CAPÍTULO III: ANÁLISIS DEL LUGAR
3.3 Análisis de transporte de la ciudad de Abancay:
3.3.3 Situación actual de los terminales terrestres de buses:
Interview first published May 2011
I have a lot of good things to say about Joscha Bach. He has a deep conceptual understanding of AGI and its grounding in cognitive science, philosophy of mind and neuroscience; a software system embodying a subset of his ideas and doing interesting and useful things; and a concrete plan for what needs to be learned to expand his system and his associated theory gradually toward human-level intelligence. What's more, I actually think his approach might work, if extended and grown intelligently from its current state. There aren’t many AGI researchers I can say all that about.
Joscha got a strong start in AGI via launching the MicroPsi project, a cognitive architecture based on an elaboration of Dietrich Dörner’s Psi model of motivated cognition. To test and illustrate the theory, Joscha used MicroPsi to control a population of simulated agents in a simple virtual world, as they interacted with each other and sought food and water and other resources in order to conduct their virtual lives. He also experimented with other AGI-relevant aspects of AI such as vision processing and robotics.
In 2005 Joscha left academia to co-found txtr, a company marketing e-reader software (and, briefly, e-reader hardware as well). But he continued to find time for AGI thinking, publishing his excellent book Principles of Synthetic Intelligence in 2009, summarizing his early work on MicroPsi and taking a few additional steps. While he found entrepreneurship an invigorating learning experience, and txtr is proceeding successfully, in 2011 he returned to academia, focusing back on AGI research full-time via a position at Humboldt University in
Berlin. Now he is building a new version of MicroPsi, aiming at both research and commercial applications.
I’ve found some of Joscha’s scientific ideas useful in my own AGI work, and the OpenCog AGI project that I co-direct has recently adopted some MicroPsi-based notions in its treatment of goals, actions and emotions. This indicates that we see eye to eye on many aspects of the AGI problem – and indeed, one amusing thing I found when editing this dialogue, was that I occasionally got confused about which parts were said by Joscha, and which parts were said by me! That very rarely happens to me when editing interviews or dialogues, and indicates that some aspects of our thinking on AGI are very closely aligned.
However, Joscha and I also disagree on many details regarding the best approach to AGI, which sometimes leads to animated discussions – for instance during his 2011 visit to the OpenCog game-AI project in Hong Kong. I’m afraid the dialogue between us presented below doesn’t quite capture the depth or the energy of our face-to-face debates. A series of words on a screen or printed page can’t quite capture the spirit of, say, our heated debate about the merits of multimodal story understanding as an AGI test domain, conducted while walking past the statue of Bruce Lee in Kowloon by the Hong Kong harbor. Also, many of our core points of dissension are too technical to be compactly discussed for a general audience (for instance, just to give the flavor of the sorts of issues involved: his preference for a more strictly “subsymbolic” network of concept nodes, versus OpenCog’s more flexible concept node network that mixes subsymbolic representations with more explicit rule type representations). But even so, I think this dialogue captures a few of the many interesting aspects of our AGI discussions.
Ben
You’ve said that your core motivation as an AGI researcher is to understand intelligence, to understand the mind. Can you explain
Joscha Bach: Understanding the Mind
why you think AGI is a good way to go about seeking this kind of understanding?
Joscha
Intelligence is probably not best understood as a collection of very particular capabilities and traits of test subjects (like in experimental psychology), nor as an emanation of a particular conglomerate of biological cells, their chemistry or their metabolic rates. Why not? Imagine you were to study the principles of flying: should you categorize birds, or examine their muscle cells? While both details might be helpful along the way, they are probably neither necessary nor sufficient for your goal. Instead, the most productive approaches have been the formulation of a theory of aerodynamics (the correct functional level for flight) along with the building of working models. This is exactly what AGI proposes: deriving, and simulating theories of
intelligent information processing, including perception,
motivation, memory, learning, imagination, language personhood and conscious reflection.
Ben
What jumps out at me here is that aerodynamics is a unified theory with a few core principles, and then a bunch of specialized subtheories derived from these principles plus various special assumptions. Whereas, in the case of AGI you give a sort of laundry list of areas for theories to address:
“perception, motivation, memory, learning, imagination,
language, personhood and conscious reflection...”.
The entries on this list seem to correspond better to subtheories of aerodynamics covering particular cases or aspects of fluid flow and flight, rather than to the core principles of aerodynamics. So this leads up to a question: do you think there will emerge some core set of mathematical/conceptual principles of intelligence, with the generality and power of, say, the Navier- Stokes equations that underlie fluid dynamics (and hence aerodynamics)? Or is the underlying theory of intelligence going to be more messy and laundry-list-like? If the former, what sorts
of ideas do you see as most likely to underlie this potential unifying underlying theory of intelligence?
Joscha
There are two main directions of approach towards AGI: One is constructive. We identify each sub-problem, perform a
conceptual analysis, derive tasks, specifications and
benchmarks, and get to engineering. Let us call the other one 'emergentist'. The idea is to specify a set of general dynamic information processing principles and constraints, and arrange them in a relatively simple top-level layout. After a period of sensory-motor learning, micro and macro structure of an artificial mind/brain are going to emerge.
There is some correspondence between these approaches and the divide between 'classical', symbol processing AI, and the currently fashionable 'embodied' AI, which eschews symbols, language, abstract decision making, planning and so on. If we understand symbol processing as logic oriented systems, then we are probably looking at the wrong paradigm. I have little hope for emergent embodied AI either. It is true that this is the way humans work: our genes define both a set of dynamic principles (for neural learning, generation of maps and clusters, reorganization according to new stimuli etc.) and the overall layout of our nervous system. But even if we manage to get the principles right in our models, and we recreate the major known features of the cortical architecture (cell types, neurotransmitters, columns, layers, areas, main pathways and so on) – and this is a big if we might not end up with something intelligent, in the sense that it is capable of learning a natural language or apply itself to an engineering problem. Nature has brought forth a huge number of these sensorimotor emergentist designs, and none, with the single exception of homo sapiens, is generally intelligent.
I suspect that we need a toolbox of the principles and constraints I just mentioned, so we can endow our system with cognitive information processing. This toolbox might be akin to the
Joscha Bach: Understanding the Mind
principles of aerodynamics, but I guess a little more complex. But just as the theory of aerodynamics does not yet imply the design of a particular helicopter, this alone is not enough. We will need to apply this toolbox to engineer the particular set of cognitive behaviors that is necessary for AGI.
By the way: “Cognitive AI” might be a very good label for what we want to do. It is not about Symbolic or Embodied AI (most embodied critters are not smart). The notion of Generality (as in 'Artificial General Intelligence') might be too difficult to define to be truly useful. (Does anybody remember the General Problem Solver?) But 'cognition' might hit the right set of problems: what is it that enables a system to think, to perceive, to learn? What representations are suitable for cognition? What motives are necessary to direct it? What kind of environmental interaction is involved? These are the questions we need to ask.
Ben
You’ve posed a lot of interesting ideas there, which we could dissect in detail. But one thing that jumps out at me is your perspective on embodiment. You seem to imply it’s not so central to the quest for understanding intelligence. Yet your own MicroPsi project did focus on embodiment of a sort – using your AI architecture to control virtual characters in a simple simulation world. Which seems to suggest you find some level of embodiment valuable.
Joscha
Some kind of embodiment is valuable for testing and teaching a cognitive architecture. However, I think a human-like AI could probably live in the body of a kraken, or in World of Warcraft, as long as the internal architecture – the motivational and representational mechanisms and the structure of cognitive processes are similar to the one of humans, and a sufficient (physical or virtual) environment is provided.
Ben
That’s an interesting direction to me. Care to share some more detailed thoughts on what constitutes a “sufficient” environment for human-like AGI? The relative paucity of stimuli in both World of Warcraft and the typical robot lab (as compared to everyday human environments) is on my mind here.
Joscha
You’re correct regarding robot labs. A lot of AI scientists defend their robot grants with the need to go beyond the paucity of simulations. And then they go ahead and put their robot in a black and white labyrinth, or in the simplistic green-white-orange environment of robotic soccer. Current robotic environments are not about complex social interaction or applied real-world problem solving. Instead, they are usually about limited locomotion, combined with second-rate image recognition. But I guess your question does not address the problems of using robots instead of simulations. So, what's necessary and sufficient, if robotics are neither?
Human intelligence is the answer to a complex and badly specified set of problems, such as: How can I find food, shelter, intellectual growth and social success? Tackling these problems over the course of one or two decades leads to a bunch of encodings in our nervous systems, along with a set of operations to manipulate these encodings, so we can infer, reflect, anticipate and plan. The encodings in our brain represent the order that our cognition managed to discern in the patterns thrown at it by the environment.
Epistemologically, it does not matter how we come by these encodings. An AI might get them by (A) proper perceptual and developmental learning, (B) by offline inference from a large corpus of environmental patterns, (C) by programming them manually, or (D) by copying them from another intelligent system with a similar architecture. I do not see how we can do (D) in the near future, or how (C) is tractable, so let us look at the other options.
Joscha Bach: Understanding the Mind
If you go for (A), you will need an environment that provides situatedness and supplies hierarchical, compositional objects, which the system can rearrange in interesting and useful ways. It should provide obstacles and ways to overcome them by applying problem solving capabilities. To make the obstacles count as such, they should block the way to needs (such as nutrition, shelter, social success, intellectual growth), and it probably helps if there are multiple problem dimensions, so that optimization is non-trivial. Free compositional thought seems to require the structuring provided by a natural language, so the environment needs language teachers and stuff to talk about. Also, we want a person-like and adult AI, the environment needs to provide social interaction and even a formal education. I do not think that any current robotic environment is even remotely sufficient for this.
Also, (B) might go a long way: statistical methods, applied on Youtube and the textual World Wide Web, might be able to infer much the same representations as humans have. We will still need to find the right representational framework, and the right operations and inference methods, but perhaps we won't need an environment at all.
That said, we should consider research trajectories. I believe that developmental learning in a complex environment could provide an excellent, competitive benchmark, and an exciting research task as well. If we move in that direction, we will see a continuous competition between (A), (B) and (C), and combinations thereof.
Ben
You mention perceptual learning, but the virtual world you used for MicroPsi was somewhat simplistic in terms of the perceptual challenges it offered. But I know you’ve also done some work on computer vision. Also, as you know, our friend Itamar Arel is one of many AGI researchers who believes modeling complex environments – unlike those in current simulated worlds or robot labs – is critical for human-like intelligence, perhaps even the
most critical thing. What do you think? How important is perceptual modeling to AGI, versus abstract conceptual thinking? Or does it even make sense to consider the two separately?
Joscha
I think there is probably no fundamental difference between percepts and concepts; rather, percepts are something like conceptual primitives.
Ben
Hmmm… But do you think everything in the mind is built up from percepts, then? Do you think there are other conceptual primitives besides percepts? Mathematical primitives, or primitive operations of mental self-reflection, for example?
Joscha
At the moment, I do not think that we would need other representational primitives. I guess that the initial building blocks of mathematics are inferred as generalizations from perceptual learning. I do not know the set of primitive operations for cognitive processing, though.
Ben
Fascinating topic indeed… But while it’s tempting to bounce my ideas about primitive cognitive operations off you, this interview probably isn’t the best context for that! So now I’d like to shift our chat in a more practical direction. As we’ve discussed before, while our philosophies of AGI are broadly compatible, I’m a bit more optimistic than you are. Assuming my OpenCog design is roughly adequate for human-level AGI (which, if so, I wouldn’t take as imply your design or others are inadequate… I think there may be many possible routes), I estimate there’s maybe 50-70 expert man-years of work between here and human-level AGI. I accept that could be wrong, but I think it’s the right order of magnitude.
Joscha Bach: Understanding the Mind
Joscha
I believe that the creation of an AGI exceeds the scope of 50 man-years by a wide margin, probably by more than one order of magnitude.
Ben
I'd love to hear more of the reasoning underlying your estimate...
Joscha
You are talking about something in the order of 10 million dollars. I have an idea how far a startup company, or a university project, can go for that much money, and your estimate strikes me as extremely optimistic. I remember when we were discussing that a budget like the one for the Avatar movie would really make a difference for AGI. 10 million – that's roughly twice the budget for Kurzweil's Singularity movie, or about as much as Sex and the City II spent on their wardrobe. Or, to transfer it to software development: 10 million dollars is an average budget for the development of a PC or PlayStation game. Even the environment to put the fledgling AI into is likely to cost more than that.
Ben
Certainly, a government agency or a large company (or a bureaucratic small company) could not create an AGI with 50 man-years of work, or with 10 million dollars. And a startup in Silicon Valley or a major US university has huge cost overheads. But a small group of the right dedicated people with the right ideas, in an inexpensive location, can do a lot with fairly minimal resources.
Joscha
Yes, that’s true, of course. However, in most cases where that happens, the small, dedicated team involved is solving a single hard problem. AGI is not concerned with a single enigmatic obstacle, but with at least several dozen problem fields. Each of these fields presents an infuriating number of puzzles.
Ben
Again this seems to come back to the “laundry list” versus “core unifying theory” dichotomy. What some AGI researchers would say is that, if you get the core unifying ideas of your AGI architecture right, then the puzzles in these various specialized areas will be dramatically simplified even if not altogether eliminated. I admit I have some sympathy for this perspective. I think the problem remains hard once you understand the core principles, but not quite as hard as you’re saying…
Joscha
To get back to my slightly crude airplane metaphor: A theory of aerodynamics will drastically cut costs and time in the design of an airplane. It is probably even a prerequisite. And yet, it won't solve the problems of constructing the airplane itself: what should the ailerons look like, or can we do without them? Where do we put the wheels, and how do we brake them? How do we distribute the payload, and which material is suitable for stiffening the wings, and so on.
A lot of the AGI problems are similar, in that they do not represent general principles of neural information processing. I think that humans share these general principles with other mammals, and yet, none of these others are generally intelligent. The architecture of the human brain is not just the result of general principles. It has additional constraints, so that particular lesions to the brain often yield specific defects. Some of these defects can be compensated for, which is an argument for general principles at work--but in general, this is not true.