Just recently we celebrated the 4th anniversary of Konverge AI. As we reflected on our 50+ AI implementations, we thought about creating a blog series based on our learnings, mistakes and experiments. This blog series will help the AI fraternity avoid a few critical mistakes, provide a point of view based on our experience and (hopefully) serve as a consultative body of knowledge.
Over the next few weeks, I will be sharing my insights based on 50+ implementations across industry sectors and project life cycles. I just took out some common themes, best practices and tricks of the trade, if I may use that phrase and sincerely wrote about them. I hope this series helps you.
In AI, well begun is half done, really!
How identifying the right use case to pursue can be the difference between winning and losing your AI gambit
The Artificial Intelligence and Machine Learning (AI-ML) revolution is truly underway. All enterprises, large and small, understand the impact AI-ML can have on their competitive edge. Nobody wants to miss out and hence we see a surge in new AI-ML projects. As an AI expert, while I love all the attention and interest in the field, there are a few worrisome statistics:
- According to Gartner, 85% of AI projects fail to deliver value
- 70% of companies report minimal or no impact from AI
- 87% of data science projects never make it into production
This makes me wonder why so many AI projects fail. So I decided to create a series of posts on why AI projects fail and more importantly, how these failures can be stopped. The first thing to do is to identify the right problem to solve.
Well, while AI seems omnipotent, not all problems can be solved using AI or not worthy of being solved by AI. So before you begin your AI journey, choose the right set of problems to solve using AI. So what kind of problems can be solved using AI?
AI can solve the following typical problems:
- Image classification, recognition and segmentation problems
- Object Detection & Recognition
- Text Classification
- Text Analytics
- Text Sentiment Analysis
- Text Similarity Problem
- Natural Language Understanding Problems
- Natural Language Generation Problems
- Classification, Clustering, Forecasting/ Prediction Problems
In general, any problem where we can use the existing data attributes to train a computer to classify the new items on their attributes or predict the attributes of new items, can be solved using AI.
So any problem that does not fit the description above will be unlikely to be solved using AI. Having said that, does it mean that a problem can be readily solved using AI if it has the attributes mentioned above, it can be readily solved using AI? No, it’s a necessary but not sufficient condition.
For AI to be successful, the training data has to be available, clean and sans any biases. Data not being available is one of the biggest impediments in deployment of AI powered solutions. If you don’t have the right data input, your output will not be correct.
The other perspective of looking at the right problem to solve is the impact perspective. While choosing the problem to solve using AI, we must look at the potential business impact the solution of the problem will have. Solving a problem with minimal business impact using AI is like using a laser cutter to chop vegetables in your kitchen. It does look cool; however, it is a sheer wastage of capacity. So choose the problem that will have a profound business impact.
In my next blog, I am going to write about how the lack of right data leads to problems in AI. Till then, I will focus on solving (the right) AI problems we have to solve for our customers.