AMA Highlights: Matthew O’Kane Head of AI & Analytics EMEA at Cognizant
Our latest AMA was with Matthew O’Kane the Head of AI & Analytics EMEA at Cognizant. Matt has two decades of experience in the analytics industry. He started his career as a data scientist and moved into machine learning very soon after that.
He was our guest on the 6th of December and answered questions on everything from AI disruption to HRtech. You can read some of the discussion highlights below or check them all out over in our Slack Community.
Question from Will
Afternoon everyone! Hi Matthew O’Kane What do you think will be the biggest AI trends in 2020? Is there anything specific you think we should keep an eye on?
It will all be about scaling. Everyone is beginning to understand the benefits of AI but currently it is really hard to implement at scale. So technologies that help with AI operations, data stores specifically built for ML (see Uber’s Michaelangelo for a great example) and governance tech will be really key for 2020.
Hey Matt, it’s interesting that you said ‘Everyone is beginning to understand the benefits of AI’. Do you still feel business in general has a long way to go before AI is adopted fully across the board? Do you still have to justify the spend on AI in top-level meetings?
It is becoming a lot easier to justify spend. A lot easier than 15 years ago when I tried to explain to businesses what a Randon Forest was! But we still need more success to really bring everyone along across industries.
Question from Anil Gunesh
Hi Matthew O’Kane, Could you please tell us more about the Analytics technology/ platforms you have been using in your day to day job and how this is likely to change in the next few years?
Hi Anil. This area is changing so rapidly. When I started my career I used SAS and SQL. Then more ML shifted to R. Now unfortunately I don’t get many opportunities to do coding but I turn to Python when trying out new techniques. There has been a big shift to cloud over the last two years and that really is the future. Then you need to consider data platforms, ML ops platforms, feature stores etc. etc. It is all getting very exciting!
Question from Charan
Hi Matthew O’Kane, Working with a client, how do you prioritize problems which the AI & Analytics team can solve? Is there an approach that you follow to measure the impact of various ideas/problems the client/team puts out?
Thanks, Charan. Value and ease of implementation are the primary areas of prioritisation. But I would also add that some use cases really get everyone excited in the organisation about AI. They might not have the highest value to the business but winning hearts and minds is key with initial use cases.
Question from Sam Davis
Morning all, hi Matthew O’Kane. I work in HR and I’m really interested in AI’s growing role in my sector. What do you think I should be looking out for in 2020 in regards to HRtech and AI?
Hi Sam Davis. HR is interesting. There are already a number of use cases in areas such as employee retention, recruitment screening, training needs identification. But there have been some missteps in HR AI – particularly around the use of screening AI.
The problem is that a machine learning model will replicate the human bias in any data we give it. So feeding an AI a load of CVs and whether those CVs were selected will just amplify the human biases to certain candidates. This is now being fixed through explainable AI and the use of real randomised experimental data to remove our inherent biases.
YES, this is one thing I’m really interested to learn more about. Do you think it is possible to create a truly unbiased machine learning tool that isn’t influenced by human bias? I like the idea of using AI to screen CV’s but I also think rigorous training in unconscious bias can solve that problem too.
I think that if there is any human selection in the process then there will be some bias in the model. But unconscious bias training and some randomisation of CVs being seen can get us towards a really fair system
True. I’m also on the fence about the ‘gut feeling’ you get about people when you’re hiring. You don’t want to be biased and AI can eliminate that but you also want to make sure you’re hiring a good culture fit for your business. Do you think AI can predict culture fit for a company or is that where the AI in the hiring process ends.
If you had 360 feedback a year into every successful candidate’s role then you might be able to start seeing what influences cultural fit. But it is hard to gather that type of data currently.
Question from Mark Pybus
Hi Matthew O’Kane, There has been a rise of ‘Low Code’ ML platforms – offering to automate the ML process. Is this a trend you’ve seen and does it excite or worry you?
I think anything that automates or improves the ease of ML development is great. AutoML has really taken off and even open source packages in Python can do amazing things. The only concern is that the business problem can be lost in the mix. For example, if you want to reduce customer churn building a predictive model that predicts churn doesn’t help the business case. Instead, you need to construct a model that tells you how to reduce churn in a prescriptive way. This is beyond most AutoML packages today so I worry that people might miss this key element.
Yep, I’ve concerns that as well as domain knowledge loss the data sourcing and prep is not given the attention it needs, leading to Models that, even if understood, are not really as suitable for the purpose as these platforms GUI’s suggest.
Agreed – half the problem is on the data side around feature engineering. I’m surprised there aren’t more platforms devoted to the data side of ML development but working with a couple in stealth mode currently
Aye, increasingly it’s possible to get decent ML results with the domain knowledge and data engineering. Do you see that tackling the ML Ops and scaling the use and deployment of models across a business can deliver much more return than chasing the last percentage improvement in fewer models?
Spot on – it might be more boring but getting simple Decision Trees, Regression or GBMs in as many use cases is a clearer and quicker way to value.