to operationalize AI.

Me: I think the Infrastructure for the AI models is very different

Gyan: Yes and one of the major factor is

Managing the Compute power: Compute power requirement depends a lot on the kind of model being worked upon. In case of deep learning, computer vision models the compute power for the training data set is high while we can use 1/nth compute power in production. But for the KNN models the compute power in production needs to be high. So optimizing the compute power in each stage of analytical modelling is key to the success of Operationalizing AI in organizations.

Me: That I have witnessed, while we were building the Deep Learning models. We had to increase the compute power Multifood to run it on training data set

Gyan: Yes and there are so many Organizations out there who are still looking for an answer to the standard compute power. Though advent of cloud and its adoption is helping Organizations in making it possible but still it’s the long way to go.

Me: Do you think one of the challenges in Operationalizing could be the Interpretability

Gyan: Yes it is, Interpretability is essential for operationalizing anything today. People need to understand What’s going on in the black box. This goes back to the first point we discussed around Senior Management understanding of AI. The first step to start any data science initiative in the Organization is to present the business case or get the budget. The moment you decide doing something using Neural Network to achieve High Accuracy it becomes a black box for the management as the high accuracy comes at the cost of low interpretability. Though as a concept you can explain the management on what’s happening within the solution but mostly the solution is a black box for the organizations. For example risk scoring using the Neural Networks can be highly accurate but difficult to explain.

Me: Very True..I think it would be easier if Senior Management understand the constraints of the Data Scientist.

(Gyan was planning to leave now as its been 3 hours since we were sitting in the balcony and discussing this)...I stopped him..

Me: I think one important issue which we missed is also around the Data Ownership. In my last consulting assignment with one of the large corporation. We faced the major issue of departments working on Silos

Gyan: And why do you think that’s the issue Mr. Consultant

Me: I think not directly but indirectly it should be the issue as these large Organizations are collecting data from years now but I think if we want to create AI models use Neural Networks, there would be more and more data required and Collaboration is the only way forward as more the data, better the training, better the accuracy.

Gyan: Well I think that’s an interesting angle to look at it. Lets keep it for next time…The Data Ownership and Collaboration

(Gyan had to leave as he was expecting few guests at his home. But as He is in India now.We have decided to meet every weekend and spend time Musing about AI)

https://www.forbes.com/sites/gilpress/2019/11/22/top-artificial-intelligence-ai-predictions-for-2020-from-idc-and-forrester/#4fcfebef315a