Traditional Definition of Artificial Intelligence Trends - PDF document

Traditional Definition of Artificial Intelligence Trends • “ Artificial Intelligence (AI) is the part of in Artificial Intelligence computer science concerned with designing and Artificial Life intelligent computer systems, • that is, systems that exhibit the characteristics we associate with Bruce MacLennan intelligence in human behavior — Dept. of Computer Science • understanding language, learning, www.cs.utk.edu/~mclennan reasoning, solving problems, and so on.” — Handbook of Artif. Intell. , vol. I, p. 3 2005-02-01 1 2005-02-01 2 Example of Propositional Traditional AI Knowledge Representation • Long-term goal: equaling or surpassing human intelligence IF • Approach: attempt to simulate “highest” human 1) the infection is primary-bacteremia, and faculties: 2) the site of the culture is one of the sterile sites, and – language, discursive reason, mathematics, abstract 3) the suspected portal of entry of the organism is the problem solving gastrointestinal tract, • Cartesian assumption: our essential humanness THEN resides in our reasoning minds, not our bodies there is suggestive evidence (.7) that the identity of the – Cogito, ergo sum . organism is bacteroides. 2005-02-01 3 2005-02-01 4 Graphical Representation Formal Knowledge- (Semantic Net) Representation Language warm- mammal • dog(Spot) • Spot is a dog blooded • Spot is brown • brown(Spot) • Every dog has four • ( � x )(dog( x ) � four-legs legs four-legged( x )) Example dog • Every dog has a tail • ( � x )(dog( x ) � tail( x )) Inference • ( � x )(dog( x ) � tail • Every dog is a mammal mammal( x )) • Every mammal is • ( � x )(mammal( x ) � Spot brown warm-blooded warm-blooded( x )) 2005-02-01 5 2005-02-01 6 1

The Cognitive Inversion Five Stages of Skill Acquisition • Computers can do some things very well that are difficult 1. Novice for people • learns facts & rules to apply to simple “context-free” features – e.g., arithmetic calculations 2. Advanced Beginner – playing chess & other abstract games – doing proofs in formal logic & mathematics • through experience, learns to recognize similar situations – handling large amounts of data precisely 3. Competence • But computers are very bad at some things that are easy for • uses developing sense of relevance to deal with volume of facts people (and even some animals) 4. Proficiency – e.g., face recognition & general object recognition • analytical thinking is supplemented by intuitive organization & – autonomous locomotion understanding – sensory-motor coordination 5. Expertise • Conclusion: brains work very differently from digital • skillful behavior is automatic, involved, intuitive, and fluent. computers 2005-02-01 7 2005-02-01 8 “The New AI” The 100-Step Rule • A new paradigm that emerged in mid-80s • Convergence of developments in: • Typical recognition – philosophy tasks take less than – cognitive science one second – artificial intelligence • Neurons take several • Non-propositional knowledge representation milliseconds to fire – imagistic representation & processing • Therefore then can be – propositional knowledge as emergent at most about 100 • Neural information processing sequential processing – connectionism (implicit vs. explicit representation) steps – =critical dependence on physical computation 2005-02-01 9 2005-02-01 10 Imagistic Representation Multiple Intelligences (Howard Gardner) • Much information is implicit in an image • But can be extracted • linguistic • naturalistic when needed • logico-mathematical • intrapersonal • Humans have • spatial • interpersonal prototype images for each basic category • musical • existential • Brains use a kind of • bodily-kinesthetic analog computing for image manipulation 2005-02-01 11 2005-02-01 12 2

Propositional Knowledge as Artificial Emotions? Emergent & Approximate • Have been neglected (in cognitive science & AI) due to Cartesian bias • System may only appear to be following rules • Importance of “emotional intelligence” now – a spectrum of rule-like behavior recognized • Recognition of situation can be fuzzy & context- • Emotions “tag” information with indicators of sensitive relevance to us • Extraction of relevant elements can be context- • Emotions serve important purposes in sensitive – motivating & directing behavior – modulating information processing • May explain subtlety & sensitivity of rule-like • Artificial emotions will be essential in behavior in humans & other animals autonomous robotics 2005-02-01 13 2005-02-01 14 Neural Information Neural Density in Cortex Processing • 100-Step Rule & Cognitive Inversion show brains operate on different principles from digital computers – “wide & shallow” vs. “narrow & deep” • How do brains do it? • 148 000 neurons / sq. mm • Can we make neurocomputers ? • Hence, about 15 million / sq. cm 2005-02-01 15 2005-02-01 16 Relative Cortical Areas Macaque Visual System 2005-02-01 17 2005-02-01 18 (fig. from Van Essen & al. 1992) 3

Hierarchy Bat Auditory of Cortex Macaque Visual Areas 2005-02-01 19 2005-02-01 20 (fig. from Van Essen & al. 1992) (figs. from Suga, 1985) How Dependent is Intelligence Neurocomputing on its Hardware? Traditional View • Artificial Neural Networks – implemented in software on conventional computers • Brain is no more powerful than Turing machine – are trained, not programmed • Human intelligence is a result of the program – “second-best way of doing anything” running on our brains (Cartesian dualism) – poor match between HW & SW • The same program could be run on any Universal • Neurocomputers TM – goal: design HW better suited to neurocomputing • In particular, it could run on a digital computer and make it artificially intelligent – massively-parallel, low-precision, analog computation • Ignores “performance” (as opposed to – electronic? optical? chemical? biological? “competence”) 2005-02-01 21 2005-02-01 22 Connectionist Natural Computation View • Computation occurring in nature or inspired by computation in nature • Information processing on digital computers ( hardware ) is fundamentally different from that in • Characteristics: brains ( wetware ) – Tolerance to noise, error, faults, damage • The flexible, context-sensitive cognition we – Generality of response associate with human intelligence depends on the – Flexible response to novelty physical properties of biological neurons – Adaptability • Therefore, true artificial intelligence requires – Real-time response sufficiently brain-like computers – Optimality is secondary ( neurocomputers ) 2005-02-01 23 2005-02-01 24 4

Importance of Embodied Intelligence • Traditional (dualist) view: mind is essentially independent of the body Embodied Intelligence – in principle, could have an intelligent “brain in a vat” • Now we understand that much of our knowledge is implicit in the fact that we have a body • Also, our body teaches us about the world • Structure of body is foundation for structure of knowledge • A “disembodied intelligence” is a contradiction in terms? 2005-02-01 25 2005-02-01 26 “Social Interaction” Embodied Artificial Intelligence Rodney Brooks’ Lab (Humanoid Robotics Group, MIT) • Therefore a genuine artificial intelligence • Cog attending to must be: visual motion – embedded in a body • Orients head & eyes to – capable of interacting significantly with its motion environment • (Arm & hand motion • We expect the intelligence to develop as a are not relevant to consequence of interaction of its body with interaction) an environment including other agents 2005-02-01 27 2005-02-01 (video < Brooks’ lab, MIT) 28 Kismet (Brooks’ Lab, MIT) Giving the Computer a Face (Brooks’ Lab, MIT) • Example of three-way conversational interaction • Models: – head & eye orientation – motion tracking – turn taking – facial expression • Does not “understand” speech 2005-02-01 (image < Brooks’ lab, MIT) 29 2005-02-01 (video < Brooks’ lab, MIT) 30 5

Autonomous Robots Starting Small • The ultimate test of intelligence is to be able to • In science, it’s generally considered prudent function effectively in a complex natural to start by studying the simplest instances of environment a phenomenon • Natural environments do not come parsed into context-free categories • Perhaps it is premature to attempt human- • Natural environments are characterized by scale artificial intelligence complexity, unpredictability, uncertainty, • It may be more fruitful to try to understand openness, & genuine novelty the simplest instances of intelligent • There is also a practical need for autonomous behavior robots 2005-02-01 31 2005-02-01 32 Mound Building Collective Intelligence by Macrotermes Termites 2005-02-01 33 2005-02-01 34 Structure of Mound Fungus Cultivator Ants • “Cultivate” fungi underground • Construct “gardens” • Plant spores • Weed out competing fungi • Fertilize with compost from chewed leaves 2005-02-01 figs. from Lüscher (1961) 35 2005-02-01 36 6