Machine learning everywhere: IBM banks on trust and open standards
AI holds tremendous promise to help technology-driven businesses soar to new heights. But getting started with AI can be a challenge. As companies build out the information architecture to support AI, it’s critical to get machine learning right so data and insights can be trusted.
Dinesh Nirmal, IBM vice president of analytics development, recently shared some thoughts on AI: where he sees machine learning and data science going, what challenges companies face and how IBM can help.
BDAH: What is the biggest challenge enterprises face around machine learning today?
DN: When I look at machine learning, I look at the “three Ds”: The data of models, the development of models and the deployment of models.
There are so many frameworks out there that automate development of models. Data scientists typically spend the majority of their time cleansing data and getting the data ready for the development of the model.
I always say, “Garbage in, garbage out.” If you can’t get data suitable for the model, you likely wont have good accuracy once you deploy it.
Once you get the model built, how do you deploy the model? If you have an enterprise running for 40 to 50 years, you will, in all probability, have third-party software that includes legacy applications and the dispersal of skills sitting in different places.
The challenge most enterprises are facing is: how do you infuse machine learning into all of it? You have to make sure all the angles come together for them to be productive in data-centric world.
BDAH: What’s been the response from clients around our approach toward machine learning?
DN: Our clients want to infuse machine learning into their infrastructure or enterprise. I think they will not survive without it. They realize that they need to serve their customer base and they are getting more data about their preferences in order to serve them better. That’s a very fundamental part of any enterprise.
Now there are two things you need to bring in machine learning or cognitive capabilities: tools and skills.
With the tools, how do you build these models? That’s where IBM Data Science Experience (DSX) comes into place. Now not only can you use the core infrastructure but you have a choice of which framework you want to use. DSX brings together the versioning and the governance of the models — and the tools become a critical piece in that journey.
The second piece is skills. How do you make sure you have the right set of skills to help you build those models? We’ve solved for both in the sense that that DSX has brought the tools in a uniform, collaborative manner and we have a team of data scientists available to help them build the models. Between those two, our customers get that IBM is serious about bringing cognitive capabilities to enterprises.
BDAH: With tools and skills now available, how close to democratization or mass adoption is machine learning?
DN: We aren’t there yet, but we’re beyond the buzzword space, because enterprises are serious about infusing machine learning into their infrastructure. We are not yet at the point where customers are reaping the benefits of say, a grocery store uses ML to know our preferences. We’re probably another 10 to 15 years from that.
BDAH: How well do you feel IBM is poised to dominate the machine learning space?
DN: The challenge enterprises have is that they need somebody they can invest their data in. IBM has been around for 100 years.. The second piece is that we bring in the skills enterprise need. We know enterprises in and out, and we serve the enterprise market well. We definitely bring skills, and our customers see it.
BDAH: What are your goals for 2018 to improve the tools?
DN: What we did in DSX is take open source tooling and gave customers a choice to build using their own framework. Some customers want to use Python, some want to use R, and some want to use Spark. They’re all available on DXS. You get strength of open source along with the enterprise capabilities like governance, management and monitoring of the model. And that is resonating really well.
If I look at 2018, the theme is “machine learning everywhere”, meaning that we intend to infuse ML throughout the IBM product portfolio where appropriate. From a tooling prospective, we’ll continue to improve tooling by bringing in the different open source capabilities and improving our enterprise strengths.
BDAH: Can you give an example of how ML is being infused in the IBM Analytics portfolio?
DN: Sure. There’s a long list including SPSS, IBM Integrated Analytics System, IBM Cloud Private for Data, Db2 Event Store, Watson Explorer, Information Server to name just a few. Another example is how we are infusing ML into Db2 where a predictive model will help predict access paths. We’ve done the work, but there’s a tremendous testing to make sure access path is continuously pulling accurate results.
BDAH: There have been a lot of revelations about the propensity toward bias in decision making across some industries. How can machine learning help tackle bias in the future?
DN: Loan processing is a good example where bias can happen. Organizations need to provide "explainability" of why a loan may or may not have been approved. Identification of bias is the first step to overcoming it. Machine learning can then help remove human, personal or emotional bias in decision making which might otherwise imply some form of discrimination. There are a number of steps teams can take - such as ensuring the training data is not biased to begin with, making sure models can explain themselves at an aggregate level as well as at individual data points (i.e predictions). So you need good data scientists to be able to check and monitor how the models are performing during training and deployment.
More than bias, explainability becomes a key piece when you use AI. That’s the benefit of using machine learning versus deep learning. With deep learning, it becomes hard to prove that there was any kind of bias or irregularity because it’s a multi-layered neural net where you cannot go explain it that easily.
BDAH: How can we then get around bias and avoid those kinds of problems?
DN: It’s a really hard problem to immediately chase and fix because of the data that you feed. Depending on who the machine is learning from, it’s going to have some level of bias, and you can only figure that during testing. To test every single permeation is going to be hard. That’s where constant modeling of the model is helpful. We built in some provisions into our monitoring tools whereby you can specifically ask for explainability in certain areas. But it doesn’t guarantee 100 percent coverage.
BDAH: With features such as tone analyzer coming into AI getting closer to decipher users’ emotions, we’re hearing a lot about the opportunity for brands to create relationships with their users.
DN: This is a new frontier to be solved involving tremendous amounts of data. How do I interpret users’ emotions? There are three ways: through the interpretation of the language used, the facial expressions conveyed and the intonation of the audio. if you ask me, that is more of a deep learning challenge which involves a boatload of data. It’s not for run-of-the-mill data science to solve. This is where IBM Watson can be leveraged to help companies succeed.
Ready to take your machine learning experiments into production? Check out the IBM Data Science Experience.