Tuesday, December 31, 2013

Secrets of the Holy Grail, Part V

Part I, II, III, IV, V

Abstract

Previously in this article, I explained how the brain's perceptual learning and memory storage system is organized. In today's post, I reveal the surprising source of my understanding. Here goes.

Liars and Thieves

In Part I, I wrote:
It turns out that the brain performs at least two essential functions while we are asleep: it purges liars (bad predictors) from sequence memory and eliminates thieves (redundant connections) from pattern memory.

Note: I will explain my choice of the liars and thieves metaphors in an upcoming post.
I am not going to go through a detailed explanation of how I arrived at my understanding because I would have to write a book. I probably will, eventually, but not today. I'm too busy. :-) Instead, let me just list one example to give you an idea of what I'm talking about. Here's where I got the idea for the liars and thieves metaphors:
Then said he unto me, This is the curse that goeth forth over the face of the whole earth: for every one that stealeth shall be cut off as on this side according to it; and every one that sweareth shall be cut off as on that side according to it.

I will bring it forth, saith the Lord of hosts, and it shall enter into the house of the thief, and into the house of him that sweareth falsely by my name: and it shall remain in the midst of his house, and shall consume it with the timber thereof and the stones thereof.

Source: Zechariah 5
After years of agonizing and thinking about it, I finally figured out, in the light of other passages, what the liars and thieves metaphors were all about. One day, some neurons in my cortex just made the right connections, and then, without warning, I understood what it all meant. Just like that.

A Little History on How It All Started

In December 1999, I made an amazing discovery, one that would transform my life. As a Christian, I had long suspected that some of the metaphorical passages in several Old and New Testament books were scientific in nature. I knew that, one day, their true meaning would be deciphered to reveal world-changing scientific knowledge. To me, it was inconceivable that Yahweh would go to great lengths to hide certain information for future advanced generations just for grins and giggles. I also felt that the various interpretations of the metaphors that I had seen in the Christian literature were not only stale and nonsensical but they did something even more reprehensible: they insulted Yahweh's intelligence with their banality.

At the time, I was researching artificial intelligence and, after many years of study, I had developed the beginning of a theory of intelligence. Things were starting to make sense a little and then I came face to face with the brick wall of my own ignorance. Progress just stopped. One evening, while reading the book of Revelation, it suddenly occurred to me that chapters 2 and 3 (known commonly as the letters to the seven churches of Asia) were a detailed metaphorical description of the organization and working of the left hemisphere of the brain. Color me crazy (I don't care) but I could see a clear correspondence between a couple of the metaphors and objects in my own AI model. I trembled with excitement.

Where others saw strangely worded admonitions to a handful of first-century Christian churches, I saw sensors, effectors, signals, hierarchies, patterns and sequences, perceptual learning, motor learning, success and failure, fitness criteria, etc. I knew nobody would believe it but, to me, the general meaning of the text was unmistakable. It just needed to fleshed out, that's all. Then my investigation of the message to the Church of Sardis soon led me to discover the book of Zechariah (The Lord Remembers), an amazing fountain of knowledge about memory organization, attention and pattern recognition. I was overwhelmed.

I remember the time I finally figured out that the two olive trees on both sides of the golden lampstand did not represent the trees of knowledge in the left and right hemispheres of the brain, as I had originally assumed. I had no real reason to make that assumption other than the observation that the brain consisted of two hemispheres. The analogy seemed to fit at the time but it was wrong. It took me years to realize I was hopelessly lost. I was forced to retrace my steps to that fork in the road and take the other path, the one that I had previously dismissed. It was like finding an opening out of the forest.

I could see clearly for the first time. I understood that the brain's memory consisted of two hierarchies, one for concurrent patterns (fig trees) and one for sequences (vine).
In that day, saith the Lord of hosts, shall ye call every man his neighbour under the vine and under the fig tree. Zechariah 3:10
I understood the importance and meaning of the Branch metaphor in Zechariah's text with respect to knowledge construction, attention and invariant recognition. I understood the metaphors of the flying scroll, the thieves and the liars. I understood why sequence memory was a hierarchy of seven-node chunks or building blocks. I understood what the brain was up to during sleep and why. I understood that the mainstream Bayesian approach to AI was deeply and fundamentally flawed, in spite of its successes. I knew why the brain went to great lengths to eliminate uncertainty.

I understood a lot of things. It felt sort of like what Mr. Anderson might have felt in the Matrix movie, when he wakes up from a virtual reality session and announced that he knew Kung-Fu.

Not long after that fateful evening in December 1999, my elation quickly faded when I realized that it was not going to be that easy to extract the full meaning of those ancient metaphors. I noticed that almost every word used in the metaphorical texts was pregnant with powerful and subtle meanings that were easily overlooked. I was forced to commit myself to carefully analyse everything in detail. I made slow progress over the years but I kept at it. Progress came in fits and starts but I think I have come a long way. I don't yet understand it all but it's just a matter of time before I do. Almost everything I wrote in this article regarding the brain came from my interpretation of the ancient metaphors.

The Future

Let me conclude by saying that I am not asking anybody to believe me. Whether you take it or leave it, is up to you. But do keep your ears and eyes open. It will get a lot more interesting in the not too distant future.

See Also

The Holy Grail of Robotics
Raiders of the Holy Grail
Jeff Hawkins Is Close to Something Big

Monday, December 30, 2013

Secrets of the Holy Grail, Part IV

Part I, II, III, IV, V

Abstract

In Part III, I described the hierarchical structure of sequence memory and I explained why patterns are the key to sequence learning. In this post, I explain invariant object recognition, the difference between short and long-term memory and how to catch a liar. But first, a word about remembering and forgetting.

Remembering and Forgetting

Unlike patterns which, once learned, remain permanently in memory, learned sequences are slowly forgotten, i.e., disconnected, unless they are repeated often. Repetition strengthens the connections between nodes in a sequence. If the strength of a connection reaches a predetermined threshold, it becomes permanent. There are two ways the connections in a sequence can be repeated, via sensory stimulation or internal playback. So even sequences that receive little sensory stimulation can become permanent if they are played back internally. The latter happens each time the brain focuses on a particular branch in the memory hierarchy.

In Part III, I wrote that the sequence learner starts with short intervals before moving on to progressively longer intervals. That is the first of the three fitness criteria used in sequence learning. Forgetting is the second. A node in a sequence under construction is often presented with multiple successor candidates. Initially, the learning system has no way of knowing which of the candidates are legit, if any. One way to eliminate bad candidates is to slowly forget them. A node cannot survive unless it is frequently reinforced via sequence repetition. But this raises a serious question, what happens to infrequent sequences or to sequences that never repeat? The answer is that they must be repeated (replayed) internally in order to be retained. However, the only way to really be sure that they are good or bad is to use the third fitness criterion: find out if they lead to a contradiction (see Catching Liars below).

Invariant Object Recognition

Hold your hand in front of your eyes and slowly rotate your wrist. As you do so, your visual cortex is presented with a sequence of images. Even though each successive image is different from the others, your brain does not think of each image as representing a different object. Somewhere, in your cortex, you still know that you are looking at your hand regardless of its orientation or distance from your eyes. This is called invariant object recognition, probably the most important perceptual capability of the brain. It holds the key to understanding several other aspects of perception such as attention and short-term memory.

Catching Liars

When you rotate your hand, your brain sees the successive patterns as one object because they are linked together within a single package called a branch. The branch is a bundle of linked pattern sequences. But how does the brain know that one sequence should be tied to another in order to form a bundle? The answer turns out to be rather simple: if two sequences have two or more nodes in common, the branch mechanism automatically links them together. The problem is that any of the shared nodes may have been acquired in error. How can we tell? The answer lies in the timing between shared nodes. If two sequences belong to the same branch, their predictions must match. If there is a mismatch, one of the nodes is a liar and must be discarded. Consider the two sequences in the diagram below.
The arrows represent the direction of pattern activations. The red circles are normal sequence nodes and the green circles symbolize shared nodes. The total recorded interval between the two shared nodes must be the same is both sequences. If not, there is a contradiction and the culprit is eliminated. If there is agreement, it means the two sequences represent different facets of the same object and thus belong to the same invariant branch.

In human and animal brains, testing for sequence contradictions occurs during sleep. The reason is that the sequences must be replayed internally during the test and this would disrupt the brain's normal activity. I believe that catching thieves (see Part II) and liars is what is happening during so-called REM sleep.

Short-Term Memory

Psychologists and neurologists have maintained in the past that memory is divided into two separate areas, one for long-term memories and one for short-term memories. My position is that there is only one memory structure for both short and long-term memories. Short-term memory is merely the currently activated branch, i.e., the one under attention. The brain can focus on only one thing at a time, that is to say, only one branch in the sequence hierarchy can be activated at a time, a phenomenon that magicians and pickpockets have exploited over the years. Furthermore, a branch can only be active for up to about 12 seconds at a time, after which attention must be switched to another branch.

Coming Up

As I promised in Part I, in my next post, I will reveal where I got my knowledge of the brain's organization.

See Also

The Holy Grail of Robotics
Raiders of the Holy Grail
Jeff Hawkins Is Close to Something Big

Tuesday, December 24, 2013

Secrets of the Holy Grail, Part III

Part I, II, III, IV, V

Sorry for the long hiatus. I posted the last installment of this multi-part article way back in March. I have been meaning to write some more on the topic but a series of unfortunate events in my life have slowed me down a bit.

Abstract

In Part II, I explained how to do pattern learning and how to prevent patterns in the hierarchy from getting bigger than they need to be. In today's post, I explain sequence learning and the organization of sequence memory. However, be advised that there are a couple of things about sequence memory that I want to keep secret for the time being.

A Few Observations

A sequence is a string of consecutive nodes representing successive pattern detections. Sequence memory is organized hierarchically, like a tree. A sequence is divided into seven-node sections although sections at either end of a sequence may have less than seven nodes. These are the building blocks of memory. Why seven nodes per sequence? It's because this is the capacity of short term memory. But regardless of its level in the hierarchy, a block is ultimately a sequence of patterns.

Sequences have many functions and any sequence learning mechanism should take the following into consideration:
  • A sequence is used to detect a unique transformation in the sensory space which is manifested as a number of consecutive pattern detections.
  • A sequence is a recording mechanism. It records a memory trace, that is, the precise timing of its last activation.
  • A sequence is a predictive mechanism. The firing of a single node in a sequence is enough to predict the firing of subsequent nodes.
  • A sequence can be used for pattern completion and fault tolerance. The firing of a node in the sequence is enough to compensate for missing signals. This is important when working with noisy and imperfect sensory streams.
  • Sequences, together with the branch mechanism (see Part IV), are part of the invariant recognition mechanism of an intelligent system.
  • A sequence is a sensory motor unit, i.e., an integral part of the goal-oriented behavior mechanism of an intelligent system.
  • The temporal interval between any two consecutive pattern signals can vary.
  • Some sequences repeat more slowly than others. Indeed, many sequences will occur only once or twice.
  • Several sequences can share one or more patterns. This is used to join otherwise unrelated sequences together and is part of the mechanism of invariant recognition.
The Power of Patterns

This may sound counterintuitive but patterns (see Part II for a description of patterns) are the key to sequence learning. This is because patterns are inherently predictive. Patterns are so unique that they normally have just a few predecessors and successors. Most patterns will have only one predecessor and one successor. This is important because it dictates a crucial aspect of sequence learning. Unlike pattern learning, which requires many frequent repetitions, the learning of a sequence (predecessor-successor) needs only one example. In other words, sequence learning can be extremely fast.

Dealing with Imperfection

How can an intelligent system learn a sequence if pattern signals do not always arrive on time? In my opinion, it does not matter that the signals are imperfect as long as they arrive on time some of the time. A single instance of two consecutive signals is sufficient to learn a new sequence. Sequences that lead to a contradiction (I'll explain this in Part IV) or that are not reinforced over time are simply discarded.

Small Things First

One of the main problems we face in sequence learning is that the interval between any two consecutive pattern signals is a variable. It can change with circumstances. For example, the notes of a song can be fast or slow but it's still the same song. It is a problem because it makes it almost impossible to determine which pattern precedes which. The solution turns out to be rather simple: the learning system should start with the smallest intervals first before slowly moving on to progressively longer intervals.

Coming up

In Part IV, I will go over the mechanisms of invariant recognition and short term memory. I will also explain how to catch a liar, i.e., how to detect contradictions in sequence memory.

See Also

The Holy Grail of Robotics
Raiders of the Holy Grail
Jeff Hawkins Is Close to Something Big

Tuesday, November 19, 2013

Did OSU Researchers Solve the Cocktail Party Problem?

Potential Speech Recognition Breakthrough?

There is incredible news coming out of Ohio State University. Speech recognition researchers claim to be able to pick out a particular speech sound from a sound track mixed with random noise and/or other speech sounds using a deep neural network. I have my doubts about the claims but, if this is true, it would mean that they have solved a major part of the perceptual learning puzzle: the ability to focus on one thing while ignoring others. If true, it would be the biggest single breakthrough in the history of artificial intelligence research, in my opinion. That being said, I will reserve judgement until I know more about the details of the research.

Wednesday, November 6, 2013

Sequences and Analogies, Part II

Part I, II

Abstract

Previously, I wrote about Professor Douglas Hofstadter's claim that analogy is at the core of human intelligence. Although I agree with Hofstadter to a certain extent, I faulted him for focusing on language and symbols instead of investigating the root of the phenomenon. In this post, I argue that there is something that is even more fundamental than analogy: the sequence. I further argue that the neuronal mechanism responsible for discovering analogies can only work while the brain is asleep.

Temporal Proportions

Analogy comes from the Greek ἀναλογία (analogia), a word that means 'proportion' or 'proportionality'. When we think of proportionality, we usually think of geometry. For example, we can say that two triangles of different sizes are analogous if their sides are proportional. I think the ancient Greeks hit the nail right on the head. In this article, I defend the hypothesis that the mechanism used by the brain to make analogies is also based on proportionality. However, in the brain, the proportions are not measured in units of length but units of time.

Proportional Sequences

One of the curious things about analogies is that we have no recollection of learning them. They seem to suddenly materialize into our consciousness out of nowhere. We can safely surmise that the process of recognizing and establishing analogies is automatic. We can further assume that memory is organized in such a way as to make it easy for a neural mechanism to examine two chunks of knowledge and determine whether or not they are analogous, i.e., proportional.

Unlike Hofstadter, I believe that the primary function of intelligence is to make predictions, not analogies. It just so happens that predictions are possible only if recorded events are stored in sequences of patterns. The chunking ability of memory that Hofstadter is so fond of is simply the result of its hierarchical structure. Each branch of the hierarchy corresponds to a chunk. In the Rebel Science model of intelligence, a chunk is a single sequence of up to seven nodes. A node can be either another chunk or a sensory pattern. Sequences do not have fixed timing. The temporal intervals between the nodes in a sequence can vary but their proportions are invariant.
In the figure above, a yellow-filled circle represents a sequence or chunk of knowledge. I hypothesize that the brain can replay the sequences during sleep and determine whether or not any two sequences are proportional. If they are, a link is established between the two to form an analogy. The reason that analogy discovery must be done during sleep is that the sequences must be played back and that would disrupt normal brain function and behavior. During waking hours, the activation or recall of one sequence triggers the recall of another if they are analogous. I believe that all types of analogies are based on sequence timing comparisons.

Note: I am hard at work incorporating all of these principles into the Rebel Science Speech program. I hope to release a demo as soon as it is ready. Hang in there.

Monday, November 4, 2013

Sequences and Analogies, Part I

Part I, II

Abstract

How does the brain solve the problem of recognizing a musical tune or an utterance irrespective of its amplitude, tempo or overall pitch? How does it determine that a pony is a type of horse, that dogs, coyotes and wolves are related, or that going on a wild goose chase is not really about running after a bird? These are important questions because they have to do with one of the most fundamental and essential aspects of intelligence: the brain's ability to draw analogies. Without it, we would have a hard time understanding each other or the world around us. In this two-part article, I will argue that the secret to making analogies lies in the timing of sequences and the ability to make predictions.

Analogies, Hofstadter and Predictions

Professor Douglas Hofstadter is the College of Arts and Sciences Distinguished Professor of Cognitive Science and Computer Science at the University of Indiana, Bloomington. Hofstadter became famous after the publication in 1979 of his book, Gödel, Escher, Bach: An Eternal Golden Braid, for which he received a Pulitzer Prize. He is one of those elitist academics who have developed a fanatical faith in materialism and naturalism. In other words, he believes that complex life somehow sprang from dirt and evolved into bats and whales all by itself and that consciousness is an emergent property of the brain. It never ceases to amaze me that some of the most brilliant and knowledgeable people on earth can be so full of crap at the same time. From my perspective, it is fitting that true AI, the kind that materialists like Hofstadter are after, will come from the one place that they least expect.

This is not to say that Hofstadter has nothing interesting to say. Far from it. Like I said, the man is brilliant, one of those highly educated intellectuals one never gets tired of listening to, whether or not one agrees with what they have to say. The reason that I bring this up has to do with analogies. Hofstadter spent the last four decades preaching to everyone who was willing to listen that the ability to detect analogies is at the core of human intelligence. He makes a pretty convincing case. Of course, analogies are not the be-all of intelligence. There would be no intelligence without the ability to learn patterns and sequences, to make predictions, to develop motor skills, to seek goals and adapt to one's environment. Hofstadter is aware of all this, I am sure, but he has chosen to focus only on aspects of intelligence that involve analogies. For example, he speculates about how thoughts in memory are organized in chunks (what I and others are calling a hierarchical structure), and how a single conscious thought can awaken another and make it available to introspection in working memory.

Naturally, one wonders why nothing spectacular has emerged from the decades Hofstadter and his students have spent experimenting with analogies. Hofstadter has only himself to blame, in my opinion. He perfectly understands that analogies are at the core of the organization of memory but, instead of digging deep to the root of the phenomenon, he spends most of his time and energy playing with pictograms, languages and words, i.e., with symbols. It is all very entertaining but it is an altogether too high a level of abstraction. Surely the ability to recognize an analogy is much more fundamental than the high level manipulation of symbols. The only way to explain how the brain makes analogies is to come up with a biologically plausible mechanism that is universally applicable to all types of analogies. By focusing on language, Hofstadter has locked himself into the same box as the symbol manipulation proponents of the last century and the Bayesian Brain fanatics of this century. That's too bad.

Next Up

Strangely, it never occurs to Hofstadter that analogies are possible, not because the brain is designed specifically to discover them, but because the brain is designed and organized to make something that is even more fundamental than analogies: predictions. This will be the topic of my next post.

Tuesday, October 29, 2013

Did Vicarious Achieve an AI Breakthrough?

Did Vicarious Solve the Cocktail Party Problem?

There is something a little weird about Vicarious's recent announcement in which they claim to have developed a machine learning program that can solve CAPTCHAs at 90% accuracy. What is interesting, from my vantage point, is that some CAPTCHAs print irrelevant words behind the actual text to be recognized. Paypal, for example, displays copies of the word 'PayPal' behind the CAPTCHA text. Take a look at this CAPTCHA, which Vicarious claims to be able to solve.
Quick and dirty techniques such as thresholding can be used to reduce or eliminate background noise but that would be cheating. Is this what Vicarious is using here? I don't know but it would seem that thresholding would not work very well in this case unless you knew in advance what to look for. This means that, if we are to believe their claim, Vicarious's AI program is sophisticated enough to be able to focus on certain things while ignoring others. Wow. Really? Are we to understand that Vicarious solved a visual analog of the cocktail party problem, which is essentially the ability to pay attention to one object within a group of many others? If the answer is yes, it would be a monumental breakthrough because this is one of the hardest unsolved problems in computer science. Even so, the question becomes, how can the program tell which letters in the picture are relevant and which are not? There is something either fishy or missing in this story.

See Also:

Vicarious Has a Winner

Vicarious Wakes up with a Bang

It has been a while since we heard anything from Vicarious. I was beginning to wonder if the company had fallen asleep. Then suddenly out of nowhere, they announced that they have a machine learning program that can solve CAPTCHAs, the sometimes hard to read letter puzzles that are meant to ward off those pesky computer bots. That's a rather dramatic awakening, I would say. Although I do not agree with Vicarious's Bayesian or probabilistic approach to AI, I have to admit that this is very impressive.

A Few Observations

There are a few things about this new development that intrigue me. First of all, why didn't Vicarious host an online demo somewhere in the cloud and release a free app that others can use to test their claim? How hard would that be? It would have added some meat to the sauce, so to speak. Second, and this is more a question than a criticism, my understanding is that the recursive cortical network (RCN) works best with moving pictures. It is hard to imagine how it learns using static pictures. Third, Vicarious's CEO, D. Scott Phoenix, claimed that RCN needs less than 10 training examples per letter whereas other visual recognition programs require thousands of examples. This is truly amazing and, if true, it tells me that they must have figured out an efficient way to do invariant pattern recognition.

Why I Still Don't Think Vicarious Is on the Right Track

Yes, I still think that the Bayesian approach to AI is a red herring. Vicarious's CTO and co-founder, Dileep George, is convinced that intelligence is based on probabilistic math. I believe that neither human nor animal intelligence uses probability for reasoning, prediction or planning. We are cause/effect thinkers, not probability thinkers. The brain has a fast and effective way of compensating for the uncertain or probabilistic nature of the sensory stream by filling in any missing information and filtering out the noise. I see essentially two competing models. The Bayesian model assumes that the world is inherently uncertain and that the job of an intelligent system is to calculate the probabilities. The Rebel Science model, by contrast, assumes that the world is perfect and that the job of the intelligent system is to discover this perfection.

In Secrets of the Holy Grail, I wrote, "nobody can rightfully claim to understand the brain’s perceptual learning mechanism without also knowing exactly what the brain does during sleep and why." I'll say it again. If the guys at Vicarious don't know why the brain's neural network needs sleep, then they are not doing it right.

See Also:

Did Vicarious Achieve an AI Breakthrough
The Second Great Red Herring Chase
Vicarious Systems' Disappointing Singularity Summit Talk
The Myth of the Bayesian Brain
The Perfect Brain: Another Nail in the Coffin of the Bayesian Brain

Friday, October 11, 2013

The Problem with Speech Recognition Models

Abstract

I recently read a paper (by way of PhysOrg) about a new speech recognition model called the Hidden Conditional Neural Fields for Continuous Phoneme Speech Recognition. The good news is that HCNF outperforms existing models. The bad news is that it does not come close to solving any of the pressing problems that plague automatic speech recognition. Problems like noise intolerance and the inability to focus on one speaker in a roomful of speakers continue to vex the best experts in the field. While reading the paper, it occurred to me that the main problem with speech recognition models is that they exist in the first place. I will argue that the very fact that we have such models is at the root of the problem. Let me explain.

Not Good Enough

Wouldn't it be nice if we could talk directly to our television sets? We should be able to say things like, "turn the volume down a little" or "please record the next episode of Game of Thrones". Indeed, why is it that we can talk to our smartphones but not to our TVs? The reason is this. Current speech recognizers are pretty much useless in the presence of noise or multiple voices speaking at the same time. The sounds coming from the TV alone would confuse any state of the art recognizer. Sure, we could turn the remote control device into a noise reduction microphone and hold it close to the mouth when speaking but that would defeat the purpose of having a hands-free and intuitive way of interacting with our TVs. What we need is a recognizer that can focus on one or two voices in the room while ignoring everything else including other voices, e.g., from children, pets or guests, or from the TV. A good TV speech recognizer should respond only to the voices of those it was instructed to pay attention to. It should also ignore any conversation that does not concern its function. Unfortunately, these capabilities are way beyond what current speech recognition technology can achieve and there are no solutions in sight.

Limited Domain

I am arguing that speech recognition models are not up to the task simply because they are limited domain models, i.e., they only work with speech. But why shouldn't they, you ask? It is because the brain does not use a different representation or learning scheme for different types of knowledge. To the brain, knowledge is knowledge, regardless of whether its origin is auditory, tactile or visual. It does not matter whether it has to do with language, music, pictures, food, houses, trees, cats or what have you. The cortical mechanism that lets you recognize your grandmother's face is not structurally different than the one that lets you recognize your grandmother's name. A good speech recognition model should be able to learn to recognize any type of sensory data, not just speech. It should also be able to recognize multiple languages, not just one. And why not? If the human brain can do it, a computer program can do it too, right? After all, it is just a neural mechanism. However, as such, the model would no longer be a speech recognition model but a general perceptual learning model.

There Is a Pattern to the Madness

The brain learns by finding patterns in the stream of signals that it continually receives from its sensors. The origin of the signals does not matter because a signal arriving from an audio sensor is no different than a signal arriving from a light detector. It is just a transient pulse, a temporal marker that signifies that something just happened. This begs the question, how can the brain use the same model to learn different types of knowledge? In other words, how does the brain extract knowledge from a stream of unlabeled sensory pulses? The answer lies in the observation that sensory signals do not occur randomly. There is a pattern to the madness. In fact, there are millions of patterns in the brain's sensory stream. The key to learning them all has to do with timing. That is, sensory signals can be grouped and categorized according to their temporal relationships. It turns out that signals can have only two types of temporal relationships; they can be either concurrent or sequential. The learning mechanism of the brain is designed to discover those relationships and recognize them every time they occur. This is the basis of all learning and knowledge.

The Holy Grail of Perceptual Learning

Many in the business assume that the cocktail party problem is relevant only to speech recognition. In reality, it is a problem that must be solved for every type of sensory phenomena, not just speech sounds. Humans and animals do it continually when they shift their attention from one object to another. The brain's ability to pay attention to one thing at a time is the holy grail of perceptual learning. In conclusion, let me reiterate that we don't need different models for visual and speech recognition. We need only one perceptual learning model for everything.

PS. I am continuing to write code for the Rebel Speech recognizer and incorporating the principles of perceptual learning that I have written about on this blog. I am making steady progress and I will post a demo executable as soon as it is ready. Hang in there.

Thursday, August 29, 2013

The Other Facet of Motor Learning, Part II

Part I, II

Abstract

Previously, I wrote that motor learning is a trial and error process which consists of making random motor output connections and testing them for fitness. There are two fitness criteria used in motor learning. First, a motor connection must not achieve goals other than the one it is associated with. Second, a motor connection must not cause conflicts with other connections on the same motor neuron. In this post, I explain what a motor conflict is within the context of motor control.

Motor Control

Motor control is a process whereby the brain's cortex sends motor commands to motor neurons in order to effect goal-directed, coordinated behavior. Motor commands are discrete signals of which there are two types, start and stop. These correspond roughly to the excitatory and inhibitory signals that are observed in the motor systems of humans and animals. A start command starts an action (such as contracting a muscle) while a stop command stops an action already in progress.
In the illustration above, the red-filled circle represents the neuron's axonic output to the muscle, the black circle is a stop input synapse and the white circles are start input synapses. Note that the diagram is only meant to illustrate the principles of motor control, not the actual neural mechanisms which can be complicated. The brain has a highly complex structure dedicated to it. It's called the basal ganglia. For our purposes, think of a motor neuron as a little motor that can be turned on or off by multiple pushbutton switches attached to it. When the neuron is on, it continually fires which causes its target muscle to stay contracted. When the neuron is off, it stops firing, allowing the muscle to relax.

An important question is, how does the motor system control the amount of force exerted by the muscles? The answer is that every motor action is serviced by a set of motor neurons and each neuron in the set is pre-wired to exert a different force magnitude. It is up to the motor learning system to determine which ones to activate for a given goal or situation.

Conflict Detection

Since motor neurons are shared by a large number of cortical programs, it is important that the command signals are not in conflict. There are two principles that govern motor coordination.

1. A motor neuron must not receive more than one command at a time.
2. No action can be started if it is already started or stopped if it is already stopped.

During motor learning, motor connections are monitored to see if they violate the coordination rules. Any violation results in the weakening of the conflicting connections. In the end, only the strongest connections survive.

Analysis

The principles of goal-directed motor learning are simple, powerful and can be easily implemented in a computer. However, they are of no use unless motor signals can be generated in an orderly, predictable and coherent manner. In other words, an intelligent system must have a sophisticated perceptual mechanism before it can even begin to behave sensibly to achieve its goals.

The topic of goals brings us to the notion of motivation. How can an intelligent system have goals unless it is somehow motivated to pursue those goals. Motivation is a huge part of intelligence. In fact, there can be no real intelligence without it. But how can a machine have likes and dislikes? In an upcoming article, I will dig deep into this topic and the importance of motivation to adaptation and survival.

See Also:

The Holy Grail of Robotics
Goal Oriented Motor Learning
Raiders of the Holy Grail
Secrets of the Holy Grail

Tuesday, August 27, 2013

The Other Facet of Motor Learning, Part I

Part I, II

Idolatry

In a previous two-part article titled Goal-Oriented Motor Learning, I wrote the following:
The Two Facets of Motor Learning

I will not go into how the brain learns sensory patterns in this article. I will, one day, but not today. What I will explain in this article is one facet of motor learning, the one that leads to goal-seeking behavior. There is another facet that has to do with eliminating motor conflicts. That, too, will have to await a future article. I just want to explain how the brain finds the right motor connections for goal-seeking behavior. As I wrote previously, I get my understanding of the brain by consulting an ancient oracle (no, I was not joking) and interpreting its message the best I can. Here's what the oracle says about goal-oriented motor learning:

Notwithstanding I have a few things against thee, because thou sufferest that woman Jezebel, which calleth herself a prophetess, to teach and to seduce my servants to commit fornication, and to eat things sacrificed unto idols.
I always burst out laughing every time I read this verse in the book of Revelation. I laugh, not just because I think the wording of the verse is hilarious, but because the choice of metaphors is so exquisitely brilliant. Jezebel is a metaphor for a predictive mechanism. This is why she is called a prophetess. That's an easy one to interpret. But the mechanism is not a good predictor because, during its attempt to achieve various goals, it causes bad things to happen along the way: fornication and idolatry. Fornication is the oracle's metaphor for making connections that cause motor conflicts, a topic for a future article. Idolatry (the worshiping or serving of other gods) symbolizes the making of connections that lead to the wrong goals. Obviously, neither fornication nor idolatry will be tolerated. :-D
I went on to explain that a goal is a pattern and that goal-oriented motor learning consists of choosing the right connections between pattern-associated motor neurons and the target motor effectors. During this trial and error learning process, the idolaters, i.e., the motor connections that fail to satisfy their associated patterns, are simply disconnected. But there is a little bit more to motor learning than just achieving goals. Below, I explain the other facet of motor learning, the one that the oracle metaphorically calls fornication.

Fornication

Fornication, as used in the Bible, is the act of having sex with a woman who is someone else's wife. It's a conflicting act. In the quoted passage above, the oracle is using fornication as a metaphor for motor conflicts in the brain. Motor conflicts are a problem because there is a fixed number of actuators (i.e., muscles) that must be shared by a huge and growing number of competing sensorimotor entities (neural programs) in the cortex. Obviously, if two or more entities try to use the same actuator simultaneously, conflicts will arise which will cripple any kind of goal-directed behavior.

The Simplicity and Power of Motor Learning

The motor learning system I described above uses a trial and error process to find appropriate motor connections. New connections are made randomly and an error detection mechanism is used to weed out idolaters and fornicators. This system ensures effective and smooth motor coordination. Robots will use it to learn sophisticated motor behavior such as walking, driving, speaking, etc. The simplicity of it all is unnerving, some would say, but its power is in the simplicity. In Part II, I will go into the details of conflict detection and elimination. Coming soon.

See Also:

The Holy Grail of Robotics
Goal Oriented Motor Learning
Raiders of the Holy Grail
Secrets of the Holy Grail

Monday, August 26, 2013

Why Speech Recognition Falls Short

It's Not the Way the Brain Does it

Current speech recognition technology, while impressive, falls short of delivering on the promise of human-like performance. The biggest problem is that speech recognizers are sensitive to noise which makes them pretty much useless if there are several voices speaking at the same time. The reason, of course, is that they do not work like the human brain. We humans have no trouble listening to a friend in a noisy restaurant because, unlike speech recognizers, we have the ability to focus our attention on one voice at a time and we can change our focus in an instant, if we wish. The human brain can also easily adapt to a given situation. A new voice may have an unfamiliar foreign accent but the brain can quickly learn its peculiarities and do a good job at recognizing what is being said.

A New Approach, Rebel Speech

The main reason that current technology falls short is that speech recognizers, unlike the brain, do not learn to recognize speech. They are hand-programmed. In other words, their knowledge (phones, diphones, senones, syllables, words and other speech patterns) is painstakingly compiled and coded by a programmer. This approach, while effective to an extent, is forever doomed to be incomplete. There are important subtleties in speech sounds that can only be detected via direct learning. If we are to make any significant progress in computer speech recognition, then learning and paying attention are key capabilities that we must incorporate into our future recognizers. My hope is that its autonomous ability to learn and to focus its attention on a given voice is what will set Rebel Speech apart from the rest.

See Also:

The Holy Grail of Robotics
Goal Oriented Motor Learning
Raiders of the Holy Grail
Secrets of the Holy Grail
The Myth of the Bayesian Brain

Sunday, August 25, 2013

The Rebel Science Speech Recognition Project

Soon, I plan to go back to work on my ongoing speech recognition project, Rebel Speech [pdf]. The goal of the project is not so much to design a better speech recognition program from scratch but to demonstrate the superiority of the Rebel Science approach to artificial intelligence. Speech recognition will be just one of many applications that use the Rebel Science Intelligence Engine. The latter will serve as the future backbone for any type of AI research project that requires perceptual and motor learning. The key word is learning, which includes adaptation. My ultimate goal is to use it in the design of a brain for a complex autonomous bipedal or quadrupedal robot with many types of sensors and degrees of freedom.

Those of you who are interested in this topic can brush up on Rebel Science AI by clicking on the following links:

The Holy Grail of Robotics
Goal Oriented Motor Learning
Raiders of the Holy Grail
Secrets of the Holy Grail
The Myth of the Bayesian Brain

I have also created a new forum to start discussing AI and its consequences.

Tuesday, August 20, 2013

ALS, Be Not Proud

Guillermina, my wife, passed away this morning from complications due to ALS. The disease that ravaged her body for over nine years finally won. She fought hard but, in the end, her lungs just stopped working. It is particularly sad because she knew before she died that an effective drug for ALS was just sitting on the hospital shelf, out of reach. A part of me just died but I will continue to fight this horrible disease in her honor.

Tuesday, August 6, 2013

What Really Cured Ted Harada?

Note 1: Mr. Harada correctly pointed out on Twitter that he is not cured. I concur and apologize but it's too late for me to change the title. This does not take away from the spirit of the original message. (8/28/13)

Note 2: This article was written more than a year ago and things have changed drastically. My wife succumbed to ALS almost a year ago. After much study, I have now concluded that it was a combination of the anesthetics and anti-inflammatory drug (dexamethasone) that she received during back surgery years ago that contributed to a near miraculous recovery that lasted almost a month. I would recommend that anybody with ALS take some kind of anti-inflammatory medicine or supplement. Even a non-prescription, over-the-counter drug like Naproxen is likely to help much. Read Anesthetics and Glucocorticoids for ALS for more on this topic. (7/21/2014)

This is a question that must be asked because nobody seems to really know for sure. Ted Harada, an ALS patient, experienced a miraculous recovery after undergoing stem cell treatments in 2011 and 2012 as part of an FDA-approved trial. Neuralstem Inc., the company that derived the stem cells from the spinal cord tissue of a fetus, wasted no time in capitalizing on the apparent success of their technology. Ted Harada became an overnight celebrity.

But not everybody is convinced that Ted Harada got well as a result of the stem cell injections. Some people pointed out that Harada's improvements occurred too quickly after the procedures. The neurotrophic factors in the stem cells do not work that fast. Others noted that his improvements did not correspond to the areas of his spinal cord that received the stem cells. Dr. Angela Genge of the Montréal Neurological Institute and Hospital, expressed doubts according to a January Alzforum article, "suggesting that any benefit might have resulted from the immunosuppressant drugs the participants received, that is, their ability to quell neuroinflammatory pathology."

Dr. Genge had, so to speak, thrown a huge fly in Neuralstem's ointment, so much so that, in May of this year, Neuralstem CEO, Richard Garr was forced to make an interesting admission on his blog:
It is also possible that an “unknown unknown” is responsible for Ted’s long term improvement and the stabilization of the other patients. The argument here is that just because WE can’t figure out what else it might possibly be, doesn't mean there isn't another explanation. However unlikely we feel this could be, it is why large, well-controlled trials are always required and justified. We need to continue and enlarge our clinical trials to refute this argument.
Indeed, in July, Dr. Jonathan Glass and Dr. Christina Fournier of the Emory ALS Center announced plans for a new study in order to eliminate the possibility that Ted Harada might have been cured by the immunosuppressants (anti-rejection drugs) that he received as part of the stem cell procedures. But those of us who have followed ALS research over the years know that it's a useless trial because the outcome is already known: immunosuppressing drugs have already been shown to be ineffective against ALS. So why the study? In my opinion, it's really Mr. Garr's way of calming the fears of his company's investors. Garr plans to wrestle that straw man to the ground and declare victory. He'll be able to triumphantly announce at the next shareholders meeting, "You see, we told you it was our stem cells that cured Ted Harada."

If I were a Neuralstem investor, this is where I would raise my hand and ask, "Uh, are you really sure about that?" I mean, did not Ted Harada receive another powerful immune suppressing drug that has not been tested in this latest study? But of course, he did. Mr. Harada was anesthetized for 5 + hours during each procedure. In other words, he received a massive dose of anesthetics in order to keep him completely immobilized during the delicate operation. Most anesthetics have powerful anti-inflammatory properties. Several ALS patients have asked Mr. Harada to reveal the type of anesthetic he received but he declined to do so. No matter. We can guess that it was probably sevoflurane, isoflurane or a similar volatile anesthetic. Why? Only because these are the anesthetics of choice used to keep a patient perfectly immobilized. So why did those in charge of the new study omit the anesthetics from the list of immunosuppressing drugs to be tested? I am asking because both Glass and Fournier were aware of reports that some ALS patients are seeing improvements after undergoing anesthesia. What's up with that? Inquiring shareholders and all that.

Something smells fishy at the Emory ALS Center in Atlanta, Georgia. One wonders what CEO Richard Garr has to say about all this. Join the discussion.

See Also:

Anesthetics and Glucocorticoids for ALS
Treat ALS with Anti-Inflammatory Drugs

Thursday, August 1, 2013

The ALS/Anesthetics Hypothesis

Note: This hypothesis has been revised. Please read Anesthetics and Glucocorticoids for ALS for the latest.
Abstract

Amyotrophic lateral sclerosis is a rare and fatal neurodegenerative disease that strikes mostly older adults. This hypothesis is based on the finding that ALS is primarily an immune system disorder. Researchers have identified an elevated inflammatory response in ALS patients that is manifested during both presymptomatic and later stages of the disease. This inflammation is thought to be responsible for its rapid progression. Identifying the cause and nature of the inflammatory response is the key to formulating an effective therapy. Even though researchers have known about the innate immune response in ALS patients for more than a decade, attempts at using traditional anti-inflammatory drugs have not been very successful. The reason is that researchers have not yet identified the cause of the inflammation. There are many types of inflammations and many types of anti-inflammatory substances. By identifying the exact cause of ALS inflammation, we can formulate an effective therapy to eliminate it. We believe that eliminating the cause will not only stop progression, but will also bring the disease into full remission, short of regenerating dead motor neurons. We hypothesize that certain anesthetics such as propofol and sevoflurane can fully eliminate the cause of the inflammation.

The Cause of ALS Inflammation

ALS inflammation is caused by a deficiency in certain neurotransmitter receptors (or neuroreceptors), primarily the GABA-A alpha-1 and glycine alpha-1 receptors. A deficiency means that the receptors lack their normal affinity for their neurotransmitters and, as a result, fail to activate properly. These receptors are used extensively by the inhibitory neurons in the brain stem and spinal cord to control the activation of motor neurons. A deficiency causes an abnormal increase in the activity of the motor neurons and this, in turn, leads to a pathological condition known as neuronal excitotoxicity. But, and this is the crux of this argument, the same receptors are also used by innate immune system cells such as monocytes. If monocytic receptors are functioning normally, they respond to the normal level of neurotransmitters in the cerebrospinal fluid and this inhibits the activity of the monocytes. During an infection, messenger immune molecules are used to block the receptors. The ensuing decrease in inhibition activates the monocytes in order to fight the infection. However, if the receptors are deficient, the monocytes are no longer properly inhibited and the result is the chronic and destructive inflammatory response we observe in ALS patients.

The Therapy

An effective ALS therapy must not only suppress ALS inflammation, it must also eliminate the cause. Doing so will kill two birds with one stone because it eliminates neuronal excitotoxicity as well. It just so happens that certain anesthetics, such as propofol and sevoflurane, can potentiate all the known deficient receptors in ALS patients. Potentiation is the key. It consists of increasing a receptor's affinity for its neurotransmitter, restoring it to its normal functioning level. But what sets these anesthetics apart is that the induced potentiation does not disappear after the drug is eliminated from the body. It can last for days and even weeks. Part of this hypothesis is that, by fully eliminating the chronic inflammatory response, the disease can be put into full remission.

Experimental Confirmation

Although no official trials have been conducted to test this hypothesis, at least a dozen patients have reported significant and, at times, spectacular improvements in their symptoms after undergoing anesthesia with the anesthetics propofol and sevoflurane. Based on their reports, we can deduce a number of therapeutic principles. The optimum propofol dose seems to be about 800 mg. Anything below 200 mg does not seem to be very effective. We also have good reasons to believe that a mixture of propofol and sevoflurane is much more effective than propofol alone.

Tuesday, June 18, 2013

ALS Update

Note (8/21/14): See my latest thinking on ALS: Anesthetics and Glucocorticoids for ALS.
Alright, here's what's going on from the ALS front line. Based on my research and my wife's experience with this awful disease over the years, I figured out that certain anesthetics, such as propofol, sevoflurane and halothane, can reverse ALS symptoms within hours after anesthesia. The effect can be spectacular. It turned out that this was the easy part. The hard part is to convince the ALS research cottage industry to take these anesthetics seriously. So far, up to five ALS patients have reported significant improvements after undergoing unrelated medical procedures that required anesthesia.

Many in the ALS community have figured out that the ALS research industry is not interested in finding a treatment. Their only goal is to continue to raise as much money as they can while making sure that a cure is never found. The pharmaceutical industry is only interested in trying out proprietary drugs which they can legally use to exclude rivals from the market while they're raping an unsuspecting public with exorbitant prices. Unfortunately for ALS sufferers, most of the patents on anesthetics have expired. It's inhuman and sickening. If I could afford it, I would pay a reputed lab to conduct a human efficacy trial but that is just a dream. Better yet, I would finance several medical clinics outside the country and immediately offer the therapy to those who can make the trip. Patients would pay only if they see improvements.

I could kick myself in the rear end for not having figured this out much sooner. All the evidence I needed was there. Years ago, my wife could still walk. I would have contacted some anesthesiologist in some other country and pay for the treatment. Unfortunately, my wife is currently bed ridden, under 24-hour nursing care and forced to use a ventilator for breathing. Our only hope is that a significant number of ALS patients in less strict countries can get the treatment. Their success stories would hopefully make the evening news and force the FDA and other health authorities to do something. In the meantime, because of the callous indifference on the part of those who are paid good money to find a treatment for this disease, thousands of people continue to die a horrible death.

See Also:

Anesthetics and Glucocorticoids for ALS
Treat ALS with Anti-Inflammatory Drugs
The Evil Lie about ALS

Saturday, May 4, 2013

Excuses, Excuses

One of the reasons that I cannot wrap my head around artificial intelligence at this time is that I am currently heavily involved in trying to find a cure or an effective treatment for my wife who suffers from ALS, also known as Lou Gehrig's disease. The good news is that I am now convinced that we are on the verge of a genuine breakthrough, not just for ALS but for other neurodegenerative diseases as well. The bad news is that AI will have to wait a little longer. One thing at a time. Hang in there.

Saturday, April 27, 2013

Darn It!

What the Hell Is Wrong with Me?

The last couple of weeks, I've been feeling like I'm in a daze. I can't bring myself to make any important decision, as if something alien had taken a hold of me and slowed my brain to a crawl. I've been meaning to post the last two installments of my article, Secrets of the Holy Grail, but I can't bring myself to do it. It feels like I'm not in charge of my own free will, as if I have been drugged or something. My ears are ringing all the time. I don't know what's wrong with me but I can't stand it. If I were superstitious, I would say that someone cast an evil spell on me. I need more time to get over this.

Wednesday, April 17, 2013

The Perfect Brain: Another Nail in the Coffin of the Bayesian Brain

The Impending Crash of the Bayesian Bandwagon

Last August, I wrote a series of posts titled, The Myth of the Bayesian Brain. I argued against the prevailing notion in the AI community that the brain uses some kind of Bayesian statistics to make decisions. I argued that, internally, the brain always assumes that the world is perfect even if its sensory space is inherently noisy. The brain does this bit of magic by filling in any missing information and ignoring irrelevant noise. This cleansing process is essential to reasoning and planning. At least one other researcher (to my knowledge), computer scientist Judea Pearl, has been saying the same thing. Well, a story out of Princeton University points to a new study that corroborates what I have been saying. Essentially, Princeton University researchers found that, when we make an error, the brain's decision making system is not at fault. The system is flawless. The fault is invariably due to faulty sensory information. Here's an excerpt:
Previous measurements of brain neurons have indicated that brain functions are inherently noisy. The Princeton research, however, separated sensory inputs from the internal mental process to show that the former can be noisy while the latter is remarkably reliable, said senior investigator Carlos Brody, a Princeton associate professor of molecular biology and the Princeton Neuroscience Institute (PNI), and a Howard Hughes Medical Institute Investigator.

"To our great surprise, the internal mental process was perfectly noiseless. All of the imperfections came from noise in the sensory processes," Brody said.
The "great surprise" of Carlos Brody and his team is understandable, given their training within the current Bayesian paradigm. But it's never too late to jump off that silly wagon. I am not one to laugh and say, "I told you so". But I did, didn't I?

See Also:

The Second Great AI Red Herring Chase
The Myth of the Bayesian Brain

Tuesday, April 2, 2013

Soul Searching Again

I apologize for the delay in posting Part III and IV of The Secrets of the Holy Grail series. I am seriously debating whether or not I should continue to publish this stuff at this time. Given the dangerous world that we live in, true machine intelligence is not something to be taken lightly.

When it comes out, AI will change the world drastically in a very short order, for good and bad. Scientific and technological know-how will no longer be concentrated within the so-called developed nations. Knowledge is power. There is no doubt that various groups will immediately use AI to gain a powerful economic and military advantage over others. This shit will get out of control real fast and this is why I am paranoid. We are either on the edge of a precipice or on the border of paradise. I hate it. I really do.

I am not claiming that I understand it all but I understand enough of it to know that the rest will mostly be about dotting the i's and crossing the t's. I also realize that this knowledge will come out sooner or later, with or without me. I just need a little more time to think about how I should reveal what I have found so far. I swear, sometimes I wish I was living in the stone age. Hang in there.

Tuesday, March 26, 2013

Secrets of the Holy Grail, Part II

Part I, II, III, IV, V

Abstract

In Part I, I gave a brief description of the brain's memory architecture. In this post, I explain how the brain does pattern learning and catches "thieves" in its sleep.

Winner-Takes-All vs the Bayesian Brain

Although it feels like I am preaching in the wilderness, I have been railing against the use of Bayesian statistics in machine learning for some time now. The idea that the brain reasons or recognizes objects by juggling statistics is ridiculous when you think about it. The brain actually abhors uncertainty and goes to great lengths to eliminate it. As computer scientist Judea Pearl put it not too long ago, "people are not probability thinkers but cause-effect thinkers."

Even though it is continually bombarded with noisy and incomplete sensory data, internally, the brain is strictly deterministic. It uses a winner-takes-all mechanism in which sequences compete to fire and the winner is the one with the most hits. Once a winner is determined, the other competitors are immediately suppressed. The winning sequence is assumed to be perfect. To repeat, the brain is not a probability thinker. It learns every pattern and sequence that it can learn, anything that is more than mere random chance. Then it lets them compete for attention. Read The Myth of the Bayesian Brain for more on this topic.

Pattern Learning

The job of the pattern learner is to discover as many unique patterns in the sensory space as possible. Pattern learning consists of randomly connecting sensory inputs to pattern neurons and checking to see if they fire concurrently. However, keep in mind that a pattern neuron will fire when a majority of its input signals arrive concurrently.
The pattern learning rule can be stated thus:
In order to become permanent, an input connection must contribute to the firing of its pattern neuron at least X times in a row.
X is a number that depends on the desired learning speed of the system. In the human brain, X equals 10. With this rule, the brain can quickly find patterns in the sensory space. It is based on the observation that sensory signals are not always imperfect. Every once in a while, even if for a brief interval, they are perfect. This perfection is captured in pattern memory.

Catching Thieves During Sleep

The pattern learning rule is simple and powerful but it suffers from a major flaw: it imposes no restrictions or boundaries on the growth of a pattern. Without proper boundaries, patterns become more and more complex and the simpler ones eventually disappear, crippling the system. Obviously, we need a way to prevent a pattern neuron from acquiring more complexity than its level within the hierarchy requires. The solution to the boundary problem consists of enforcing the boundary rule:
A pattern may not have duplicate sensory input connections.
Here is another way of putting it: a sensory signal may not arrive at a pattern neuron via more than one path. For example, in the illustration below, pattern neuron A behaves as if it were connected directly to sensors a, b, c, d, and e.
Suppose sensor c was connected (dotted red line) to pattern neuron C. This would mean that pattern neuron A would have two duplicate inputs from sensor c, one via C and the other via D. This is forbidden by the boundary rule. What this means is that somewhere along the paths leading from sensor c to pattern neuron A, there is a bad connection. The culprit is always the most recent or weakest one. It is called a thief because it "stole" something that does not belong to it. By purging thieves from all branches of the pattern hierarchy, the growth of every pattern is automatically limited to a degree of complexity commensurate with its level in the hierarchy.

The power of the boundary rule is betrayed by its simplicity. It prevents runaway pattern growth while facilitating the discovery of every possible unique pattern in the sensory space. It is indispensable to pattern learning and works for any type of sensory patterns, not just visual.
Note: As far as I know, the boundary rule is not in any books. Please make copies of this page on your computer. This is intended to serve as "prior art" in the public domain, i.e., it cannot be patented. :-D
The brain cannot eliminate thieves while it is awake because it must test fire all untested connections. This could cause problems during waking hours. This is one of the reasons that sleep is so important. An intelligent machine, by contrast, is not so limited. During learning, a computer program can examine a branch on the fly to see if a new connection is a thief.

Coming up

In Part III, I will show how learning occurs in sequence memory.

See Also

The Myth of the Bayesian Brain
The Holy Grail of Robotics
Raiders of the Holy Grail
Jeff Hawkins Is Close to Something Big

Sunday, March 24, 2013

Secrets of the Holy Grail, Part I

Part I, II, III, IV, V

Abstract

According to brain and machine learning expert, Jeff Hawkins, goal-directed behavior is the holy grail of intelligence and robotics. He believes that the best way to solve the intelligence puzzle is to emulate the brain. Hawkins is right, of course. There is no question that we can learn everything we need to know about intelligence by studying the brain. The only problem is that some of the answers are so deeply buried in an ocean of complexity that a hundred years of painstaking research could not uncover them. In this multi-part article, I will describe some of the amazing secrets of the brain before revealing the surprising source of my knowledge (no, it's not the brain, sorry).

Liars and Thieves

Let me come right out with a bold statement: nobody can rightfully claim to understand the brain’s perceptual learning mechanism without also knowing exactly what the brain does during sleep and why. Sure, we know what neuroscientists and psychologists have told us, that the brain uses sleep to consolidate recent memories, whatever that means. Unfortunately, that is pretty much the extent of their knowledge on the subject. Hawkins doesn't know either, although he should. That is, assuming he wants to stay in this business. It turns out that the brain performs at least two essential functions while we are asleep: it purges liars (bad predictors) from sequence memory and eliminates thieves (redundant connections) from pattern memory.
Note: I will explain my choice of the liars and thieves metaphors in an upcoming post.
Without these frequent purges, the brain would get confused and eventually stop working. But why is that, you ask? That, my astute and inquisitive friend, is one of the secrets of the holy grail, which is why you must read the rest of the article. But before I can answer your question, I must first say a few things about how memory is organized.

Pattern Memory

I did not always think so but the brain has two types of hierarchical memories: pattern memory and sequence memory. My original objection was that a pattern hierarchy cannot do invariant object recognition. That was before I realized that it doesn't have to; that's the purpose of sequence memory. Pattern memory is a hierarchy of pattern detectors that send their output signals directly to sequence memory. A pattern is a transient group of sensory signals that occur together often and a pattern detector or neuron is best viewed as a complex event sensor. Pattern detectors (red-filled circles) can have an indefinite number of inputs.
A hierarchy makes sense for several reasons. First, it gives us a very compact storage structure because of the inherent reuse of lower level patterns. Second, and just as importantly, it provides a way to automatically limit the boundaries of patterns. This, in turn, makes it possible to discover all possible patterns in the sensory space. I'll have more to say on this later.

A peculiar but critical aspect of pattern memory is that the time it takes an incoming signal to propagate through the hierarchy must be very fast. The cortex uses electric synapses to do this. The end result is that signal propagation through the hierarchy appears instantaneous to the rest of the brain. And the reason for this has to do with timing integrity. For instance, if a high level neuron (A) fires, all the pattern neurons in the branch below A in the hierarchy are assumed to have fired concurrently with A.

Sequence Memory

It would be accurate to say that sequence memory is the seat of intelligence. It is used for many functions such as recollection, prediction, attention, invariant object recognition, reasoning, goal-directed motor behavior and adaptation. Sequence memory contains sequences of patterns organized hierarchically just like pattern memory. Note that, in the diagram below, the pattern hierarchy is shown as a single flat layer (red circles). This is because sequence memory (yellow circles) does not see pattern memory as a hierarchy. That is to say, the system must act as if sensory signals could travel through pattern hierarchy instantaneously. Otherwise, pattern detection timing would be askew.
One of the more interesting design characteristics of sequence memory is that a sequence detector has a maximum of seven nodes or inputs. Why seven? For one, it explains the capacity of what psychologists call short-term or working memory. Second, it is a compromise that aims to minimize energy usage while maximizing the breadth of focus. As it turns out, the brain can focus on only one branch of sequence memory at a time. A branch should be seen as a grouping mechanism that represents a single object or concept. No need to look any further. The branch is the mechanism of both attention and invariant object recognition.

What is even more interesting from the point of view of invariant object recognition is that multiple sequences may and do share patterns. In fact, every complex recognized object in memory consists of multiple, tightly intertwined sequences. This will become clearer later.

Coming up

In Part II, I will explain how learning occurs in pattern memory and how to catch a thief.

See Also

The Holy Grail of Robotics
Raiders of the Holy Grail
Jeff Hawkins Is Close to Something Big

Wednesday, March 6, 2013

Raiders of the Holy Grail

Numenta Aims High

In my recent article, The Holy Grail of Robotics, I wrote that I was impressed with Jeff Hawkins' description of goal-oriented behavior in his book On Intelligence. In a recent blog post about the Obama administration's $3 billion science initiative known as the Brain Activity Map Project (BAM), Hawkins wrote the following:
The activity and connection maps envisioned by BAM will be useful, but brain theorists today are not lacking in empirical data. We haven’t come close to understanding the tremendous amount of data we already have. If we want to understand how brains work, then a better direction is to focus on brain theory, not brain mapping. We should set goals for brain theory and goals for machine intelligence tasks based on those theories. That is what we do at Numenta. For example, we set goals to understand how neurons in the neocortex form sparse distributed representations and then how they learn to predict future events. This resulted in Cortical Learning Algorithm (CLA) which is the heart of our Grok streaming prediction engine. The next big theoretical challenge we are working on is how the cortex generates behaviors from predictions, what is sometimes called the sensory-motor integration problem.
This is even bigger news than Numenta's recent announcement of Grok, in my opinion. I would not think so had it come from anybody other than Hawkins. As I wrote elsewhere, Hawkins is no dummy. This tells me that Numenta is aiming to solve the entire intelligence problem single-handedly. Why? Because, once you have figured out how to do both perceptual and motor learning, there isn't much more to add other than an appetitive/aversive learning mechanism. This is psychology lingo for a pain and pleasure (reward and punishment) mechanism, a must for adaptation. But this is a rather trivial problem once you've gotten this far down the road.

The Challenge

Creating a viable model for sensorimotor behavior is not an easy task. It starts with designing a working perceptual learning system (both pattern and sequence recognizers) and an attention mechanism. I don't think Numenta has perfected either of those, regardless of the hype emanating from Redwood City. An attention mechanism is a must because there can be no coherent motor behavior without the ability to focus.

There is more to motor learning and behavior than what happens in the neocortex, however. Mammals and birds have an additional sensorimotor control structure known as the cerebellum. Humans use it to help with a bunch of automatic tasks such as walking, standing, maintaining posture, and even driving. The cerebellum works on completely different motor control principles than the neocortex. It is needed because it frees the neocortex from having to handle routine motor behavior so it can focus on other things. But even without a cerebellum, a robot could still learn some sophisticated skills.

Even without a good perceptual learner, one can still build a very impressive learning robot with multiple degrees of freedom. This is assuming one has the motor learning part right. Hawkins can save his company a boatload of money by reading my recent article on goal-oriented motor learning. In it, I  gave a biologically plausible definition of goal and explained how the neocortex finds the right motor connections for any given goal. I already did a major part of the homework on motor learning. It took me a while to arrive at my current model but it's easy to explain to others once you know how it works. So, in my opinion, Hawkins would do well to skip to something else. For example, he will have to figure out how to handle motor conflicts but, from my perspective, this is not that hard either.

The AI Race Is On

There is something new in the air. There has been a frenzy of activity in AI and brain research in the last couple of weeks. A lot of money is suddenly being allocated for research by both the government and the private sector. It's strange but there is a sense of desperation in the air, as if time was of the essence. I don't know what but something must have happened to trigger this. What is certain is that the race is on to be the first to understand how the brain works.

My prediction is that these initiatives will fail. Like Hawkins, I don't think that throwing money at the problem is the way to go. At this time, I think Jeff Hawkins has the best chance to unlock the secrets of the brain. He knows a few already. However, unless he can fully grok perceptual learning and attention (he doesn't, even if he thinks he does), his efforts will also fail in the end. He may come up with a useful gadget or some other product but the holy grail of intelligence and robotics will remain out of his grasp.

In the meantime, I continue with my own efforts and I don't need a million dollar budget. All I need is a personal computer and some spare time. May the best model win.

See Also:

The Holy Grail of Robotics
Goal-Oriented Motor Learning