# Can AI research lead to new findings in general cognitive science?

Can AI be used as a tool to investigate our minds?

To be more precise, what am I specifically asking for here are examples of discoveries on artificial intelligence (so algorithms, programs and computers that try to implement intelligent systems) that brought to light facts about intelligence and cognition in general. Have this ever happened? Is it frequent? How influential and important were these discoveries, if any?

A possible example of what I mean could be the PSSH, which states that a formal system is sufficient to simulate general intelligent behaviour. I believe that this is relevant to Cognitive Science in general because it entails our understanding of this phenomena. (Of course, this is just an hypotesis, but I believe that its importance in the AI debate makes it a really compelling result).

• I think it's the other way round in most cases, by most I mean almost all. – DuttaA Mar 28 at 13:13
• Science as we know it has always been limited by our perspective, and as our perspectives have grown and changed so has our ability to understand our world. In this way I think what you ask is very likely. – JSON Mar 28 at 23:32

There are several vectors of scientific investigation and attempts to use the output of research to both comprehend the human mind and productize that knowledge for industrial, business, and consumer use, which run in parallel toward similar and largely overlapping goals.

• Artificial network research — convergence on behavior based on a criteria and a method, such as stochastic gradient descent with back propagation using a ReLU network cell function
• Reverse engineering the DNA mapping to forms of adaptation in the brain and its relationship to learning and what humans have come to consider intelligence
• Cognitive science via imaging and behavioral and cognitive experimentation
• Computer vision and auditory recognition in the context of robotics and automated vehicles
• Natural language comprehension research, usually related to semantic networks
• Natural language expression research, again, usually related to semantic networks

NPL without semantics is merely a trick but does not relate at all to cognition, although it is theorized that a sufficiently deep artificial network combined with the right heuristics and perhaps genetic algorithms can reach semantic processing via a natural route of emergence as it did in biological history on earth.

There are some past discoveries that lend much information to our comprehension of cognitive abilities in mammalian brains, especially humans, that are so integrated into our technological culture that we dismiss them as too mechanical to be related to artificial intelligence. These are phenomena we can verify and therefore qualify as the basis of fact, not conjecture.

• That arithmetic is a computer's job when it was once the sole domain of humans — We know from this that neural networks in the human brain, but not those of other mammals, can be trained to perform numerical computations and are born with the ability to accept such training, usually in the form of math classes, but they perform such inefficiently and require years of training to obtain the skill which a program can accomplish in a few lines of source code.
• That mail sorting is now normally done using automated address reading systems — From this we know that poor penmanship can still be deciphered by machines, indicating that reading handwriting is probabilistic, giving a hint about the whole of natural language.
• That the best artificial chess players regularly beat human champions — We know for a fact that excellence in board game play, by itself, cannot lead to adaptive survival. Put a person and a chess computer deep in a war zone or a dangerous jungle. The person will develop new abilities vastly outside the scope of board games and outlast the computer — Unless the computer is switched off, not particularly desirable to the soldiers or the lions, as the case may be, and is waterproof. There is something about biological intelligence that is integrated in with the body of the organism that makes it fundamentally different from machine intelligence, thus far.

We also know that normal human logic is approximately Boolean in that conjunctions "and" and "or" are like binary operators in computer science and unary operators like "not" are rough approximations of the natural language use of the word. It is also likely that Boolean algebra has honed the approximations of their natural language counterparts in industrialized countries and the emergence of programmers and philosophers have honed those approximations toward mathematical perfection, but these are not artifacts of human advancement from AI proper.

And it is not from the field of AI that we have learned the most important things about human cognition. There are several other related fields that provide a much larger set of data.

• Education

It is likely that AI will provide more information as these parallel tracks of investigation listed above continue to interact with one another. It is also likely that new hardware approaches to computing will bring to light more about how machines can think, which may or may not be related for any given specific case to how mammals think or how humans think. The development of cybernetic concepts into computers, mobile phones, and robo-investors have also shown us how humans don't seem to think very well, which segues into the other topic in the question.

General intelligence is a fable based on the idea that what humans consider intelligence is one dimensional. That college boards still differentiate mathematical achievement from linguistic achievement is just one artifact of the genetic truth. Those researching the relationship between genetic information and intelligence test results have found 22 independent genes with statistically significant correlations to smarts, and there are likely more to be found. There are other kinds of intelligence not yet investigated from a genetic standpoint that become evident upon examining these statements.

• The intelligence J.P. exhibits on the court shown through dramatically at last night's game.
• Sergei Rachmaninoff's brilliance is obvious to any advanced pianist or concerto lover.
• The business intelligence of Jeff Bezos has led him to the top of the heap of billionaires for two years in a row.

People exhibiting exceptional intelligence in various areas do not necessarily have high college board scores and may not pass a Mensa membership test. Although there is some correlation supporting the idea of a g-factor, the genetics simply indicate that a single dimension of intelligence is a highly oversimplified comprehension of what humans normally call intelligence in writings and conversation.

Furthermore, a simple application of inductive reasoning sheds significant doubt on the notion that general intelligence of an absolute nature is achievable in a human brain or a machine. Consider a thinker $$T$$ that can solve all problems in set $$P$$. There always exists a set $$P'$$ containing at least one problem for which thinker $$T$$ lacks sufficient resources to solve. If we develop $$T'$$ to solve all problems in set $$P'$$, then a set $$P''$$ may be found to contain problems of a complexity above the capabilities of $$T'$$.

Therefore, we can speak of more general or less general intelligence sensibly, but that only if we consider intelligence to be a vector, which it is. It cannot possibly be a scalar considering the full set of disruptive and compelling genetic and cultural observations that support multidimensional intelligence. We have not yet seen a proof that there can be a thinker $$T_{absolute}$$ that can solve a set of all possible problems $$P_{comprehensive}$$. In fact, Gödel's two incompleteness theorems may have already proven the opposite.

• wonderful answer! – olinarr Apr 1 at 21:05

If narrow AI research discovers new algorithm to solve a problem this can be used for improving General AI as well. For example, if an advanced A* pathplanning algorithm was invented to control a robot, the same algorithm can be put into a cognitive simulator to increase the realism of a virtual brain. The same thing is working in the other direction. If AI research doesn't produce notable results, the cognitive science has nothing to model and as a consequence the simulation accuracy of brain models is low.

An interesting question is, which kind of AI tools can be used for cognitive architectures. A recent examples are predictive models. They were invented for technical reasons to control a plant, but they can be identified in the human brain as well. Most humans have a detailed understanding about their environment and are able to estimate the result of an action. A typical example is, if somebody assumes that the bottle is full and tries to raise it up, but then he notices the bottle is empty and is surprised how lightweight it is.

This is about hard AI and soft AI: proponents of hard AI work on systems that simulate the way human cognition works, with the eventual (hypothetical) goal of replicating it. This presupposes that you know how cognition works, and presumably you will learn about it as you attempt to replicate it.

Soft AI, on the other hand, tries to emulate the outcomes only. For example, Weizenbaum's ELIZA is clearly on this side, as it uses simple pattern matching, and does not 'understand' anything about the conversations it is having.

Obviously, we don't even know fully what it means to 'understand' something, and building working systems is not really possible with a hard approach. Hence, soft AI is more common, as researchers are usually measured by their outcomes rather than their ideas. As far as I am aware, the hard AI approach has been all but abandoned long ago.

As current AI seems to be dominated by statistical approaches, I doubt that we can find out many useful things about cognition this way.

One interesting side-note: it seems to me that the capabilities of modern AI systems have developed away from human capabilities. A three-year-old can do some things that a sophisticated AI system cannot do, but in some areas (chess, translation, ...) the capabilities of AI systems surpass what humans are capable of. Maybe imitation is indeed not the right way to approach AI.

• Weak AI is doing more than only replicate the outcome of human cognition. A typical example for a Weak AI is an inverse kinematic solver which controls a robot arm. The sinus/cosinus equation isn't used by humans. It's completely artificial not available in nature. – Manuel Rodriguez Mar 29 at 11:00
• Yes, exactly: it is not replicating human cognition, but achieves a result that a human would also achieve, but by other means. You are basically agreeing with me! – Oliver Mason Mar 29 at 11:53
• hi @OliverMason great answer! So there's no chance in a Hard AI / AGI revival? I'd like to end up researching that in life, should I pick another focus? If that may help, I like both evolutionary computing and also Symbolic AI. Is there any chance I can follow this path? – olinarr Mar 29 at 12:07
• @NetHacker Yes... though it always leaves the question of how you can be sure you have replicated the way cognition works... TBH, I think there's a continuum between the two extreme hard and soft poles, so you can always slowly work your way towards the hard AI goals. – Oliver Mason Mar 29 at 14:02