Spoofing AI

This an interesting new scientific meme. It made it into Science on the basis of a presentation at the International Conference on Machine Learning here. The idea is that hackers can easily defeat AI’s (think “social engineering” used on a machine).

Meanwhile there is the contrasting meme of us getting spoofed by AI, in the FT, here. In this case AI’s are able to make videos of people doing things that they did not do.

All of this gets to the cybersecurity aspects of AI that potentially put society at risk.


AI and Video Gaming

Today’s FT has an interesting article(behind paywall) about AI being deployed into the video game space after its success at Chess and Go. What interests me here is that such video games are more open ended and ‘noisy’. They typically don’t have compact rule sets and strike me as capturing more of the flavor that smart machines are going to encounter in the real world (say when they are autonomously driving on the Washington DC beltway). Of course, the typical algorithm right now involves reinforcement learning and the AI plays against itself. That’s perfectly sensible in a gaming environment, but not really applicable to autonomous robot roaming out in the wild.

There’s a different approach out there and it’s based on reverse-engineering the brain processes that sub-serve hu man child language acquisition.  The key idea is that human children acquire language with great ease and not a lot of reinforcement. We know quite a bit about the neurobiology of mnemonic function, both at the molecular level and at the neuro-algorithmic level. That this existence proof manifests so saliently suggests to me that this is where the next paradigm is going to be revealed.

Rules of Life: SBE Version

Many readers are aware of NSF’s 10 Big Ideas. One of them, Rules of Life: Predicting Phenotype originated in the Biological Sciences Directorate while I headed it up. We also used a similar set of words to frame all of the Directorate’s investments—from scale of an individual ion channel up to that of an ecosystem: Understanding the Rules of Life (URL). The intellectual idea here was that simple rule sets can, on the one hand, constrain nature and yet on the other produce vast complexity. An example of a very simple such rule is the Pauli Exclusion Principle from Chemistry. Pauli constrains atomic configurations by requiring electrons occupying the same orbital to have opposite spins. That simple rule produces the Period Chart of the Elements and by extension carbon chemistry (i.e. organic chemistry, the backbone of living things).


Biology itself has many such examples. Evolution itself consists of a rule involving history and contingency. Neuronal synapses (the connections between nerve cells) in the brain are constrained by the tree-like morphology of neurons: if branches of adjacent neurons aren’t close enough, then there is no possibility for the formation of a new synapse. The DNA dogma itself is a compact rule set that leads from base pairing through the genetic code to the construction of polypeptides that we call proteins.


The NSF has another Directorate for Behavioral, Social and Economic Sciences (SBE). It deals with all things human, particularly the emergent properties of human beings interacting with one another in constructs such as cities or, in a more abstract example, markets. Wars, mass migrations, stock market crashes and the World Cup are the types of emergent properties that are referred to here. They are concrete, consequential and produced as a result of many individual human agents behaving together in the biosphere. The current climate disruption on the Earth is thought by many of my colleagues to be anthropogenic in nature, an emergent of human development since the Industrial Revolution.


Not surprisingly, SBE was (and presumable is) enthusiastic about the Rules of Life Big Idea at NSF. After all human beings are living things, embedded and integral to the biosphere. If you are investing in social, behavioral and economic sciences, then by definition, you must be curious about the rules that govern these disciplines. And I think such an outlook can only strengthen the social sciences (writ large). Rules of Life as a framework can help create a theoretical scaffolding for the SBE fields in the same way that quantum mechanics does for physics and chemistry. Scientists seek to do more than collect and describe. Above all, they seek to predict and generalize.


A larger question though is, what are the rules that govern the production of human societal emergent properties? Is it possible that we could write them down in a compact fashion as we can for the game of Chess?


As I look out over the global political landscape these days, with the populist electoral success extending from the Philippines to Brexit Britain…and certainly including my own country…. I am curious whether there is a hidden rule set that relates these movements to a certain societal incivility that seems to be spreading as a social contagion. Another phenomenon that seems to be recently emergent is an increasing acceptance of lying on the part of political leaders. Instead of being viewed as shameful, such actions seem to viewed by many as reflecting strength and genuineness. Is there a human societal rule set that governs the acceptance of deception?


I had lunch yesterday with a colleague from our economics department yesterday and we both wondered whether the decline of organized religion had something to do with the recent political landscape, however humans have been in such dark places before in times when organized religion was very strong. In any case, a lunchtime conversation is not the way to elucidate a rule set for human societal behavior.


What would be the way to reveal such rule sets? One notion is to use agent-based modeling. In this approach, human beings are modeled in silico as software agents. The agents interact according to rule sets created by the experimentalist (a computational social scientist) in a massive manner, limited only by Moore’s Law. The emergent behaviors of the whole system are what is measured and the idea is to understand the relationship between the designed social rule set for the agents and the resultant emergent behavior of the model. The problem with this approach is that humans are very complex—much more complex that the modular pieces of software that comprise agents.


Another approach is to use college students as experimental subjects in behavioral economic experiments. This was the invention of another former colleague, also an economist, who won the Nobel Prize as a result of this idea. In such experiments, human subjects are paid real money as they interact with each other or computers under designed rule sets, similar to those used in agent-based modeling. The famous Prisoner’s Dilemma is an example of such a designed rule set. Here, the experimental results are quantifiable (how much money each student has at the end of each experiment) and the agents are real human beings (albeit a bit young). A neuroscientific bonus to this type of research is that the human subjects can be brain scanned as they interact revealing the neural substrates for their actions. The problem with this approach is that the number of experimental subjects is orders of magnitude less than the number of human agents interacting in real social phenomena such as stock markets. Hence, in general, such behavioral economic experiments are statistically under-powered relative to the social behavior they try to explain.


I think it’s time for by SBE friends to invent a new rule discovery approach. The timing is ripe: the relevance of such rule sets to our survival on the planet is clear. With the advent of ubiquitous AI, such rule sets will be of crucial importance to the engineering of ethical, legal and social frameworks for robots and the like as they interact with human beings. And it would be interesting to discover how human history relates to our social natures, not just in a qualified way, but one that is predictive and generalizable.

What is Strategic Surprise and how can AI help us?

We have been considering the notion of strategic surprisein the context of a report that we are preparing for the USAF. The AI part of this so far had been that US might suddenly face the reality of something like general AIbeing used against us at some point. But it doesn’t have to be like that. It could be that AI could help us delineate strategic surprises in other non-AI domains that human analysts haven’t anticipated (e.g. commercial airliners being used as very smart cruise missiles during 9/11). This intrigues me. Imagine if an AI had guided the original investigators who discoveredan adaptive bacterial immune system (CRISPR) that it could be used as a gene editor way back in 1987. So could AI act as a discovery accelerator?