AT&T's goal is to combine big data analytics, deep diagnostic tools, software-defined networking and predictive capabilities to create an infrastructure that will react to customer needs instantaneously -- and, in some cases, in anticipation of future events.
Artificial intelligence (AI) is a key to this goal, according to Mazin Gilbert, AT&T Labs' assistant vice president for intelligent systems and platform research. The road to such a network is a long and fascinating one. Telco Transformation spoke with Gilbert about just what AI will do for AT&T and where fits into its network.
Telco Transformation: What is AT&T's approach to artificial intelligence?
Mazin Gilbert: There are three levels of sophistication in the journey. They ultimately lead to hyper-automation in which artificial intelligence really plays a role.
The first generation is automated systems. These are not just simple manual systems. They collect large amounts of data and analyze this data in a very sophisticated way and may perform some [task such as] anomaly detection, some form clustering, some form of data analysis. Then they try to infer some action that needs to be done.
The second generation of this paradigm is where we are implementing automation. The sophisticated automation here is that I am still collecting a massive amount data and sifting through it at the edge of our network at AT&T. We are processing massive amounts of data at the edge of our network, a competitive advantage of ours. In this new second generation we are able to tell -- based on sophisticated analytics -- that there is a potential threat. The system is able then to decide that if there are these types of threats what should be our next course of action.
They are pre-programmed on what action to take. It could be telling the system that if there are these types of threats then we want it to take this action. For example, maybe close down an IP address and stop the traffic from these people. Then there is an application or controller that goes and performs the action and it closes the loop. These smart very data-intensive control loops are starting to be implemented and deployed in our network. We are implementing a factory of these smart automated systems in our network today.
The third generation is hyper-automation. This not just advanced analytics and machine learning, but also very advanced artificial intelligence. The metaphor is playing a chess game. Your goal is to win the game. At any point you have to predict what step to take. But when you take a step the AI element is strategizing the next five [moves]. Even though I take only one step now, I am really thinking five steps ahead to win the game.
TT: So AT&T's AI is an extension of what came before?
MG: We are collecting massive amounts of data, not just from the network, but from the customer, the device, from everywhere. We can process it in a distributed way and now we are applying advanced deep learning to predict an event before it happens. Imagine I have these machines supporting you as a customer or supporting our network traffic or the systems supporting security. I need to predict in the next minute, the next hour or the next day what could happen. How confident are you? Are you confident that these systems supporting customers across region X or region Y are going to degrade in performance? Are you confident we are going to have potential threats in a particular region or IP address? Are you confident that we potentially are going to have customer X wanting three times the traffic, ten times the traffic in the next hour?
It is like dynamically adjusting our network. I have to predict what consumers or businesses will want in the next hour or next minute and completely change the paradigms. In generation three I am predicting the future. I am saying, "This what I think is the action we need" before it's a disaster. The AI is saying that based on everything I know, based on all predictions, these are the next three actions we need to take. And these are not hard wired. These are dynamic actions. Those decisions are changing every second of the day for our entire network.
TT: But anything that is predictive by definition may be wrong.
MG: There is a probability, a chance that something is going happen, many things. Millions of things could happen in the next second. Instead of waiting, we now are forecasting and attaching probability and saying, "Okay, what is the best path take? Do we wait? Or do we take an action, and when do we take that action?" That's the beauty of artificial intelligence. Even when you take an action you want to learn from the action. Some of the actions could be a failure. Not every action is a success because it is based forecasting and possibilities.
I could have 90% confidence that something is going to happen -- a particular router in our network in the next minute will fail. I could find after a minute that it did not fail. The idea of the data-driven machine intelligence in AI is that it is going to learn from that and change the confidence score. The same for a firewall, the same for a switch or a virtual network function of any sort.
TT: How does AI fit into the big picture at AT&T?
MG: We recently announced our transformation to a software-defined network where we are writing the network as software and driving it on commodity cloud-based hardware. There was one piece of that transformation called ECOMP. We are embedding AI and machine earning as a platform is the core of ECOMP that is deployed all across our network. That means as we create applications, the technologies and the platforms already exist. You are simply writing a design. (See AT&T's Rice: ECOMP Reaches Critical Mass.)
So what we putting the platform in our network? When have to create [applications] to do X, to do Y, to support these virtual network functions for customer Z. We are creating a model, a design. But the platform already is there. We are just distributing the model to this platform and the magic happens.
— Carl Weinschenk, Contributing Writer, Telco Transformation