How Intelligent is Artificial Intelligence?

We’re standing here together in the Gewerbemuseum (Museum of Industry) in Winterthur in the midst of the “Hello, Robot.” exhibition. Will robots soon enable us to find more time to visit museums?

Pascal Kaufmann I hope so. Because there really would be more fascinating things than work for us humans to do, such as visiting a museum for instance. But it will still take a while for that to happen in hands-on professions, in all professions that involve a lot of fine-motor manual labor, such as nursing for example. It will happen faster in knowledge- and rule-based professions. Many tasks performed by the human brain today could be taken over by robots and algorithms.

Algorithms, artificial intelligence, machine learning, deep learning – can you two untangle the terminology confusion a bit?

Jeremy Callner I’ll start with machine learning. People often do a lot of fantasizing about this term along the lines of science fiction novels and Hollywood films. We could avert exaggerated expectations and even fears, though, if we called machine learning what it really is: automatic parameter calibration.

Kaufmann I think you need to explain that.

Callner As a physicist, I conceptualize the brain as receiving an input through sensory perception, experiences, etc., and effectuating an output that manifests itself by me moving, gripping something, etc. I fundamentally believe we can model that mathematically. A model has parameters. Machine learning means that for each of those parameters there is an algorithm, a kind of instruction manual on how the system processes new information – how it should sort it, for instance. The “learning” part of machine learning, if we even want to call it that, is limited to the algorithm that the programmer predefined. Deep learning, a special type of machine learning, requires somewhat less rigid algorithmic presets because it employs neural networks.

Kaufmann The image of an input-output machine is what instinctively comes to mind. It corresponds with what we know from the computers that shape our everyday technological lives. But this image doesn’t necessarily bring us closer to the understanding we seek. A few centuries ago, people believed that the brain was composed of tiny cogwheels and coil springs simply because clocks defined our everyday technological lives back then. In all likelihood, though, the brain actually isn’t an input- output machine because we humans constantly alter the input. I can’t grab a glass of water without continually manipulating the input in countless different ways. So, machine learning and thus also deep learning are purely statistical processes, in my opinion, and are not what I would consider intelligence.

Views undoubtedly diverge the most over the term “artificial intelligence.”

Callner One definition I like starts with machine learning as I described it. On top of that, there’s an added ability to simulate something and to thus formulate predictions and, lastly, to make a decision based on all that. I think that’s very similar to the way humans learn. Aside from being a physicist, I’m also a musician, so I know from firsthand experience how important it is to practice the right sequence of notes in a melody. If I repeatedly play it incorrectly, that forms memory traces in the brain. On the stage, under the spotlight, in front of an audience, I’m guaranteed to play the wrong note because it has become etched in my mind.

Kaufmann The theory of neural networks that you mentioned in the context of deep learning is arguably the best theory that we have at present to explain that. We neuroscientists say: “What fires together, wires together.” Neurons that fire at the same time in reaction to something form preferential connections with each other in the brain.

Callner And that’s exactly what we’re able to model. In “my” automatic parameter calibration, the parameters would be the connections between neurons. For example, there are methods by which we can teach a computer to recognize a cat. We work backwards from the image of a cat to individual pixels. The more images we input, the better we can calibrate the parameters and the stronger the connection is between the artificial neurons.

Kaufmann But how successful are we really at recreating the structure of the human brain in a computer? To me, the brain still possesses something magical. Neither better statistics nor faster computers enhance the quality of artificial intelligence, in my opinion. I also don’t place much stock in Big Data. Do I really need 300 million images of cats to be capable beyond a doubt of telling them apart from cows? Intelligence to me means to a far greater extent learning from Small Data. A toddler sees a cat once and knows what a cat is for the rest of his life. Don’t get me wrong, I’m unshakably confident that we’ll be able one day to create artificial intelligence on par with the human mind. After all, research groups around the world, like the Swiss Mindfire Foundation, are attempting to crack the brain code.

Why is that worth pursuing? Do I need to understand the brain in its entirety to create smart applications?

Kaufmann We have to specify more precisely what we’re talking about here. Even though automation and the quest for artificial intelligence are often conflated, there’s a difference between them. Humans have always sought automation. Back in ancient times, Archimedes’ screw made it easier for people to pump water. Today we automate by digitizing – thanks to machine learning, Big Data, etc. That occasionally has yielded some spectacular results. But we don’t have to crack the brain code to be able to automate. I’m trying to do it for other reasons. I want to understand, for instance, how humans plan, sometimes even beyond their own life spans. I don’t have to understand human biology to do that, by the way. Loosely along the lines of Leonardo da Vinci, I don’t want to replicate a bird, I want to build a flying machine. I’m interested in the principle of human intelligence.

Callner Which brings us back to the definition of artificial intelligence. I have a somewhat different view in the context of automation, perhaps also because I work for SIX. When we, for example, develop a service designed to automatically detect anomalies in market data, that definitely requires a certain form of intelligence. The result is a correspondingly large efficiency gain. JACOB, short for Jacob’s Automated Compliance Bot, is another example. In the past – before machine learning, Big Data, and today’s computing power existed – we would never have been capable of creating such a tool. JACOB helps compliance specialists at SIX, and will soon be helping our clients, the banks, to more efficiently keep track of regulatory changes in thousands of documents. So, to that extent, enthusiasm for artificial intelligence in a broad sense definitely has its justification.

Kaufmann I think we can agree that there is of course intelligence embedded in such a service – human intelligence. Starmind’s technology utilizes this as well to make the know-how of employees in large companies more accessible organization-wide. But at root, isn’t it the stored intelligence of the programmer that we mean whenever we talk about artificial intelligence? A service with a “life of its own,” particularly one from an operator of infrastructure for the financial market, arguably wouldn’t be tenable.

Callner I agree. And that’s also why terms like “machine learning” and “selflearning” are not just inaccurate, but even dangerous. As I said tonguein- cheek, that’s why I prefer to speak of automatic parameter calibration.

Kaufmann Speaking of autonomy, I would like to add something important. Once we have cracked the brain code and have replicated it in an algorithm, we won’t leave it purely to statistics and thus to its own devices. We will embed it in some kind of value system because genuine intelligence is always goal-oriented. And we set the goal. Artificial intelligence should play on the human team, not on the robots’ side.

Callner I’m all in favor of that. In the meantime, let’s allow machines to work for us as intelligently as currently possible.