EFTA Document EFTA01793418

From: Joi Ito Sent: Tuesday, October 22, 2013 11:26 AM To: Epstein Jeffrey Subject: Fwd: MDF Attachments: signature.asc BTW, getting going with Joscha. He's smart. Let me know if you're =nterested in joining the brain threads. Begin forwarded message: > From: Joscha Bach > Subject: Re: MDF > Date: October 21, 2013 23:56:09 -0400 > To: Joi Ito > Cc: takashi ikegami > Kevin Slavin Greg Borenstein Ari Gesher Martin Nowak > Hi Takashi, hi Ari, hi all, > finally I got around to look at Takashi's talks and his 2010 ACM =rticle. The first thing that came to mind was the distinction between =neat" and "scruffy" AI, which might be described as the clash between =olks that wanted to construct Al by adding function after function, vs. =hose that want to take a massively complex system and constrain it =ntil it only does what it is supposed to do. > The idea of starting from massive data flows is very natural and =heoretically acknowledged, even it is often practically neglected. =ognition, by and large, is an organism's attempt to massively reduce =omplexity, by compressing, encoding, selectively ignoring, abstracting, =redicting. controlling it. Thus, it seems natural to focus on the =echanisms that handle this complexity reduction, which I think is =xactly what most research in computer vision, machine learning, =lassification, robot control etc. is doing. A lot of the work on =roblem solving and learning within cognitive science even works _only_ =n the highest level of abstraction, i.e. grammatical language, regular =oncept structures, ontologies and soon. > If I understand Takashi correctly, he points towards another > =erspective: (please forgive and correct me if I should oversimplify too =uch here) 1. Cognitive systems do not only need to reduce complexity, but also =uild it (for instance, take simple cues or abstract input and use it to =eed a rich, heterogenous, ambiguous and dynamic forest of =epresentations). > 2. Cognitive processes that work directly on and with high complexity =ata are under-explored. > 3. The study of systems that are immersed in such complexity might =pen the door to understanding intelligence and cognition. > There is really much more in Takashi's talk, but let me respond to =hese in turn: > 1. I believe that cognition is really about handling massive data =lows, by encoding it in ways that the cognitive agent can handle and =se to fulfill its demands. This works mostly by identifying apparent =egularities and turn them into perceptual categories, features, =bjects, concepts, ontologies and so on. Our nervous system offers =everal levels and layers of such complexity reduction, the first one of =ourse at the transition between sensory inputs and peripheral EFTA_R1_00126387 EFTA01793418 nervous =ystem (for physiological, tactile, proprioceptive input), or, in the =ase of visual perception, the compression we see between retina and =ptic nerve. The optic nerve transmits massively compressed data from =he retina to the thalamus, and from there to the striate cortex (the =rimary visual cortex, V1). V1 is the lowest level of a hierarchy of =isual and eventually semantic processing regions: from here, the dorsal =nd ventral processing streams head off into the rest of the cortex. V1 =ontains filtering mechanisms, which basically look for blobs, edges, =ovements, directions and soon, based on local contrasts. V2 organizes =hese basic features into a map of the visual field, including contours, =3 detects large, coherently moving patterns, V4 encodes simple =eometric shapes, VS seems to take care of moving objects, and V6 =elf-motion. The detection of high-level features always projects back =nto the lower levels, to anticipate and predict the lower level =eatures that should be isolated based on the higher-level perceptual =ypothesis. The story is similar for auditory processing, and eventually =he integration of basic visual and optical percepts into semantic =ontent: at each level, we take extremely rich and heterogeneous =atterns and reduce their complexity. > The transformation from concepts to language also represents another, =ncredible level of complexity reduction. > The highest complexity reduction, however, takes place at the =nterface between conscious thought and all the other processes. I =elieve that the prefrontal cortex basically holds a handful of pointers =nto the associative cortical representations, skimming off only a =andful objects, relations or features at a time, and bring them into =he conscious focus of attention. > The perspective of the need for staying at a complex level is entirely =arranted, though: there are many intermediate representations that =Ilow cognitive processes only if the complexity stays high, and might =ven need to increase it. This includes many sensor-motor coordination =rocesses, but also most creative, more intuitive exploration. > This is not the same complexity as the one at the input, however! This =s a level where data is already split into modalities, semantically =rganized and so on. On the other hand, it is much more complex as =inguistic or cognitively accessible types of mental content. > 2. Scientists tend to have a fixation on thinking with language, and =t is quite natural to fall for abstract, a-modal representations, such =s predicate logic systems or extensions of these when it comes to =odeling cognition and problem solving. This might explain the fixation =f cognitive architectures like Act-R and Soar on rule-based =epresentations, and the similar approaches of a lot of work in =lassical Al. > On the other hand, there is a lot of work on learning and > =lassification to handle vast complexity, with the goal of reducing > it. =A particular beautiful example was Andrew Ngs work on deep > learning, =here his group took 30 million randomly chosen frames from > Youtube, and =rained an unsupervised neural net to make sense of them. > They ended up =ith spontaneously emerging detectors for many typical > object =ategories, including cats and human faces. I could not avoid > to think =f that paper when Takashi mentioned his fascination with > looking at TV =ixels directly...) --> http://arxiv.org/pdf/1112.6209.pdf > Thus, the typical strategies seem to encompass "abstract 2 abstract" =ognition, and "complex 2 abstract" cognition. What about "abstract 2 =omplex" and "complex 2 complex"? Most of the existing approaches on =complex 2 complex" cognition are not really cognitive, such as Ansgar =redenfeld's "Dual Dynamics" architecture, or Herbert Jaeger's Echo =tate Networks. The current proponents of such complex cognition are =lso often radical embodimentalists (cognition as an extension of sensor =otor control, neglecting dreams, creativity, imagination, and =apabilities for abstract thinking). > 3. The idea of getting to artificial intelligence just_ by "looking > =t" (blind deep learning) on complex data flows is not new. I think that =here are at least two aspects to it: deriving a content structure that =Rows the identification and exploitation of meaningful semantic =elationships (for instance, discerning space, color, texture, causal =rder, social structure, ... for instance simply by analyzing all of =outube, or by 2 EFTA_R1_00126388 EFTA01793419 collecting data from a robotic body and camera in a =hysical world), and the integration of that structure with an =rchitecture that is capable of thought, language, intention, goal =irected action, decision making, and so on. The former is tricky, the =atter impossible. Complexity itself does not define intentional action, =nd the differences between individuals and species should not be =educed to differences in complexity perceived by the respective agents. a I agree that we need to gain a much better understanding of "complex 2 =omplex" cognition, but that must integrate, not replace what we already =now about the organization of cognitive processes. I am certain that =ur current models are a long way off from capturing the richness of =onscious experience of our inner processes, and even more so from the =uch greater complexity of those processes that cannot be experienced. > Another interesting point I gathered from Takashi's talk is the idea =f something we might call "hyper-complex" cognition. The complexity =andled by our human minds (as well as the one of Andrew Ng's deep =earning Youtube watching networks) builds on very simple stimuli. But =hat if the atoms themselves are abstract or highly complex, for =nstance because they are already semantic internet content? The =ognitive agents handling those elements may essentially be operating at = level above human cognition if they are capable of operating on that =omplexity without reducing it. Unlike humans which are forced to =ranslate and reduce all content into their individual frame of =eference, and access it only through a single perspective at a time, =rtificial agents do not need to obey such restrictions. Today's Big =ata moniker probably marks just the beginnings of the abilities of =achines to make sense of abstract and complex input data. > Cheers, > Joscha »» Fascinating. Ikegami is taking a very interesting tack: >>» »» http://www.youtube.com/watch?v=tOLIHhjNIBc »» http://sacral.c.u-tokyo.ac.jp/pdf/ikegami_ACM_2010.pdf >>» »» For me, this is similar to the discussions that you and I and Kevin =ave been having about auto-didactism: starting from complexity rather =han abstraction (which is generally antithetical to academic learning). =It would seem to me that most artificial intelligence research has =tarted from abstraction (and forgive my ignorance if I'm off base here) =nd attempted to build up to complexity. My very cursory look at the =oscha's MicroPSl work seems to show an approach moving in the direction =f the what Ikegami did with the MTM from the classical =bstraction-first approach. MicroPSl places its constructs in a reduced =idelity virtual environment, has lower-level abstractions, and brain =tructures/dynamic pre-synthesized for things like motivation, emotion, =please correct me if I'm off base - like I said: cursory). The brain =tructures in living systems have have evolved as low-energy means of =rocessing brain signals (both sensory data flows and internally routed =treams) once they have showed fitness - ultimately, they were =and-blasted into their shape by generations of massive data flows. We =ave an understanding of what purpose they serve but not a good =nderstanding of how they work (maybe I'm behind on the state of the art =n neuroscience on that point?). >>» »» Ikegami is starting from the complexity and seeing what emerges - =hich seems to me to mirror the rise of consciousness in natural =ystems. Mind is the surfer that hangs on the eternal wave of the =assive data flow of sensory input without wiping out. Somehow, the =eality of the temporally continuous observer arose from exposure to =ensory data flows and the evolution of the complexity of the brain. =kegami is shortcutting the snail's pace of the physical evolution of =atural systems by synthesizing a neural network of sufficient =omplexity as well as high-resolution sensors. >>» »» Thinking about modern synthetic data flows (you know.... the =nternet!) as being as rich as sensory data leads one to imagine some =nteresting possibilities in a) whimsically, the spontaneous emergence =f consciousness and b) 3 EFTA_R1_00126389 EFTA01793420 practically, new techniques for dealing with =hat massive data flow that mimic something like natural consciousness. =here's nothing in the practical world of big data that really looks =ike the MTM (that anyone is talking about - who knows what lurks in the =igh frequency trading clusters busily humming in the carrier hotels). =verything that Google and Facebook and the like seems to be doing is =uch simpler than anything like this. >>» >>» »» On Oct 19, 2013, at 9:37 AM, Joi Ito >>» >>>» >»» http://www.dmi.unict.itiecal20B/workshops.phpft4th-w >>>» >»» - Joi wrote: Please use my alternative address, [email protected] to avoid email auto =esponder 4 EFTA_R1_00126390 EFTA01793421