Abdessamad: "chinese room argument" has many problems. For example, It's based on an "axiom" that a book does not have consciousness. However, this "axiom" is not always right when a book is big enough to answer all the question in natural language. Researchers of AGI do not need to be limited by existing philosophy conclusions. Actually, I believe with the development of AGI, some of the philosophy conclusions will be broken.

Personally I do. I feel that there's a huge need for much more generalized artificial intelligence systems, but there's a kind of dichotomy here. The general way I find these going is that researchers will choose a very specific problem area to focus on (Named Entity Recognition for instance) and develop very advanced approaches for that specifically. However, with the amount of generality that is inherent in statistical machine learning models, the findings from specific areas are often generalizable to other areas.

There are currently a number of companies working for more general artificial intelligence and data science solutions. One such company is Data Robot, they have a large suite of general data science algorithms that they run on an arbitrary data set. They haven't gotten very far, but they've raised a few million USD to try to solve this problem.

That said, academia has been very slow to take up the helm. Industry has been getting further and further ahead of academia in this respect, and while academia informs most if not all of the algorithms that go into these algorithms, the papers that are prevalent in the academic space generally have a very narrow focus.

Right Dr Victoroff and thanks for joining the discussion. Indeed, just 20 papers were accepted for the normal session of the conference (http://www.agi-conference.org/2013/accepted-papers/) and titles give an idea supporting your critique about academic focus in the field of AGI.

I do and share also the view with S. Victoroff of academic obsolescence regarding this perhaps the most epochal enterprise in the history of mankind. I have a conviction related to perception and abstraction patterned in the operational complexity of district conceptual mathematics that we are presently really nearing faster than generally may be expected to initial breakthroughs in AGI. The rest depends of the general public acceptance rate of adjusting itself to and taking benefits of this transition in a sense to present human abilities passing systems.

I’ve just noticed and I acknowledge a former thread in RG dealing with “the future of artificial intelligence”.

This thread is about DIRECT paths to Artificial General Intelligence : “When explaining Artificial General Intelligence to the uninitiated one often hears the remark that, surely, everyone in AI is working to achieve general intelligence. This indicates how deeply misunderstood intelligence is. While it is true that eventually conventional (domain-specific) research efforts will converge with those of AGI, without deliberate guidance this is likely to be a long, inefficient process.”

The Direct Path to Artificial General Intelligence, Peter Voss in Artificial General Intelligence, Springer 2007 “The first book on current research on artificial general intelligence (AGI), work explicitly focused on engineering general intelligence” (http://www.springer.com/computer/ai/book/978-3-540-23733-4)

The original goal of AI was achieving a sufficiently advanced understanding of intelligence so that engineered systems could be built that harnessed human-level intelligence. However, the initial optimism faded quickly when slow progress was made. Fast forward to today, and we have Narrow AI. These are almost without exception highly specialized systems that work only for specific, reasonably well defined tasks and can not tolerate substantial changes in tasks or operating environments. The lack of progress when it comes to the original goal is not really surprising, partly because computer science has largely abstracted away time and energy, which are quite central to human-level intelligence.

So when you ask if we share the call for this transition, i.e. that we should refocus on the original goal vs. continuing to work on Narrow AI, that depends. Do we want things such as, for example, fully autonomous household robots, rescue worker robots, space exploration robots? If the answer to this question is yes, we must pursue AGI as I firmly believe (and can easily argue for) that narrow AI is fundamentally not capable of delivering on these kinds of systems.

As for me personally, the answer is clearly yes - I have a paper at this conference this year and had one last year as well. The conference usually has about twice as many papers accepted, we can speculate that the location of the conference is a considerable journey for many AI researchers and also that the last conference was not too long ago, in December last year.

I would also be careful do dismiss all AGI work for academic focus. There are many researchers working in this space with different research goals, and this does not hold for all of them.

Artificial General Intelligence is not that far from us as many researchers believed. I am working on a general NLP system. Actually, the NLP semantic model I proposed two years ago is general knowledge focused. And I agree to Dr. Victoroff that the problem for academic researchers is the research always be limited in a specific problem. However, to process the general knowledge, the different problems such as dependency parsing, co-reference resolution, name entity recognition and most important semantic analysis depend on each others. Each one of these problems can not be solved separately. An integrated solution is necessary for AGI.

I agree with many of Shanfeng's points, in particular the last one regarding integration. However, if one truly wants to address the complexities of human intelligence, one cannot ignore time and resource constraints. Modern computer science has made this very challenging for AI researchers: The Turing machine completely ignores time, resources and parallel processing. So the problem is not just developing all the necessary cognitive functions in a general fashion, it is also making them work together in a way that meets requirements for real-time processing under conditions of multiple simultaneously active goals, time constraints and resource limitations.

For an in-depth look at these topics, the curious and interested could do worse than check out my recently published dissertation ("General Attention Mechanism for Artificial Intelligence Systems").

I believe the complexity problem can by solved by pruning. Semantic space of natural language is unsymmetrical. But most the conclusions of complexity theory are based on symmetrical data space. Complexity theory can not be the guide of the research of AGI.

Shanfeng, I am sure you are doing magnificent work on NLP, but this is just one part of intelligence. A robot that is extraordinarily well equipped with NLP does not make for a robot that is capable of making a cup of coffee in a random household, rescuing survivors from a random burning building or exploring the hidden oceans of Titan (the most likely location of extraterrestrial life in our solar system). There is not a single reference to complexity theory in my dissertation, I recommend you examine it and get back to the main thread of this discussion.

I believe that both fields need to progress interacting with each other, but at the same time seek their independent goals. As long as narrow AI directs its focus on small incremental developments in narrow fields it will only progress towards optimizing algorithms on narrow subjects instead of general problem solving. Nonetheless, there is still value in highly specialized algorithms that carry out precise discriminative tasks. We can focus on general artificial intelligence to make systems such as those described by Helgi and also have specific instances where they invoke narrow AI for precision discriminative work.

Arturo, sure, what we are talking about is how to practice AGI. Narrow AI has their own value for specific applications. But in NLP area, without general knowledge, even the domain specific problems can not be practically solved.

Heigi, I believe the knowledge representation model should be integrated. It's not only for NLP but also for graph recognition, robot control ... etc. We need to find a way to comparatively process the parallel information from video and natural language input.

I first understood Dr Hu arguing the necessity of integration as path to AGI, I mean not only Semantic integration but a kind of Cognitive integration ... am I wrong?

Arturo is actually describing a space where I think narrow AI is highly relevant to AGI: Sensors. For example, if we consider the case of vision, an AGI system will be overwhelmed when dealing with individual pixels. However, I can easily envision an AGI system that uses a "fat sensors" to supply symbolic features from the raw image it is observing (think edges, blobs, etc.). But these features must not be pre-specified; the ocean-bed on Titan may be slightly different from those on Earth. And we have a very good reason (from cognitive psychology) to suspect that such sensors must also be configured by top-down influence from the mind (i.e. expectations).

As my understanding, "Cognitive integration" is a similar opinion with my "comparatively process the parallel information from video and natural language input". am I wrong? :D

Shangfeng: We certainly agree on integrated and unified knowledge representation.

Attention ...and intention? GAI is to defy searl's argument "chinees room"...

Abdessamad: The Chineese room only makes sense when the system has no symbol-grounding. Clearly, AGI systems cannot work like this. Anything an AGI system experiences will not simply be a string of bits, but a string of bits that get their meaning from past experience.

Abdessamad: "chinese room argument" has many problems. For example, It's based on an "axiom" that a book does not have consciousness. However, this "axiom" is not always right when a book is big enough to answer all the question in natural language. Researchers of AGI do not need to be limited by existing philosophy conclusions. Actually, I believe with the development of AGI, some of the philosophy conclusions will be broken.

Shangfeng : What I addressed in my previous post and as Helgi highlighted with narrow AI still being important is that I would not like AGI to fall into the same trend as narrow AI. I believe that AGI is very important for the future of AI, but in trying to differentiate itself so hard from narrow AI it will end up being the same as narrow AI and not achieving its full potential. That is my fear when we try so hard to draw boundaries instead of being holistic in approaching the problem.

Arturo: I believe the most important feature of AGI is the affordable general knowledge learning when we differentiate AGI from narrow AI. I agree with you that we do not need to focus on the boundaries definition. However, AGI needs its own evaluation standard. And integrated evaluation standard is necessary for AGI research. In NLP area, the separate benchmark standards for dependency parsing, coreference resolution or word sense disambiguation made the research apart. It's the time to create an integrated standard for "understanding".

Taking a holistic [Geigel] or integrative [Hu] viewpoint on intelligence is indeed among points in which AGI differs from mainstream AI. The problem is may be the lack of a commonly accepted "theory of intelligence"...

Abdessamad is right about the "commonly accepted" part, but there have been very notable efforts towards this goal. See "Rigid Flexibility: The Logic of Intelligence" and "A General Theory of Intelligence", both by Pei Wang.

Yes and thanks Dr Helgason: Pei Wang is indeed a reference in AGI and proposes working definition of intelligence in his “General Theory of Intelligence” http://www.cis.temple.edu/~wangp/GTI-book/ .

I guess my concern stems from two points of view

1) Semantic- when I define the concept of General and put constraints on it it becomes semi general and not general. This is not the same problem as in 3.2 of Introduction: Aspects of Artificial General Intelligence by Wang and Goertzel

2) There are some very specific types of problems that do not fall under AGI but may help to achieve it and should be contemplated. For example, I am building a programming language which is designed from the ground up for AI and has process interruption, dynamic state saving (I became tired of putting manual interrupts in for loops on long running data mining tasks), and process migration from interpreter to interpreter. This is part of an immune response I am building into the interpreter design that with other measures will compensate if it somehow damaged (simulating brain compensation). Also, the ability to query the states of the virtual machine and to modify its execution dynamically (controlled meta programming). These are not part of the system that I am building but are the same as providing a "brain" on which to put a "consciousness".

I guess that I read to much into the scope of fields, and it is my sincere hope that AGI does not fall into the same trap as AI did.

What about the “definition” provided by Gubrud in 1997 introducing (for the first time?) the term “artificial general intelligence ”:

“By advanced artificial general intelligence, I mean AI systems that rival or surpass the human brain in complexity and speed, that can acquire, manipulate and reason with general knowledge, and that are usable in essentially any phase of industrial or military operations where a human intelligence would otherwise be needed. Such systems may be modeled on the human brain, but they do not necessarily have to be, and they do not have to be "conscious" or possess any other competence that is not strictly relevant to their application. What matters is that such systems can be used to replace human brains in tasks ranging from organizing and running a mine or a factory to piloting an airplane, analyzing intelligence data or planning a battle.”

It is also to be noted that Gubrud introduces the term in a conference related to advances in the HARDWARE (http://www.foresight.org/Conferences/MNT05/Papers/Gubrud/ ). And by the way, the Japanese Fifth Generation Computer Systems project failure (http://en.wikipedia.org/wiki/Fifth_generation_computer ) was due among other problems to the fact “that existing CPU performance quickly pushed through the "obvious" barriers that experts perceived in the 1980s, and the value of parallel computing quickly dropped to the point where it was for some time used only in niche situations. Although a number of workstations of increasing capacity were designed and built over the project's lifespan, they generally found themselves soon outperformed by "off the shelf" units available commercially.” A hardware trap to be aware of?

A definition is a good beginning. However, to support a new academic branch of artificial intelligence, a practical benchmark is needed. To replace human brain is the final goal but it is too hard to be used to evaluate recent AGI systems. It has the similar problem of Turing test. Turing test expect the AI system to pretend a human chatter. But pretending a human chatter is not always necessary for many applications. An AGI system which can answer questions to solve real world problems in natural language could be highly valuable. I believe a well designed benchmark with high discrimination to quantitively evaluate different AGI systems will be very helpful for the AGI community.