Artificial intelligence (AI) is transforming how federal agencies can put their data to work to advance the mission. Matt Langan, Government Technology Insider’s podcast host, discussed how AI has evolved over the years and what challenges agencies are facing when implementing AI solutions with Kal Voruganti, Senior Fellow and Vice President at Equinix, and Scott Andersen, Distinguished Solution Architect at Verizon.
Matt Langan (ML): Thank you both for joining us today. How have artificial intelligence solutions evolved over the years? What are the three generations of AI?
Kal Voruganti (KV): AI has been around since the 1940s and 50s. Historically, in the first generation of AI systems, the subject matter expert would specify rules explicitly about how a system should behave. Those systems were brittle, since they were unable to think ahead of time about all the possible rules and corner cases that would need to be covered.
Then, the second generation of AI picked up steam with the era of big data. Once we started collecting a lot of data, people started to use statistics and statistical analysis techniques. Also, another key development are the graphics processing units (GPUs), which are historically for graphic cards, were repurposed for AI. The combination of big data, use of GPUs, and clouds democratized AI by making these large compute forms and algorithms available at the click of a button.
Now, AI has started to percolate and be pervasive in every aspect of our life. However, the problem is that most agencies don’t have all the data that they need to build accurate AI models within their four walls. Third generation AI is where companies have to collaborate in order to get all the data that they need. There is a need for governance, otherwise people are hesitant to share their data because AI is like the precious jewel. They don’t want to share their data and have the receiver use it in unauthorized ways. Now we’re entering the era of third generation AI, and governance, sharing, and distribution will be very key.
Scott Andersen (SA): First, I want to add that there’s a 4th generation of AI coming; I know it doesn’t exist today, but it is coming. In the 4th generation when an AI responds to a unique variable within a solution set and creates a new AI on its own to solve a new problem. When we think about the impact of AI and what AI is doing, we really go back to the beginning. If you think about AI in the stages, first of all it is automation. The very first thing is “let’s automate things and make things better.” If you think about the automation processes that really were the most successful in the beginning, they were replacing human beings because sometimes humans make mistakes. Machines don’t necessarily make mistakes, unless you consider the fact that a human being created them and might have put a mistake in it. The reality is that the machine is able to build the solution without making the mistakes that a human normally would. So, the first thing is automation. The next thing is solving issues.
Many years ago, I was sitting in a government forum and a director of an agency was out speaking and said, “do you know what the problem is? The definition of big data is wrong. The definition of big data is any amount of data that overwhelms the system that’s taking it on.” As you think about what AI allows us to do, AI allows us to actually create a more broadly spread approach towards the concept of solving AI or solving big data problems. We can begin to chunk data. We’ve gotten smarter and better in the data age as we enter that third phase of the next component of AI. That’s where you’re going to see some fairly significant changes. If you look at how networks operate today, the concept of application aware routing is the first implementation of AI. This application is critical within the organization.
If you think about how applications operate, some users are going to be more critical. My favorite example is that there’s an ambulance barreling towards a hospital. With 5G, that ambulance is live connected with no latency to the hospital, so the doctors have real time information about what’s going on in the back of that ambulance, what is happening to that patient, and what are the impacts of what’s being done to that patient. When that ambulance pulls into the bay, the doctors are prepared and already know the course of action and treatment for the particular problem of that patient. That’s the real time reality and what we’re going to be able to do now with AI is begin the concept of even filtering the information, because today there is information that that doctor doesn’t need. So, let’s build an AI infrastructure that allows us to filter that operation. Not only is that doctor real time informed, as it is today, but then the next iteration is that they’re going to be informed about what needs to happen directly for that patient.
ML: What are some of the challenges that organizations face as they are trying to implement AI solutions?
SA: The reality of AI today is that there’s a learning curve. We’ve built a lot of automated and AI solutions to solve the big data problem, but the reality we’re facing right now is that there are many other issues that we need to take a look at. As I said before, the initial phase of AI was ’let’s get to automation, and automate to make things smarter.’ Automation has to be able to respond to a variation in a variable within the solution, and the same for consuming data. We need to be able to take a look at how data is applied. As organizations begin to look at how to impact the reality of AI, there needs to be a change.
First of all, 5G buys you low latency. I’ve worked with developers for years and I’ve never met a developer that built latency into their app. Taking latency out of the network and out of the world around us is critical for improving the quality of AI interactions. The other thing that’s evolving that’s going to be a game changer for companies as they try to implement, and the big initial barrier is the reality of edge computing. The reality is edge computing is this amorphous growing thing, just like AI was ten years ago. The reality of edge computing is that it lets us get the user, the device, and the data into the same physical location, which is relative.
For example, geologists can be studying the edge of a volcano. Physically, the sensors in the volcano and the device the scientist is holding is at the bottom of the volcano. We don’t want the scientist standing up at the edge and visually looking down. Instead, we can send a drone over to do that visual look and see. The scientist at the bottom of the volcano is studying the data from the IoT device, and overtime, the scientists can use the uniqueness of the human mind to actually begin to evaluate what an AI has started to put together.
An AI will pull together all the sensors across the volcano, all the hotspots of that particular volcano, all the different things that that volcano has always done right before it erupts. All of those sensors will be provided to the scientists, and they can look at it and can begin to assess and deliver a much better view of when that volcano is erupting. The barrier to date for most organizations is simply that the application of AI hasn’t been as easily implemented. As AI gets more mature and more broadly utilized, it’s going to be up much easier application for customers, but today it’s more specialized in what it solves.
KV: I completely agree with Scott as the amount of data increases. For example, airplanes generate logs today with around four terabytes of data per day per plane. It’s hard to move all the data to a central location. It’s very expensive and in many cases, you don’t want to move that data. Instead of moving data to compute, we are now saying we want to move compute to where the data is. If a plane lands in San Francisco, you don’t want to move the data all the way to Atlanta processing. Instead, you want to process the data where the plane is landing. This is the same in the case of a tank in the field, or a car, etc. The key point to note is that people are looking for new architectures to cut down on data transfer costs for real time latency. People want to do processing or analysis in real time, where the data is generated. Also, for security and privacy reasons, you don’t want to be moving raw data outside your four walls. For all those reasons, it is now time and necessary to move compute to where the data is residing. That’s one of the major challenges agencies are having.
The second key challenge companies are having is the whole idea of having enough data scientists, and the way the data scientists and the subject matter experts interact with each other. AI requires data scientists who are experts in AI algorithms and subject matter experts who really know how the system should behave. For example, how an airplane should behave, or a car should behave, but they don’t have a good grasp of the AI statistical algorithms. There is an impedance mismatch that is leading to delay in the implementation of projects. It’s hard to find enough data scientists in a company who can go and solve the problems for the finance group, for the legal team, or for the product team. There is a need to simplify it, so that experts can use AI directly or both democratize AI for the larger subject matter experts in all the different industry verticals. We are at a cusp where that needs to happen, otherwise we won’t be able to scale the implementation of AI solutions.
The third main area is that people are now beginning to start to think about the area of security, privacy, and auditability. AI is essentially building models. There are two phases to AI, the training phase and the inference space. In the training phase, you take the raw data, you create a model, and then you use that model for prediction. Subsequently, that’s called the inference phase. The type of data that you used to create your model, if it is biased or if it has security issues, your AI model will be compromised.
It’s extremely important to see how that AI model has been created and what data has been used to create that AI model. If you create an AI model for traffic flow patterns in Dallas, you cannot use that in New York, because the data is different. You have to know the linear of the data, and in many cases, people are not building AI from scratch. They use models that have been prebuilt as a starting point, and then they use their own context to customize those AI models.
You have to know the lineage edge of that model, who built that model, what systems were used, and what data was used. For example, there was a company where the HR team was using an AI model as part of their hiring application, but certain groups of people were not getting hired into that company, because the model was biased. It wasn’t trained in a certain way. I think the issues of security, governance, auditability, and privacy will become very difficult, and they need solutions. Taking models into industrial grade production and addressing these issues is a challenge. I think those are the three main challenges most organizations are facing today.
To learn more about artificial intelligence, click here.
Click here to learn more about Verizon’s Professional and Managed Services.