Are you smarter than a four-year old?

RPI researchers have apparently created an avatar in Second Life driven by an AI equivalent to a four-year old human child. Sounds pretty interesting, and could be a good reason to check out Second Life, although I imagine it would be difficult to locate this particular character in the "game."

Currently, the team is grappling with computational tractability issues to do with the sorting of growing amounts of knowledge that is collected as a artificially intelligent character matures.

Massive parallelism perhaps?

A possible future application of this technology could be cost-effective solutions for pesky child-labor laws for TV game shows...

Outline of Conscious Modeling Framework

First post in a long time. I had trouble falling asleep last night, and this is the idea that was floating around in my mind:

1. Scan input for modeling (this input could be a sentence, image, or even perfect knowledge of a closed system provided by some form of perception).


2. Identify & label things (nouns) as "objects," actions (verbs) as "functions," etc.


3. Use (& dynamically update) dictionary/knowledge base to fill out this working model by expanding characteristics & features of the component items to some reasonable limit.


4. Cross-reference with Google-like search of memory using something akin to PageRank to generate connections, etc.


5. Update PageRank scores with new data collected from current modeling.


6. Commit new ideas into permanent model.

But can something as simple as this be robust enough to build a foundation of consciousness from?

Don't tell me I'm lost

Following up on my previous post regarding a flawed critique of AI published in MIT's Technology Review, I came across the blog of Larry Hardesty who makes some very good points revealing the poor logic behind the arguments of the original author.

Blind Impress - Uplift the bytecode!

Blind Impress - Gelernter Wrapup

Attack of the P-Zombies

Digg points us to a New York Times article about the increasing support for physicalism coming from our growing understanding of the material origin of thought. More and more research indicates that physical structures and processes in the brain are solely (no pun intended) responsible for the emergence of emotions, morality and consciousness, challenging the proponents of dualism who claim a distinction between mind and spirit.

A quote by Pope John Paul II in 1996 struck a chord with me; he said these materialistic ideas were "incompatible with the truth about man." Facts and evidence are incompatible with "the truth?" Sounds like something Galileo would hear...

We’re not lost, we just need a map...

MIT’s Technology Review has an article written by a Yale Computer Science Professor who is awfully pessimistic about the possibility of creating a conscious artificial intelligence. It seems to me, however, that his arguments against conscious AI are grounded in our inability to achieve success in the past 50 years and classical notions of subjective experience. But if the Wright Brothers had been deterred by previous failures they would have never built their flying machine, and if Magellan were bound to ancient concepts of a flat earth he would have never circumnavigated the globe.

Professor Gelernter frames the discussion around the differences between what he describes as "a ‘simulated conscious mind’ versus a ‘simulated unconscious intelligence’" and the mysterious "stuff" that makes us human. The professor has a special distaste for "conscious" AIs. He makes unsupported assertions that consciousness cannot be reproduced in software, simply because he believes this to be so. These arguments remind me of those of John Searle and others who contend that conscious AI is impossible because it is utterly inconceivable italic to them italic. The language of these philosophical arguments is often heavily-weighted with emotional triggers and sentimental definitions, rather than the cold, empirical judgment of hard science. Yet we can use the same philosophy to defend our science. I demonstrated the fallacy of many of these arguments (or at least attempted to do so) in my undergraduate honors thesis, "In Defense of Artificial Intelligence." Indeed, Gelernter points to Searle’s famous "Chinese Room" rebuttal to the Turing Test, which is addressed at length in my aforementioned thesis.

He also lumps emotions together with consciousness, and claims that since a machine cannot have emotions or a ‘sense of beauty,’ it therefore cannot have an ‘inner mental life’ or consciousness. Despite the fact that it has not been determined that emotions cannot be duplicated with software, it is not difficult whatsoever to envision a fully-conscious artificial intelligence completely devoid of emotional sensation. Indeed, sci-fi writers have had much success creating believable stories of conscious robots with no emotions at all. Although many AI researchers hope to recreate human intelligence, it is not the only conceivable model for intelligence. We may one day encounter a form of extraterrestrial intelligence completely unlike our own, but if it acts conscious, we will describe it as such.

The author ties the possibility of conscious AI to the performance of digital computers. He claims that consciousness is one of the many incomputable problems of the universe. This may be the case, but our present definition of computability may be as incomplete as our knowledge of that universe. Perhaps in the future we will build non-digital computers that are not bound by these limitations. This is another example of using current shortcomings to predict future failures.

Furthermore, how could an objective observer discern the difference between a ‘simulated conscious mind’ and a ‘simulated unconscious intelligence’? If consciousness is subjective, and one has no method of observing the subjective experience of another entity, then it is impossible to determine whether or not a given individual possesses a subjective conscious experience. The only method of observation available would be based on behavior, and if the behavior of a given system is indistinguishable from that of a conscious entity, then we have no other logical option but to ascribe consciousness to that system. We cannot know with certainty whether our fellow human beings possess the same type of consciousness we experience, but since we can observe that their behaviors are consistent with how we believe a conscious being would act, we do not hesitate to assign the attribute of consciousness. AI should be judged by the same standard.

The reason we’ve failed at using our biological brains as the model for AI development is precisely because we have very little understand of how our minds actually work. If you don’t know how something operates, its awfully difficult to recreate it. This does not, as Professor Gelernter suggests, imply an insurmountable technological hurdle, but merely a contemporary inability to implement our ideas. History is filled with examples of insufficient knowledge causing temporary roadblocks to progress, which are later overcome via persistent, hard work and breakthroughs.

Again, past failures are not good predictors of future results and traditional wisdom is often wrong.

AI News for the Masses...Down Under Style

n extensive article about robots, androids, and artificial intelligence was featured in the Sydney Morning Herald in Australia. Although designed for a layman, it still provides a decent perspective about the current state and future of AI.

Hubris Spells Doom for Grandmaster

Sophisticated AI chess programs are continuing to beat the world’s greatest human players, and the humiliated Grandmasters are continuing to gripe & complain. First, Deep Blue defeated world champion Gary Kasparov in 1997, and just five months ago the current champ was beaten in a one-move checkmate by a new AI program.

"I rechecked this variation many times and analyzed quite far ahead," Kramnik protested. "It seemed to me I was winning."

Indeed. Chess had been long regarded as a game requiring much creativity & ingenuity to excel, something that machines would never dominate. But as Deep Blue and others have shown us, our unabashed confidence was no match for the brute force mapping of all possible outcomes. Eventually, human players adapted their game play to try to throw off the computer programs, but the AI have adapted as well, going so far as to invention previously unknown variations, to maintain their chess supremacy.

How long until other domains previously described as bastions of creativity & consciousness give way to algorithms & AI? Many have spoken about the “God of the Gaps” concept, but what about a Consciousness of the Gaps?

Dartmouth’s New Approach to AGI

Dartmouth College’s new Institute for Computational Science seems like a very promising approach to creating Artificial General Intelligence, combining elements of engineering, philosophy, neuroscience, and traditional computer science.

While the student paper article seems little more than a press release regarding the new Institute, the author raises and interesting issue about algorithms used by computational scientists to “re-create” elements of the real world for research & experimentation. Could sufficiently-complex modeling & simulation systems serve as the mechanism behind internal mental modeling for a conscious artificial mind? We regularly use similar tools for testing aerospace & fluidic systems, as well as modeling astronomic & quantum phenomena, so why not a generalized apparatus for predictive conscious thought? Such a system could provide the means for planning, intentionality, exploring cause & effect, as well as guessing what one will find around the next corner. Of course, expanding the domain of our modeling & simulation algorithms to be thoroughly general in nature is no trivial task, but it does open an interesting path to explore.

Using Guesses to Build Mental Maps

Purdue Univ researchers have constructed a robotic navigation system that utilizes predictive mapping to chart the environment of unknown areas. The robots create “mental maps” of unseen regions based on observed patterns recognized in previously-charted territory. Simulations have shown successful navigation while exploring as little as 33% of an environment.

This system seems well-suited for Grand Challenge driverless vehicle applications, which in the past have mostly relied on interpretation of real-time sensor telemetry. If this predictive mental mapping could be expanded beyond environmental navigation to contextual awareness in general, building abstract mental models of ideas and concepts, it could be another step on the path towards artificial consciousnesss.

Emotion a Prerequisite for Thought?

A group of scholars gathering at Harvard to celebrate the 50th anniversary of the “cognitive revolution,” are promoting the idea that understanding human emotion is absolutely essential to understanding human thought. The scholars go on to argue that the reason all attempts to recreate human thought in AI have failed up to this point is because AI researchers have ignored the role played by emotion in intelligence. While it is true that emotion is central to human thought, it’s awfully anthropocentric to claim the human model as the only path to intelligence.

Although minds and computers may not be completely analogous, to say that emotion is this missing link to artificial general intelligence seems rather naïve and simplistic. There is much work to be done before we can even come close to developing strong AI, and it is quite easy to imagine thinking machines completely devoid of emotion. Intentionality may perhaps be integral to any intelligent system, and emotion may be deeply-seated in human intentionality, but any number of goal-oriented schemes could be devised to replace emotion as the driving force behind an intelligent machine.

An Artificial Conscience for Artificial Warriors

The Economist has more on the U.S. military's attempt to build an ethical decision-making framework for autonomous weapon systems. This particular initiative seeks to map out all possible actions & outcomes in order to select the most appropriate or "ethical" behavior. It sounds a bit like Deep Blue's brute force calculation of all possible moves & consequences, however, the decision space of a battlefield is slightly more complex than a chessboard...to say the least.

Dumbo has become self-aware

As scientists ascribe more & more conscious traits to animals, when will machines receive similar considerations?

Bronx Zoo researchers recently announced they observed that elephants can recognize themselves in a mirror. Thanks to this evidence, we can now attribute a level of self-awareness formerly reserved for dolphins and primates (including humans). Other behaviors, including the use of tools, communication, social interaction and even apparently emotional responses lead experts to credit these animals with a relatively high order of consciousness.

But if machines were to exhibit the same behaviors, would people be willing to assign the same characteristics to? The likely answer is no. Consciousness and self-awareness are highly emotive topics that play a central role in the identity of the human race. The resistance to extend such labels to artificial intelligence is akin to the resistance to accept the implications of macroevolution with respect to the origin of our species. People often resort to a sort-of “mystery of the gaps” argument, claiming that since we can fully understand the internal processes of a “white box” artificial intelligence, it cannot possibly contain the same “stuff” as human minds. But as we increase our understanding of human consciousness, eliminating the gaps and turning the “black box” gray, the mystery disappears. This entire line of reasoning strikes me as being analogous to the argument which contends that we cannot possibly be descended from apes, because human beings possess a soul while chimpanzees do not.

Perhaps one day our society will accept that our own consciousness is merely the product of highly complex systems and processes of the physical mind, and therefore grant the same definitions and descriptions to artificial entities that mimic the same processes and behaviors.

Darwinian Survival Machines or Anatman?

USNews.com has a lengthy article exploring the soul’s role in human consciousness which has some interesting implications for machine consciousness as well.

The article spans a variety of subjects from psychology to philosophy, and the most interesting passage is in the midst of a discussion about the neurophysical processes behind consciousness and language with philosopher Daniel Dennett:

"This is what I've meant over the years when I've said that the brain is a syntactic engine mimicking a semantic engine." By that, Dennett presumably means that consciousness produces orderly, grammatical representations of something out there in the world that is meaningful, but it does not create meaning. It is not necessary to meaning.

This argument rejects the notion that a Turing machine is not actually conscious since it merely mimics the understanding of language by suggesting that this mechanical procedure is the only system at work within the human mind. By reducing the physical processes in the brain down to their fundamental components, we see that they are no more remarkable than the internal workings of a digital computer. Critics of AI argue that machines & algorithms cannot be conscious since we can observe precisely how they reproduce human behavior. Perhaps once human behavior itself is de-mystified to the point where our consciousness is understood completely, these people will be willing to apply the label of conscious thought more broadly.

via KurzweilAI.net

More Details about the UAV COS

Larry pointed out some good info to me about the UAV Common Operating System on the DARPA J-UCAS website.

There is a COS FAQ (which unfortunately seems to be more focused on answering Boeing & Northrop's questions about protecting intellectual property than explaining what COS is!) and a Common Systems & Technology presentation which includes a very interesting section on the COS.

Some items of note from the 15.2MB PPT slideshow:

- The COS is more of a POSIX-compliant middleware than an 'operating system,' as it resides on top of an embedded OS.

- Integrates sensors, weapons, communications & control systems to generate decisions and provide a real-time understanding of the operating environment

- Allows autonomous operations, eliminating need for human operators

- Enables accelerated software development... 'best of breed' algorithms shared between platforms (survival of the fittest?)

- Automatic determination of flight path based on mission parameters and threat environment

- Open architecture, non-proprietary source code can be freely used, copied, modified and redistributed

- Initial development timeframe begins Dec 05 and ends Dec 08, with 4 builds planned using a spiral approach

The most significant COS component, in my opinion, would have to be the Discovery/Mediation functionality...

The discovery component provides a set of services that enables the formulation and execution of search activities to locate data assets (files, databases, directories, web pages, streams) by exploitation metadata descriptions stored in and or generated by Information Technology (IT) repositories (directories, registries, catalogs, etc). This also includes the ability to search for metadata semantics to support mediation.

This mediation component provides a set of services that enable transformation processing (adaptation, aggregation, transformation, and orchestration), situational awareness support (correlation and fusion), negotiation (brokering, trading, and auctioning services), and publishing.

What this sounds like to me is a method for storage and retrieval of shared memory resources (by way of metadata [xml?] and semantic analysis), and using those memories to enable adaptation to dynamic situations. That bears a striking resemblance to the way a conscious being uses personal and communal experiences to deal with real-world environments and react to change in real-time. The more I think about it, the more I see the possibility for something truly intelligent to emerge from this architecture. I'm definitely going to have to keep tabs on this project, and if it ends up being open source (as the presentation claims), taking a look under the hood. Now do you understand why I want to work for these people?

Inside the Mind of a UAV

DARPA is currently developing a common operating system that will serve as the 'brains' for the military's next generation of autonomous systems.

This common platform will allow UAVs to share information and work together more efficiently, and port new advances from platform to platform. Additionally, it will merge inputs from all of the sensor systems to create a single integrated picture of the environment, and interface with the on-board weapons array as well. Could this internal model of the outside world eventually be the seat of consciousness for an intelligent machine? DARPA's own report is not surprisingly vague...

The Mechanism of Human Consciousness

KurzweilAI.net picked up an article from The Economist about a possible breakthrough in the search for the neural correlate of consciousness.

In his latest (and final) published paper, Francis Crick (the late co-discoverer of DNA) speculates that the neurological basis of human consciousness is the claustrum, a thin layer of gray matter with two-way connections to every region of the cortex. It is possible that the claustrum is responsible for processing all of the various inputs from different parts of the brain to create a "cohesive, conscious experience." If research demonstrates this conclusion to be true, it will unequivocally prove the Materialist view that consciousness is merely a product of a physical system, and therefore can be artificially reproduced by mechanical means.

A Formal Framework for Conceptual-Level Cognition Based on Objectivist Epistemology

I've been doing a lot of research into Objectivist Philosophy lately, and its given me a lot to think about. I've studied philosophy pretty intensely over the past 5 years (including a year at Oxford), and I don't understand how this could have escaped my attention for so long. I think I may have finally found something that encompasses most of the conclusions I've arrived at regarding reality and the existence. If you want to understand my personal view of the universe, Objectivism is unquestionably the perfect place to start. This website, however, is not a 'philosophy blog,' so I will quickly move along to my point.

One component of this philosophy, namely Objectivist Epistemology (the nature of knowledge), has led me to an idea for an internal conceptual model of a conscious machine. This system of knowledge representation would closely mirror the Objectivist structure of concepts and instances. Simply put, Objectivism contends that the senses are the only source of valid information, that 'reasoning' is nothing more than the classification & organization of the information provided by sensory perception, and that consciousness is the act of transforming that information into a system of concepts and propositions consisting of knowledge about the world. A conscious mind creates concepts by identifying and generalizing similar characteristics of things it encounters through the senses. Once a concept has been formed, an unlimited number of instances of that concept can be recognized by the agent.

This framework immediately reminded me of object-oriented programming, with concepts acting as classes and instances as objects. Just as a programmer can construct an unlimited number of objects once its class has been defined, a conscious mind can recognize endless instances of 'airplane' once the concept of 'airplane' has been properly formulated. The key is to identify the relevant characteristics that define the concept (man-made machine, has wings, enables flight), and to recognize those characteristics in similar phenomena.

We could implement this framework in a machine using these steps:

1. Define the machine's sensory perception. How will it observe the world? Will it have eyes? ears? tactile sensors? The limits of its perception will directly affect the scope of the machine's conceptualization.

2. Construct the mechanisms for input/output. Connect the cameras and microphones that serve as eyes and ears, and provide the machine with some method to interact with or otherwise manipulate its environment.

3. Formulate the aforementioned conceptual framework. Write the software that controls the operation of these devices, analyses the data they collect, and classifies & organizes the information accordingly.

4. Release into the environment. Allow the machine to observe and interact with its surroundings, building concepts and collecting knowledge about whatever it can discover.

For example, a machine designed with these capabilities might one day encounter what we would recognize as an 'airplane.' First, the machine would compare this object to its memory files to see if it had encountered something similar before. Failing that, it would make note of the distinguishing characteristics of the object and assign it an arbitrary identifier. Later on, when confronted with another instance of an 'airplane,' the machine would recognize distinct similarities between this object and the unidentified one in its memory, and would construct a concept of 'airplane' based on the obvious relationships between them.

This framework doesn't solve some of the major challenges of artificial intelligence like pattern recognition, machine vision or natural language understanding, but it does provide us with a system to integrate these systems once they are accomplished. And it wouldn't do much good starting from scratch. This agent would require some background information, a 'seed of knowledge' if you will, in order to get up and running. Even if this included things as (relatively) simple as language skills or a basic means of communication, designers will need to provide the machine with a small amount of starting information in order to 'prime the pump.' Humans are born with many 'instincts;' artificial minds like these must be programmed with them. Additionally, we will need to sort out the details surrounding memory access and storage requirements for this agent to function efficiently. An intelligent system which maintains a full audiovisual record of its experiences would likely be impractical, so perhaps this is where some contextually-relevant summary software would be useful.

Is It Even Possible?

I was recently introduced to an intriguing speech on the plausability of conscious machines given to a club of Objectivists a few years ago.

I appreciate many of the author's positions, but am not sure what his ultimate message is. The simple fact is that we are trying to recreate consciousness when we don't even know what consciousness is! Some attribute it to the soul, which doesn't address the issue and is simply an appeal to superstition. Others argue that thought is nothing more than calculation, but that fails to account for comprehension and understanding. Even though we are unsure of the direction we are going, we continue in our endeavor to build a thinking machine. In the end, perhaps our greatest success will not be the accomplishment of that goal, but in fact that we discovered the true nature of our own consciousness along the way.

From Deep Blue to Deep Thoughts

This week I began exchanging emails with an author who is writing a series of novels based on humanity's march towards the future.

One of the major events in the story is the emergence of a conscious AI. This process is presented from the point of view of the newly conscious entity, an approach that I find to be a wonderfully intriguing and novel concept. The process of becoming self-aware from the first-person perspective is certainly uncharted waters... even we humans cannot recall how or when we first became conscious as small infants. I have often tried to imagine what subjective experience would be like for a non-human intellect, both lower orders of intelligence (animal) and higher, more developed orders. I'm sure it must be very challenging to try to capture and describe that experience using human language and concepts. The small sample included in his most recent email has made me even more eager to read the story when it is completed.

He suggests that the key to subjective experience is likely the realization that the individual is separate and distinct from the rest of the environment, and I would be inclined to agree. He referenced a pdf slideshow by Owen Holland titled Machine Consciousness and Creativity. The central theme is that it is not the 'self,' but rather the 'internal model' of the self that is the conscious entity. If this is truly the case, it would fit well with the evolutionary development of consciousness (in that the capacity to accurately model future events would provide a strategic advantage in the 'survivial of the fittest' arms race), as well as cause a shift or refocus in AI research. It would seem that a conscious mind is more of a 'universal prediction engine' as opposed to a 'universal reaction engine.' So we might say that it would be impossible for a conscious entity to know what to do in every situation (or to program a machine for every possible circumstance), so instead we have developed the ability to predict possible outcomes and act accordingly. Interestingly, it is the ability to predict outcomes and formulate courses of action which is something IBM's Deep Blue has shown a remarkable aptitude for. Not that I am suggesting that a chess program is conscious, but if a machine was developed that was able to similarly model a much broader sample of reality, we might be much closer to accomplishing that goal.