Zero Brane

Roomba and intelligence

2008-06-29T21:22:04Z

Few days ago I attended one of the Google Seattle Tech Talks; this one was given by Helen Greiner, iRobot's co-founder and chairman. She was visiting Seattle to sign the licensing agreement for UW's Seagliders and stopped at Google Kirkland to talk about iRobot and its robots.

Her talk was interesting, not so much for the descriptions of robots that iRobot develops, but mostly for the history of the company and their learning experience as they were trying to design and build robots and find their market to sell products to survive. This somewhat reminded me the story of Webmind AI Engine (Waking Up from the Economy of Dreams), albeit with a happy ending as iRobot is a live and profitable company (although its stock has been trading recently close to 52-week low).

There was a brief Q-and-A session in the end and I had an opportunity to ask a few questions, one of them was "When do you see your robots becoming more intelligent?" Helen's response was that intelligence is a complex term, but their robots are intelligent to some degree: they don't fall off the stairs, can sense objects around them, interact with virtual walls, can re-dock themselves, and, most importantly, they clean carpets and floors and do a good job at that. Essentially -- what follows is obviously my interpretation -- they are well designed for a particular task and there is no obvious need or an easy way for them to become more intelligent.

As I own several generations of Roombas I'm familiar with how they operate and what they are capable of. Still, I struggled to counter Helen's definition of intelligence without falling prey to Tesler's Theorem -- "AI is whatever hasn't been done yet" (Godel, Escher, Bach: an Eternal Golden Braid, p.601). Do Roombas possess even limited intelligence? I do not believe so and here is why.

It's not so much the ability to solve a problem that is critical to intelligence as it is the ability to learn how to solve problems. Most of the tests for intelligence seem to be focused on testing ability to solve problems, relying on the assumption that if you can solve those problems you learned the ability to solve them, hence you're intelligent. A better way to test intelligence would be have some sort of dynamic test when you learn to solve problems as part of the test and the progress you make would define your level of intelligence. Otherwise we don't really test intelligence so much as we test experience.

Wang From NARS to a Thinking Machine, p.2: "...whether the system is intelligent, or how intelligent it is, is not determined by what the system can do, but what it can learn to do."

Roombas don't learn; they don't improve their actions. For example, I have 5-inch space between floor transition in my kitchen and thick carpet in my living room and Roomba keeps getting stuck there (literally; it can't move and needs to be picked up). With more intelligence, it could possibly learn to avoid those spots or maybe go through them using a different direction. They also don't acquire new actions and don't generate actions to do their job more effectively. Even their initial "experience" is not acquired, but hard wired based on experience of its designers. While there are noticeable differences between the first and last generations of Roomba -- more autonomy, ability to dock itself and to sense objects and slow down before hitting an obstacle, and other things -- this is still not enough to call it intelligent.

The more difficult question is, if we add more sensors, mapping capabilities, working memory, and other components, would it make it intelligent then? I will not attempt to answer this question here, but will do a separate post on how I define intelligence and what it may take to make an intelligent robot.

Blind Chess and Working Memory

2006-03-31T08:04:16Z

When I was a kid I liked playing blind chess with my dad and my brother. This is the type of chess when you make moves without looking at the board at all (you just tell your opponent what piece to move and where). I couldn't beat my father (to my defence I couldn't beat him very often in regular chess either), but I could play through end-game with 30 moves or more.

Now I want to compare this skill with my inability to remember a 10-digit phone number after hearing it one time or copy a 9-digit loan number between two computer screens from memory (I need to look at it at least twice). We all know about a span of immediate memory and its limit of seven plus or minus two items (THE-MAGICAL-NUMBER-SEVEN), which (partially) explains my difficulties with numbers. I just don't see how to reconcile it with my ability to play blind chess.

For those who never played blind chess, here is a brief summary of what you need to do. You need to 1) keep the current position in your head (initially 32 pieces on 64 squares), 2) be able to evaluate this position, and 3) consider possible moves and your opponent's responses (for any serious game this analysis needs to be done for several iterations/levels).

One possible explanation is that blind chess has more to do with mental imagery and less with working memory. While imagery is definitely involved here, I don't think this solves the problem. For starters, I don't imagine seeing the board as you'd see it in real life. The board that I see has no texture and no color; even squares are not colored black and white. I can "paint" it any color I want, but this will require some effort and "by default" it's completely feature-less. Much to my own suprise (this is the first time I examine this aspect of my own mental imagery), pieces too have no colors and even no shapes. I just "know" that this square has my pawn and that square has opponent's bishop. As with board colors, I can make them look anything I want, but this doesn't change the fact that by default they don't have any look. It's difficult to explain, but I just know conceptually that this is my piece on this square. Which makes me think that this has less to do with mental image and more with working memory that not only stores all those concepts and their relationships, but is also involved in analysis of the position and future moves.

Another possible explanation is that each of those tasks involves a different type of working memory; alternatively, there may be no separation, but working memory may have different task-specific capacity limits. For example, this article provides seven views on this capacity limit:

There are capacity limits but that they are in line with Miller's 7+2.
Short-term memory is limited by the amount of time that has elapsed rather than by the number of items that can be held simultaneously.
There is no special short-term memory faculty at all; all memory results obey the same rules of mutual interference, distinctiveness, etc.
There may be no capacity limits per se but only constraints such as scheduling conflicts in performance and strategies for dealing with them.
There are multiple, separate capacity limits for different types of material.
There are separate capacity limits for storage versus processing
Capacity limits exist, but they are completely task-specific, with no way to extract a general estimate.

Yet another explanation is that somehow I trained myself to store and process all this information. I don't buy this argument simply because I have been dealing with short sequences of digits for much longer than with blind chess, but without any significant progress.

I think it has something to do with the fact that "chunks" that I'm trying to remember in the first case (digits in a phone/loan number) are not related, while in the second case they are tightly related to each other. Even more importantly, what needs to be captured is a sequence of digits, rather than a group. This may be caused by mutual interference between elements in the sequence; as a result the sequence requires significant effort to maintain itself. If those number can be somehow related to each other (for example, they may be similar to a social security number) it may be much easier to remember by capturing "social security number with some modifications...". I think this also supports the idea of a hierarchical organization of (working) memory as while the chess play requires to maintain significantly larger number of elements, they all seem to be maintained "inside" that concept without much bear on working memory. Yes, you need to focus and pay attention when you play blind chess, but I'd argue that the effort required is probably less than when you need to remember a 9-10 digit sequence for 30 seconds.

Note that this effort doesn't seem to be proportional to the number of pieces to maintain in memory: it's easier to keep the position in memory in the beginning of the game when all 32 pieces are present rather than during end game where only few pieces may be left. This may be related to the fact that the initial position is more constrained (has fewer degrees of freedom) than any end game position, which makes it easier to track.

Working Memory: a Black Box?

2006-03-13T06:01:21Z

Chris Chatham posted a brief overview of the classic working memory model and offered a new diagram for that model. There are several interesting points, but I'd like to emphasize two of them: processing and memory are two sides of the same coin and gating functions are likely to be present at every intersection of arrows on the working memory diagram.

Having said that, I side with O'Reilly and his collaborators who proposed a completely different working memory model outlined in their article BANISHING-THE-HOMUNCULUS. The model is based on tripartite architecture, which is composed of the posterior cortex (PC) that performs majority of "automatic" sensory and motor processing, the hippocampus (HC) that is responsible for rapid learning that binds together arbitrary information, and the prefrontal cortex and the basal gandlia (PFC/BG) system that maintains internal contextual information (PFC), which can be dynamically updated by the BG.

In this model working memory is defined as an emergent property of the interactions between these three brain areas. These mechanisms not only support those basic memory functions that are generally associated with working memory, but also those controlled processing functions that are typically associated with a "central executive". The critical difference between this model and the model covered in Chris's post is that all information is viewed to be distributed in a relatively stable configuration throughout the cortext with working memory being represented by the controlled activation of those distributed representations. Authors' view is that working memory and executive function are two sides of the same coin, based on the fact that processing and memory functions are typically distributed within and performed by the same neural substrates.

As a side note, while a framework that separates working and long-term memory may look familiar and computantionally appealing, think about "simple" operation of copying a concept from long-term to working memory. What mechanism would support this copying that (to properly reproduce the concept) would need to include not only sensory and motor information associated with it, but also (potentially) large number of related concepts? How would "activation" of such a copy look like?

The authors also identify six key functional demands underlying working memory and describe how those are addressed in the proposed model:

Rapid updating
Robust maintenance
Multiple, separate working memory representations
Selective updating
Top-down biasing of processing
Learning what and when to gate

There is one more interesting point in the paper that I'll expand on later:

...it is likely that working memory may represent a kind of phylogenetic extension of the same kinds of mechanisms that underlie all forms of complex motor coordination and planning. (emphasis mine)

update 2006/03/20: More on working and long-term memory processes and assessment from Intelligence Testing.

Classes vs. Instances

2006-01-30T23:02:58Z

I've been long puzzled by the fact that our brains can easily switch between working with classes and individual instances that represent those classes. Leveraging the idea that polychronous groups may represent symbols in the brain we can try to answer questions related to representation of classes and instances:

Are there symbols that represent classes, instances, or both?
Can a single symbol represent both, depending on how it's activated?
How do we define/measure "closiness" of two symbols? For example, is the sun closer to a tree than a submarine? Do we measure "closiness" in terms of activity/effort to undertake to go from one symbol to another one?

Jeffrey Elman in his papers FINDING-STRUCTURE-IN-TIME and AN-ALTERNATIVE-VIEW-OF-THE-MENTAL-LEXICON provides answers to these questions by using a model based on a simple recurrent network (SRN) where hidden unit patterns are fed back to themselves serving as the context for subsequent input patterns. The network was trained to predict the next word based on a corpus of sentences that were generated by a simple artificial grammar. The network was presented with words one by one and was tasked with predicting the next word. The network learned to predict words that were grammatically possible depending on the context.

Elman then looked at how the network categorized captured information. To analyze the similarity structure he averaged hidden unit activations for each word+context combination (after completing the learning phase) and then calculated Euclidian distance between generated vectors for each word. The network learned to partition network's "mental" space into major categories, which were further subdivided into smaller categories (like humans or food).

There are several interesting things about this category structure.

It appears to be hierarchical.
This structure is "soft" and implicit, with some categories being quite distinct and others having less distinct boundaries and sharing properties with other categoires.
The content of the categories and their structure is not known to the network.
Network's representations are highly context-dependent, which means that there could be separate representations for the same word that occurs in every different context. All concepts are expressed in a distributed manner as activation patters over a fixed number of nodes. A given node participates in representing multiple concepts. The activation of an individual node may be uninterpretable in isolation -- it is the activation pattern that is meaningful in its entirety. (FINDING-STRUCTURE-IN-TIME; p.21)

One can clearly see some parallels between this description and ideas expressed in the previous post on symbols and polychronous groups.

Capturing information about instances rather than classes provides even more benefits because "this notion of categories as emergent from the location in a high-dimensional state space, in which at any given moment different dimensions might be attended to and others ignored, suggests that different viewing perspectives on that space might yield new categories." (AN-ALTERNATIVE-VIEW-OF-THE-MENTAL-LEXICON; p.4) Also, once in place, this category structure supports generalization as every new word, identified as belonging to an existing category, inherits all the properties that are assigned to that category. Compare this with the prototype principle expressed in GODEL-ESCHER-BACH (p.352) as:

The most specific event can serve as a general example of a class of events.

You may notice that all those categories that "emerge" as the result of this hierarchical clustering are not explicitly represented in the model; an external observer attaches a label of "animates" or "inanimates" to groups of instances that represent those classes. How would a label like this, for example, "animates" be represented? I think it would not necessarily belong to the same space that is occupied by the instances in the class it represents; rather its relatioships with items in that class or category would be rule-based ("I know if something moves on it's own, it's part of the animates class").

Elman's paper doesn't specifically address the related case of category splitting: how do we acquire "labrador" or "cocker spaniel" category after acquiring the "dog" category? From that we can go to "Snoopy", which may be an instance of the "cocker spaniel" category, or even to "my dog Snoopy".

Overall, it seems like even a simple recurring network can produce quite complex results. It might be interesting to replace hidden and context units with polychronous groups and see whether it's possible to reproduce Elman's results. I see at least two challenges with using the group approach: one is proper assignment of input and output units (how do you read the state "group A is active") and the second is measurement of "closiness" of two groups to do clustering analysis.

Symbol Polychronization

2006-01-27T22:24:07Z

So far I've been focusing on the modeling of individual neurons and their networks, but after reading more on throught processes in GODEL-ESCHER-BACH (pp.348-50), I thought it might be interesting to look at the level of neural complexes that represent symbols in the brain, rather than at the level of individual neurons and signals between them.

There are many questions that come to mind concerning symbols and their representations in the brain:

How many neurons need to be connected to form a neural complex?
Can a neuron belong to more than one complex?
To how many complexes can a single neuron belong?
How many neurons need to be activated for the symbol to become active?
Are there certain core neurons for each symbol that become active every time the symbol is activated?
How many neurons can two symbols share?
How similar are complexes that represent the same symbol in different brains?
What regions of the brain these complexes extend into?

Then Hofstadter offers a hypothesis about symbol representation in the brain (pp.356-7):

...overlapping and completely tangled symbols are probably the rule, so each neuron, far from being a member of a unique symbol, is probably a functioning part of hundreds of symbols.

...in order to distinguish one symbol's activation from that of another symbol, a process must be carried out which involves not only locating the neurons that are firing, but also identifying very precise details of the timing of the firing of those neurons. Thus perhaps several symbols can coexist in the same set of neurons by having different characteristic neural firing patterns.

These ideas come very close to ideas expressed by Izhikevich in his paper on polychronization, which is defined as reproducible time-locked but not synchronous firing patterns with millisecond precision. Neurons are spontaneously organized into polychronous groups by the spike-timing-dependent plasticity (STDP) changes that select conduction delays to allow the groups to form. As each neuron participates in many groups, firing with one group at one time and with another group at another time, the number of coexisting polychronous groups could be far greater than the number of neurons in the network and even greater than the number of synapses.

What is the significance of polychronous groups? We hypothesize that polychronous groups could represent memories and experience. (p.261)

...the system has potentially enormous memory capacity and will never run out of groups, which could explain how networks of mere 10¹¹ neurons (the size of the human neocortex) could have such a diversity of behavior. (p.270)

Let's entertain the idea that polychronous groups are, in fact, neural complexes described above that represent symbols in the brain. Now we can try to answer questions on the list above:

How many neurons need to be connected to form a neural complex? Probably as few as five. Izhikevich's paper shows five neurons with optimized delays that generate 14 groups.
Can a neuron belong to more than one complex? Yes.
To how many complexes can a single neuron belong? Potentially more than the number of synapses it has.
How many neurons need to be activated for the symbol to become active? From the paper (p.267): "When a group is activated, whether in response to a particular stimulation or spontaneously, it rarely activates entirely. Typically, neurons at the beginning of the group polychronize, that is, fire with the precise spike-timing pattern imposed by the group connectivity, but the precision fades away as activation propagates along the group. As a result, the connectivity in the tail of the group does not stabilize, so the group as a whole changes."
Are there certain core neurons for each symbol that become active every time the symbol is activated? According to this model, there are always some neurons that need to be activated for a group to become active.
How many neurons can two symbols share? I don't think there are any limits. Two groups can share all their neurons and have separate activation patterns.
How similar are complexes that represent the same symbol in different brains? Not similar at all.
What regions of the brain these complexes extend into? I can't give you any definitive answer, but probably not all. Can it be that we have two types of regions in the brain: some that are formed with specific neuronal circuits and others that have randomly connected networks of neurons that self-organize into groups based on their inputs?

After thinking about some of the answers I came up with more questions:

Can groups be replicated? It seems like all we need to do is to provide the same input and then the group(s) emerge after STDP selects proper delays. It may not be the same group, but it will represent the same concept in a different place in the brain. Or will it? Is a symbol defined by its neuronal representation or by its connections with other symbols?
How do we think of something we never saw before? How do we connect those symbols to form a new concept? Do they need to form one group? Do they need to be active at the same time? How do we imagine three dogs in a teacup? Do we imagine three dogs somewhere else and then make it look like a teacup when we focus our attention on it? More on this from GODEL-ESCHER-BACH (p.362): "As we imagine a hypothetical event, we bring certain symbols into active states -- and depending on how well they interact (which is presumably reflected in our comfort in countinuing the train of thought), we say the event 'could' or 'could not' happen."
How do we keep symbols active? How do we morph them into something they are not? Do we create a new symbol based on the existing one every time we do this or do we temporarily change the existing symbol?

It's also interesting to note that, while any particular neuron can be a part of many groups, it can't participate in all those groups at the same time; at any given moment it can only be a member of at most one (?) group.

24 H in a D

2006-01-25T07:19:52Z

Try to guess what the letters mean.

Intelligence test, part 1:

24 H in a D
26 L of the A
7 D of the W
... and 30 more

Intelligence test, part 2:

5 to 7 C of the W
1 O G E 4 Y
... and 22 more

Definitely fun, but doesn't look like a substitute for the Turing Test.

Liftability of Intelligence

2006-01-22T21:52:53Z

GODEL-ESCHER-BACH on two basic problems in the unraveling of thought process in the brain:

One [problem] is to explain how the low-level traffic of neuron fitings gives rise to the high-level traffic of symbol activation. The other is to explain the high-level traffic of symbol activation in it's own terms -- to make a theory which doesn't talk about the low-level neural events. If this latter is possible -- and it's a key assumption at the basis of all present research into Artificial Intelligence -- then intelligence can be realized in other types of hardware than brains. Then intelligence will have been shown to be a property that can be "lifted" right out of the hardware in which it resides -- or in other words, intelligence will be a software property. This will mean the the phenomena of consciousness and intelligence are indeed high-level in the same sense as most other complex phenomena of nature: thet have their own high-level laws which depend on, yet are "liftable" our of, the low levels. (p.358)

KINDS-OF-MINDS echoes this idea of liftability and realization of intelligence and the mind itself in other types of hardware:

Why couldn't artificial minds, like artificial hearts, be made real -- realized -- out of almost anything? Once we figure our what minds do (what pains do, what beliefs do, and so on), we ought to be able to make minds (or mind parts) out of alternative materials that have those competences. And it seems obvious to many theorists -- myself included -- that what minds do is process information; minds are the control systems of bodies, and in order to execute their appointed duties they need to gather, discriminate, store, transform and otherwise process informatin about the control tasks they perform. (pp.68-9)

If you opt for this sort of system -- pure signaling system that transmits information and almost no energy -- then it really makes no difference at all whether the signals are electrons passing through a wire or photons passing through a glass fiber or radio waves passing through empty space. In all these cases, what matters is that the information not be lost or distorted because of the time lags between, [for example,] the turning of the wheel and the turning of the rudder. (p.71)

At the Core of Intelligence

2005-12-10T22:21:34Z

I've been reading GODEL-ESCHER-BACH and came across the definition of intelligence (p.26):

Essential abilities for intelligence are:

to respond to situations very flexibly;

to take advantage of fortuitous circumstances;

to make sense out of ambiguous or contradictory messages;

to recognize the relative importance of different elements of a situation;

to find similarities between situations despite differences which may separate them;

to find distinctions between situations despite similarities which may link them;

to synthesize new concepts by taking old concepts and putting them together in new ways;

to come up with ideas which are novel.

Rodney Cotterill in ENCHANTED-LOOMS defines inteligence in a similar way (p.6):

...the defining features of intelligence are the faculty for learning from experience, and the ability to apply acquired knowledge to fresh circumstances. ...[it] entails the ability both to imagine a variety of scenarios and to discern between consequences that are similar but not identical. The handling of novel situations thus implies categorization, association and generalization.

In his interview MIND-A-MOVING-STORY that was taken after the book was published, Cotterill refined his definition of intelligence into:

...a measure of the ability to link elementary motor elements into more complex movement scenarios.