I, AI (On the Unlikelyhood of Achievement of Human-Equivalent Cognition in Computational Neuroscientific Explorations)
April 18th, 2007 | Published in Professional, Personal | 1 Comment
I have to preface this post by saying that I happen to be reading Douglas Hofstadter’s latest book, I Am A Strange Loop
I believe that it will be extremely hard - well beyond the current estimates of futurists and Singularitarians like Ray Kurtzweil - to duplicate human cognition in silicon and code. While I was, and still am to some degree, wildly optimistic about the future of the computational study of cognition, reading Hofstadter’s new book has made me reflect more deeply on what has been called “Strong AI,” its possibility, its likelihood, and the form it would take.
I have become skeptical of predictions of human-equivalent artificial intelligence for three reasons, explained below. First, though, I have to say that I don’t think the limits will be strictly technological. In other words, I believe we will, within a couple of decades, have the raw processing power to emulate a sophisticated (if still somewhat simplified model) of a human mind, the sticking point being how to coax the hardware and software to give rise to such a thing.
I’m skeptical, first, because of the way in which human minds are bootstrapped with a huge number of concepts in their infancy, which due to the malleability present in the formative years enables them to categorize and manipulate symbols from times prior to the formation of real episodic memory. In other words, we are unfortunately doomed to forget exactly what made our a priori symbol set the symbol set we first adopted, and how we came across it. Researchers will have significant difficulty formulating and formalizing the initial structure of mental concepts, though I will take it for granted that once the framework is established, it will not be nearly so difficult to expand the symbolic “reportoire” of the artificial mind.
Two, the benefit of a “self” model. Many of Hofstadter’s writings to date have had substantial passages on what it means to be concious and to carry out cognition. Our “self” is not the focus of constant attention, but it is ever-present, and more importantly it is what shapes all of our perceptions. But, unlike a scientist at a microscope or an astronomer at a telescope, we cannot simply pull back from the instrument and take it in at a glance. Our reasoning about our self is always circumspect and incomplete, and so to think that we can imbue a machine with something so utterly essential and yet almost illusory - at least elusive - may prove beyond our abilities.
It is worth noting that I don’t expect current models of the tiniest structures of the brain (computational neural nets, for example) will scale far enough to be of aid. The approach will probably be top-down, modelling macrostructures and conceptual structures as they are deemed necessary. (But again, what of the emergent “self”? Can an otherwise “thinking” machine be compelled to do the work we wish it to do without a self with goals and drives to compel it?)
The third reason I somewhat doubt our ability to emulate human cognition relates to the richness of our environment. Imagine the stark contrast between the world we inhabit (where we take for granted trees, water, color, laughter, love, light, and concrete) and the world inside a non-thinking computer. Computers have only one built-in abstraction: binary symbol shunting. Ones and zeros. To elevate beyond this without the benefit of sensory input should strike us as much more than just inhuman. It may well prove impossible to inform a computer we wish to think about all the abstractions necessary to even begin the process of true human-like cognition. No one in the world can sit down at a computer terminal and act as the eyes and ears of a computer program, painstakingly breaking down the world into ever-simpler categories, descriptions, and relations. Without a significant revolution in our ability to shape code and silicon into the shape of a human, we will be dead in the water.
April 20th, 2007 at 12:53 am (#)
Many views of the brain portray it as an intrinsically complex system which resists reductionistic approaches. The failure of science at comprehending its workings allows broad, popular books like Hofstadter’s and Minsky’s to be successful without having solid facts or understanding to back up long-winded musings. Brain “science” is still in the domain of philosophers.
You don’t have to believe that it will be hard to duplicate human cognition — it *is* hard! But then again, what is this “cognition” you speak of? For that matter, how can we precisely define consciousness? Intelligence? We cannot even precisely define what we are aiming to achieve!
Our brain contains 100 billion neurons with 100 trillion connections, yet very little is known about how networks work. In fact, we currently only know about individual, isolated cells due to the limitations of recording technology. Hence, limitations in simulating the brain are not technological. Indeed, projects such as Blue Brain use supercomputers to attempt to simulate cortical columns using single-cell models. Yet, we do not know what portions of the models are important for “neural computation”, how information is encoded, or even what “neural computation” exactly entails. Few (if any) fundamental brain principles are known.
I disagree that top-down approaches will bear much fruit until models are falsifiable. To date, many models of neural function have been proposed (e.g. Self-Organizing Maps), but these models cannot be adequately tested using current technologies. Hence, theory has had few successes. The field of AI is inundated with top-down models of how we interpret the world, but none of them are even close at being as general like the brain. Hence, I believe it is unlikely that this approach will be successful (except for in very specific domains) without more understanding from the biologists reverse-engineering the real deal.
Your third argument against our ability to emulate cognition regarding the richness of our environment is not strong. We do not know what world we inhabit if not for our perception of that world. Perception is purely a product of the brain. We have millions of receptors throughout our body which only record *contrast* in the outside world. For instance, the mechanoreceptors on our skin only record changes in pressure rather than absolute pressure, and from that our perception records a rich, seemingly absolute tactile sense. And as for binary symbol shunting, isn’t it true that *any* discrete symbol can be encoded as ones and zeros (and indeed, almost anything can be considered discrete!)?
PS. The blog is great so far — your writing is always top-notch. Keep up the good work!