Artificial Intelligence in Geospacial Intelligence in 2019 – From Experimentation to Operationalisation
We sat down with Omar Maher, Director of Artificial Intelligence at ESRI to ask a few questions about how ESRI takes AI models from experiment to actionable workflow. Here’s what he had to say:
What is GeoAI?
GeoAI is the intersection between location intelligence and artificial intelligence. In my opinion, it's a very important branch of AI that few people are currently tackling. If you look at the AI landscape today, progress is mostly happening around high dimension data, highly unstructured data, and of course, operationalising machine learning. Fewer people are focusing on problems which have a location element by design.
There are some specific problems that are location-centric by definition, for example, Activity Based Intelligence (ABI) to find interesting and abnormal spatiotemporal patterns, and Structured Observation Management (SOM) for Comprehensive Situational Awareness for Objects of interest. GeoAI is this branch where we intersect both the power of GIS to understand location, along with the capabilities of machine learning and artificial intelligence.
How do you train or build a staff capable of leveraging AI?
First, you need to have data scientists who are able to build models. Those are people with a good background in, mathematics, machine learning, and data engineering. Then you need to have analysts – people who will be able to work collaboratively with the data scientists, consume their output, and build meaningful workflows connected to AI. The third skill set is business analysis; the ability to model problems into technical workflows.
Another crucial aspect of a staff that can effectively leverage AI is their culture. The vast majority of the projects I work on don’t have a perfect model or solution from day one. We need patience to produce valuable results from AI. It's a very iterative process, so the culture must encourage a ‘fail fast’ environment.
How can we close the gap between AI experimentation and operationalisation?
Experimentation is the stage where a data scientist attempts to build the most accurate model, experimenting on their laptop or machine, to get the most accurate results. Operationalisation, on the other hand, is the idea of taking that model and deploying it into production.
AI is a means to taking action. To make AI actionable we need a way to put models into production. Then we need a way to easily connect these models to different workflows using APIs. We also need a rich set of tools to enable analysts and decision-makers to build applications and workflows on top of these APIs. For this to happen, we make sure we have all of those mechanics in place, instead of inventing them anew every time we want to operationalise a model.
What are the most important things you'd recommend for organisations looking to start capturing real value from AI?
A pitfall I often see, usually with customers that are really excited about AI and its capabilities, is that they follow a technology-first approach. You need to focus on the specific problem you are trying to solve before you decide the best way to solve it. Many problems today don't require AI or machine learning; they can be solved with traditional tools.
I'm in favour of simplicity. The best way to find a simple solution is to think of the business problem first. For example, if you want to predict something or if you’re trying to find a pattern, then recognise that these are two different things. Each one has different machine learning algorithms, and requires a specific kind of data. If you want to find abnormalities and outliers, that is another set of algorithms. Think about the problem, break the problem into sub-problems, then think about what tools or pattern of AI you want to use.
Omar Maher will be speaking at DGI 2019 on how GIS will evolve to cater for the human machine team. Download the agenda for more details.