Certificates of achievement from the Israeli defense establishment cover the wall in Professor Oded Maimon’s office at Tel Aviv University. All the attempts to glean details on the awards come up against a stone wall. Maimon is unwilling to discuss whether he was involved in developing systems that the former head of Israel Security Agency (ISA - “Shin Bet”), Yuval Diskin, referred to when he said: “We need the sophisticated technology for databank collection and compilation, technology that can put special algorithms to use for data mining and sound warnings in real time on irregular behavior profiles. We’ve made enormous strides in these areas – even by global standards”.
The common link in the fields that Yuval Diskin mentioned is their connection to data mining and AI. In other words, they enable computers and robots to think in place of humans, and not only according to how they were programmed in the past.
Prof. Maimon, formerly of Harvard Univ., is considered one of world’s leading experts in these fields. He holds certificates of outstanding scientific achievement and originality in the area of the AI mathematics.
One of Professor Maimon’s ten scientific books is the 1500-page Handbook of Data Mining and Knowledge Discovery (Springer Publishers, 2005). Maimon began his career in operations research, and his doctorate dealt with robot control.
“Intelligence’s problem used to be how to obtain information”, Professor Maimon continues, referring to intelligence in general. “Today, almost everything has an electronic signature that can be detected. All you have to do is interpret its meaning. In effect, the amount of information is infinite. The hard part is to analyze the data and assemble a behavior pattern from the interesting parts, so you can know at an early a stage that an irregular event is about to happen.
“For example, it may have been possible to receive an early indication of the attack in Mumbai, India, based on suspicious signs, such as the terrorists’ unusual use of satellite phones, out of the ordinary money transfers, and so forth. Without knowing the important information, in order to gain an early warning, is like looking for a needle in the haystack”.
Professor Maimon, let’s clarify things. How does the process that data mining is a part of, actually take place?
“The whole process is made up of six stages. In the first stage, we collect data from sensors. This can be done in many ways: from a camera, HUMINT (interpersonal contact) or SIGINT (signals interception, generally from listening devices).
“The second stage is the ‘acquisition’. Its purpose is to convey the information to you. The third stage organizes the data (database). For example, if we’re collecting intelligence from the field, then we need to know which information came from which camera at what time, and so forth. If it’s a matter of listening, then we have to know the source and the time the conversations took place.
“The next stage is called ‘fusion’ – combining the data. All of the information collected from various sources should produce a single, rich, integrated picture, just as for humans the sense of smell, taste, and sight combine to give us the singular, rich experience of a select wine. The various sensors also have to be combined and integrated. When we link up the data coming from a UAV, ground camera, and ground-based listening device, we want to know if they’re tracking the same object or picking up different objects in the field.
“The fifth stage - the data mining stage - is designed to gain us insight. For example, if we have a print-out of a million phone conversations, we want to know if conversations number 1000 and 760,000 are interesting (and that all the rest can be disregarded). Or take a city with twenty thousand cameras collecting photos of license plates. OK, so the databank is filled with license plate numbers and routes, but we have to know what this means, if it interests us, besides the fact that thanks to the cameras we can check something retrospectively to a certain time in a certain place. Humans cannot cope with the mountains of information; we need computers to help us. The trick is not just to collect the data, but to separate the chaff from the grain, to arrive at the correct insights.
“Today you can walk into any computer shop and buy sensors and drivers with memories of several terabytes (one terabyte = one trillion bytes). The major intelligence organizations have hundreds of terabyte memories, and are on the way to obtaining petabyte capability– (1000 terabytes, a number followed by 15 zeroes). We need the insight to provide us with an efficient way to search through the material, and retrieve only the relevant information according to something known as ‘smart printouts’. Ninety percent of the information can be thrown out, and with the interesting elements that are left, we can make interconnections and crosschecks. At the end of the process, the trick is to say that something ‘interesting’ is going to happen.
“Only then, in the sixth stage, does the human enter the picture and make decisions based on the data. The final decision of how to act rests in human hands. If a helicopter has to fire a missile based on collected data, the human has to decide whether or not to press the trigger”.
Bottom line, is the intelligence in this process comes from the human or the machine?
“The intelligence comes from the machine. Humans build the algorithms that find the action pattern - the ‘pattern’. But a human cannot weed out the one interesting link from a million possibilities. This system was also used in the past, but somebody had to write the stages for deriving relevant information. Now it’s called AI. When you’re working with a terabyte of information you need a machine learning process to think for you, to glean the important information. If left to humans, they’ll miss it”.
Does the computer have to be programmed to discern what’s interesting, in order for it to warn of an irregular event?
“Sometimes yes and sometimes no. For example, when someone is at a border crossing, the computer can automatically define whether the person should be stopped for a security check, without having been previously programmed to identify this person. The goal is to design robots with the ability to ‘think’ independently, without programming them for every conceivable situation. In the future this ability may enable robots to be integrated into, say, driving vehicles on roads, driving not as humans but as super-sophisticated ‘fusion’ devices.
“These examples are only a fraction of the latent possibilities in ‘data fusion’. Today’s possibilities continue to enliven great interest in the world in both the military and civilian spheres”.
To what extent can we rely on ‘reasoning and discretion’ in computerized systems?
“Since everything is computerized, algorithms’ effectiveness can be measured and given numbers. Two numbers are given to every insight in the system: the first denotes the extent to which it is a specific matter; the second, to what extent interesting things in the process may have been overlooked.
“The user isn’t always told the degree of effectiveness that was determined based on these two numbers, so he/she won’t be confused, but we can decide that in a situation with over 90% accuracy, a target will be defined; between 70%-90% - that another check is necessary; and under 70% - the system gives no warning so as not to create an overload of false alarms. Let me make it clear: a human still needs to employ discretion in stage six, but I believe that in the future even discretion and judgment will be removed from us. In fact, some degree of human discretion has already been taken, but we don’t talk about it”.
“Since every algorithm uses much more than human data processing and thinking, it’s become a kind of smart filter. The machine often decides more than the human because there’s no other possibility. If the machine errs, then the human errs”.
There are currently Sentry Tech (Sees-Fires) Systems capable of firing automatically at potential targets in the field, for example, along a border. Can the machine make an error that would result in the loss of innocent people?
“This is why, in cases of firing systems, a human has to press the trigger, even if the person is located a distance from the sensors and firing mechanism. In the future, however, this too may change. You have to understand that in operating systems, based on data mining and AI, a learning- training process is needed. First the insights are created, and only later employed. Training is continual because the world is always changing”.
For many years now, American intelligence bodies have been operating a global listening network dubbed “ECHELON”. To what extent is Israel progressing, at the world level regarding insights that enable obtaining such systems?
“It’s hard to make comparisons between organizations and states in the area of intelligence, but in general, ECHELON is a collecting and sorting system. The question is what insights does the United States gain from it. In the area of insights, I would say that Israel is at a very advanced stage”.
Can someone evade even the most sophisticated intelligence system - Osama Bin Laden, for example, who the Americans have failed to capture to date? Maybe it’s impossible to track an individual ensconced in a cave, who doesn’t use a phone.
“The Americans seem to have missed a number of opportunities to pick up the trail of Bin Laden. Even if he doesn’t use a radio or phone, there are still people working for him and bringing him food and money. These people must be leaving some kind of electronic signature. Finding the relevant information on them is the greatest challenge facing intelligence bodies today”.
Photos: Meir Azulay