AI and Machine Learning in the Enterprise today – “You can’t have AI without IA”
If you are deep into building the Intelligent Enterprise like I am, I think you will appreciate this talk by IBM’s CEO on the topic @ Davos. It is available in the Wall Street Journal – The Future of Everything Podcast, and is entitled ‘IBM CEO Ginni Rometty on Reskilling Workers in the age of AI’. The talk covers a lot of ground. My take-aways focus on where Enterprises are with AI today, and less on the re-skilling part.
Take-aways (GR - Ginni Rometty) with my commentary where appropriate (JC):
GR: Customer Service is Nr.1 area of applications today. HR is the ‘sleeper’ area, but with great potential in showing the value of AI to the enterprise.
GR: Many companies performing ‘Random Acts of AI’ because they don’t want to be left behind.
GR: 85% of effort in just getting the data
JC: Absolutely agree. Finding, collecting and cleansing data is the grunt work of data scientists. Most practitioners are using predefined algorithms. I would add that the art of AI/ML really lie in 2 areas: 1) Knowing which algorithm(s) (and their parameters) to use, and 2) Feature engineering – this, in my opinion, is where the real domain knowledge comes to play, and where the real secret sauce lies.
GR: You can’t have AI without IA - Need to have an Information Architecture
JC: No doubt. This is precisely why so much time is spent finding and preparing data. A new category of products are coming to market under the moniker of MLDC (Machine Learning Data Catalogs), attempting to address parts of this challenge.
GR: Applying AI without considering/re-thinking your processes/workflow misses the point and value that AI can bring.
JC: This is probably the most important point of all. Part of the challenge in many enterprises is knowing, in detail, what the processes and workflows really are. Part of the reason is that they cut across multiple systems and include manual/human steps. It would seem so obvious, but this knowledge is held in silos and not necessarily mapped out end-to-end.
GR: AI should be explainable – not black-box. People need to trust it and believe there are no embedded biases.
JC: Certain algorithms are far more ‘explainable’ than others (e.g. decision trees vs. neural networks/’deep learning’) – so this will be a challenge. Philosophically, converting ‘data’ into information/knowledge necessarily involve assumptions, and in assumptions, invariably, there are biases. So I don’t really buy into a ‘bias-free’ world view. It is more a question of minimizing bias.