IBM's Watson plays Jeopardy “Unconsciously”

The video above shows Watson, a computer system with artificial intelligence developed by IBM, participating in a televised game show against Jeopardy record holders Brad Rutter and Ken Jennings. The computer uses algorithms (which have been around for a while) to parse out phrases from the jeopardy questions and match them against information in its internal memory. Watson has open access to a memory bank of 200 million pages of text including dictionaries, thesauri, newswire articles, literary works and all of the encyclopedic entries of Wikipedia. This information takes up four terabytes of disk space in its RAM based memory. It is not the first computer to solve these kinds of questions but it is the first to do so in the order of a couple of seconds with very high accuracy. 

Watson parses the clues into different keywords and sentence fragments in order to find statistically related phrases. Through the simultaneous execution of thousands of language analysis algorithms, Watson comes up with a number of potential answers (mostly proper nouns or verbs), that are statistically most likely to match the cues given by Jeopardy's Alex Trebek. Watson chooses the answer that has been found most frequently by different, independently operating, algorithmic searches.  

Watson is not conscious. It has proven that it can best the most skilled human trivia practitioners in the world but it doesn't "understand" the questions that it answers. Watson might be able to narrow its associative search down to a single city but it cannot appreciate why it chose the city, cannot visualize what it would be like to visit this city and has no desire to do so. Although it can come up with a precise list of the most closely semantically related words and concepts, it cannot reflect on these in a creative, insightful or emotional way. In some ways though, Watson's method of processing is very much like our own. Its multiassociative approach is much like the way our unconscious mind can select a single answer when given a set of related cues.

Like Watson we have mental representations of words, perceptions and concepts organized in our head and these are associated to one another in various ways. The associations are determined by our experiences and each concept, is associated with many other concepts. To access the memory of a concept, several other closely related concepts must be activated at the same time. In fact, when a set of concepts are coactivated (like the words twinkle, distant and night sky), sometimes there is only one representation that all of these descriptions sum together to activate. Interestingly, even our brain cells, at a very fundamental level are organized for multiassociativity. Individual neurons must have messages being sent to them by many other neurons, for them to increase their firing rate. In the cognitive, neuroscientific and AI sense, representations made active by a jeopardy question send out outputs to many different dormant representations and the ones that are converged upon the most become active. In other words, multiple cues activate tons of different representations, but only the representations that best match the suite of cues - taken all together - are activated maximally to become conscious.  The important question is, are the associative links in your network of associations wired up in such a way that the representation corresponding to the correct answer will become active?

An answer on Jeopardy a few days ago was: The name of a planet, an element and a Roman god. Within three seconds, the question, "what is mercury," popped into my head. It didn't pop in within a few hundred milliseconds so I must have done some conscious thinking beforehand. But I also know that I did not go sequentially down the list of elements and planets in my head in search for a commonality. The node in my brain (which might correspond to a cluster of neurons or even of cortical minicolumns) that corresponds to the concept of mercury, was converged upon and made active by these other concepts in an automatic and unconscious way. 

Snap judgments are made, unconsciously, in this way. But humans can do more than just make blind, automatic multiassociative guesses. Given time and the motivation to deliberate, humans can contemplate questions from a variety of different perspectives. It seems to me that most Jeopardy questions offer a number of associative cues that will automatically cue up the answer in a contestant’s head if they know it. Some Jeopardy questions seem to offer cues that require additional processing and the use of inferential and deductive thinking. Interestingly, Watson seems to be able to make up for its lack of human logic, on questions like these, with processing power and extensive memory.

To me the important question is, how does one imbue a system like Watson’s with human-like consciousness beyond instantaneous multiassociation? I think that the answer lies in endowing a computing system like Watson with the equivalent of a prefrontal cortex to keep certain representations active through the span of several seconds. Of course, this form of artificial intelligence would have to have a prolonged series of developmental experiences, similar to a childhood, to learn which representations to keep active in which scenarios. By maintaining the activiation of a representation, something that happend moments ago can impact future activity. This may not help much in Jeopardy where each question is meant to stand on its own independent from the other questions. However, in my mind, extended activation is totally responsible for the creativity, insight and emotionality that is necessary for most human activities. 

An Analogy Between the Neurophysiology of Thought and the Polypedal Locomotion of an Octopus

Here is an unabbreviated version of the abstract that I submitted for a January symposium at USC:

The present analogy for the neurophysiology of thought involves a many-armed octopus grabbing and releasing footholds as it pulls itself from place to place. This is meant to illustrate that the thought process involves a cyclical pattern of cortical activation, coactivation and deactivation. Coactivations (footholds held simultaneously by the octopus) fluctuate as cortical areas that continue to receive sufficient activation energy are maintained, areas that receive reduced energy are released from activation and new areas that are tuned so as to receive sufficient energy from the current constellation of coactivates are converged upon, recruited and incorporated into the remaining amalgam of active areas from the previous cycle. Newly recruited areas contribute their inputs to those of the remaining previous inputs altering the mental representations that are produced. Such a newly activated area, or primed node, corresponds to a cortical module (composed itself of neural assemblies) that, when coactivated with other such modules, unites discrete features of long-term memory into composite, global mental imagery. This model defines an individual thought as the imagery produced in primary and secondary sensory areas in response to a particular set of coactivates in association areas. The thought changes once association areas respond to this imagery with newly activated modules and send their new sum of inputs back to the sensory areas for the creation of modified imagery (reperception). The process whereby these modules fluctuate spatio-temporally is taken to be analogous to the nonlinear stride of an octopus that plants the majority of its arms temporarily while actively repositioning arms that lie behind it, toward the front - in the direction of its movement. The fact that some modules are conserved (the arms remain planted), during these reciprocal oscillations between top-down association areas and bottom-up sensory areas, is taken to account for the continuity found in successive thoughts. The result is a stream of consciousness where each thought is slightly different than the ones preceding them as new modules are added and old ones are taken away. Some modules are retained even after a number of thought cycles. This happens when one’s thoughts transition and change but hold a common element constant and is usually due to module potentiation by the PFC. Sometimes modules are not conserved from thought to thought and the octopus drops most of them all at once. This happens when one abandons a train of thought and quickly reorients to a new, salient, perhaps emotionally laden stimulus. This model predicts that someone with a working memory deficit has fewer of these allegorical octopus arms and that the arms cannot maintain their grasp for as long. The longer top-down (higher-order) modules can be continuously activated, over a series of thoughts, the longer they can influence sequences of bottom-up imagery in a sustained and consistent way allowing modeling, planning and working memory in general.

To find out more visit my new website: http://www.cognitivemechanics.net/.



Read the full article that I wrote on this topic here:

http://www.sciencedirect.com/science/article/pii/S0031938416308289

Grandmother Cells

 


The term "grandmother cell" was coined by Jerry Lettvin in 1969 to describe a hypothetical neuron that can be shown to represent a specific psychological concept. This cell would become active every time a person thinks about a complex thing, such as his or her grandmother. The question was: Are there any cells, anywhere in the brain, that are dedicated specificlly to processing information about one's grandmother? It seems that this may be the case despite the fact that neuroscientists were convinced, for decades, that this was a gross oversimplification.

It is known that millions of cells work together to help us visualize even the most simplistic visual objects. The retinas of the eyes relay information about what we look at to early visual processing areas in the back of the head. This data passes through a series of neural areas before objects are recognized. These areas, which have been fine-tuned by life experience, act as filters that allow the visual data to matchup against the brain's best existing representation of what is being seen. In this sense, when we see, we are not really looking at what is out there, but instead piecing together a collage of things that we know to try to recreate the scene. When our brain does this, large numbers of neurons in the simple visual areas that correspond to the lines and contours of what we are seeing send their information to a smaller number of more complex neurons that deal with shape and form. These neurons, in turn, converge on even smaller populations of neurons that code for recognizable objects like people, cars and animals. It is thought that populations of cells, in these high-order processing regions, can be dedicated to processing very specific objects. Grandmother cells are the theoretical limit to this convergence where the activities of a large interconnected structure of networks meet together to activate a single neuron that in some senses, holds much of the information of the entire network (because much of the network must be activated for it to fire).

Some cells that come close to meeting the requirements of a grandmother cell have been found. One study by Rodrigo Quiroga and colleagues used patients undergoing treatment for epilepsy where 100 tiny electrodes were implanted in their brains. Each subject saw around 100 images of famous people, places and things. Overall, almost 1,000 neurons were sampled and 132 of these reacted to at least one of the images. The objects that elicited increased activity were used in another round of recognition except different images of these objects were shown. For example, if a head-on photo of a truck elicited a response, then a profile picture might be shown next. The researchers were able to find grandmother-like cells in some of the participants. One female participant had a neuron that only responded to the actress Jennifer Aniston. The neuron did not respond to other pictures, even of similar-looking female actresses, with one exception. This neuron responded to some pictures of Lisa Kudrow, Jennifer’s costar on the show Friends.

Another female participant had a neuron that responded only to Halle Berry. Pictures of the actress, a line drawing, a profile, even the words of her name, all made this neuron fire. These neurons were generally in convergence areas – such as the hippocampus- where a lot of processing, in areas such as vision and audition, meet up. This neuron must have been highly tuned, not only to the physical aspects of Halle, but to the abstract representation of her overall persona because even a picture of Halle’s catwoman character made the neuron fire despite the fact that she was masked.

Early critics of the idea pointed out that there are not enough neurons in the brain to account for every possible sensory object in one-to-one correspondence. Grandmother cells don’t necessarily work this way though. It seems that sensory objects are represented by networks of neurons many of which overlap and intersect. The individual neurons that make up these networks usually have the capacity to contribute toward the perception and recognition of several different objects of sensation. The further down the processing stream these neurons lie - the closer in the brain they are to the retinal inputs - the more fundamental they are in the process of recognition and the more objects they contribute to. The contribution of an individual neuron is actually very weak. One neuron usually does not have the capacity to send a global message that can be perceived consciously. In fact, many, many neurons would have to be removed to abolish the ability to recognize your grandmothers. Even more (tens of thousands or perhaps millions), would have to be removed to ensure that you could not remember anything about your grandmothers.

Experience fine-tunes neurons and the inputs that they are receptive to. The brain will even tune some neurons so narrowly that they become dedicated processing specialists for things that you recognize frequently, especially things that you see over and over again. They may not become active to every representation of a grandmother, they may become active to totally unrelated representations and there may be a number of them in each person’s brain, that vary in their responsivity to grandmothers. Even so, grandmother cells, at least loosely defined, do seem to exist. It is probably a safe bet that the cartoon drawing of a grandmother at the beginning of this entry activated some of the same cells in your brain that respond uniquely to images of your parent’s mothers.

Quiroga, R. Q., Reddy, L., Kreiman, G., Koch, C. & Fried, I. 2005. Invariant visual representation by single neurons in the human brain. Nature 435: 1102-1107.



Read the full article that I wrote on this topic here:

http://www.sciencedirect.com/science/article/pii/S0031938416308289