The Language Based Interface (LamBDA) as a Communication Medium

 In this post, I summarize some of the conclusions that come out of the LamBDA issue controversy. To find out what LamBDA actually is, check here.  This post deals with the current controversy, which is whether "LamBDA" or any similar creation can be "sentient." Spoiler alert: noting we are doing now places us on the road to creating a sentient machine.

A sentient computer would be like a sentient book.

Computers talk to us like books, TV, and phones talk to us. 

Talking computers descend from centuries of effort that have given us talking parrots.

When we see intelligent words coming out of a computer, our natural instinct is to attribute sentience to the computer. We will get over it, just as we got over our amazement at words coming from the first telephone.  Or a parrot.

Under normal circumstances, I use words to cause mental events in your mind. This depends on both understanding the language and having a reasonably similar experience of life. In this way, language is a means of communication between two sentient beings.

If I use a megaphone or a telephone to communicate with you, this has not changed.  Here, I use an incredibly sophisticated suite of technologies to communicate with you.  Here, it will not occur to you to think that these words are coming from some element of that technology, such as a word processor AI. For example, you will assume there is some kind of conscious intent behind all this, and likewise, I assume that you are a sentient being.

Google developed LamBDA for an automated way to summarize what you get when you submit a Google query.   That would be part of "many-to-many" communication that originates with human beings who have the intention of communicating something in perhaps many ways to many human other beings. Yet, the technical "bucket brigade" between them and you is still no more "sentient" than your TV set.  And, again, the "information" under the system matters more than the presentation. One might argue that wrapping it in a "conversation" may make the whole thing more credible than it should be.

LamBDA is not only an "intelligent" speaker but also an "intelligent" "listener" and "learner."  The ability of LamBDA to "listens" and "learn" represents the current state of the art in natural language processing that is half a century old. "Training" consists of tweaking a vast array of parameters until the machine starts getting the right answers (or talking) in the desired way. Apple's SIRI and Amazon's Alexa are products of this engineering effort. This would be fairy magic in 1970. It is now routine.

Our family tree includes leaves. The computer's family tree includes door knobs*.

The building blocks of sentience are a special case of what happens everywhere in our bodies--electro-chemical reactions and large molecules like proteins. This is the same "took kit" we see in photosynthesis. 

What happens in a brain does not, as commonly imagined, resemble what happens in a computer. 

Our brain has more in common with a tree than with a computer. To put it yet another way, the simulation of a single protein reaction, let alone a cell or a tree, is far beyond the current state of the art. The problem is that "calculation" and "life" are fundamentally different things.

A computer belongs to a family of objects we regard as "tools" - useful for some human purpose. A simple example of an ancient "computer" is a doorknob. It is locked or not--a switch. The diagram at right shows how transistors implement a "gate." This is what a computer "means" by "yes or no." The Internet is a vast collection of electronic doorknobs.

It is only a human's understanding of a doorknob that makes it a doorknob. A doorknob doesn't "know" it's a doorknob.

LamBDA - A Stochastic Parrot

The more you know about LamBDA, the less impressive it becomes. 

LamBDA is the "front end" of a system cobbled together by Lemoine.  Blake Lemoine, one of Google's most skilled engineers, used years of effort and hundreds of interactions to create what should really be called something like a "Lemoine Machine."It is wrong to think of Blake's LamBDA as a single "thing" or entity. It is a setup that includes a LamBDA instance (version). Let me call it LM - Lemboines machine.  LamBDA itself is an experimental technology that implements a natural language interaction with a body of knowledge. The "body of knowledge" is what we are "talking to." It is equally important to know how Lemoine trained LM to talk the way we hear in the only public conversation available.  He has "cherry-picked" just one of many hundreds of "chats." Nobody will ever see more samples of these chats, nor will anyone ever see a new sample. LM, like Blake, no longer works for Google. 

LM is an instance of a long series of LM's, each tweaked by Lemoine to talk about an unspecified of knowledge that is interesting to Lemoine, who is a bit of a mystic. As an engineer, Lemoine tweaks not only what his creation says but how it says it. Lemoine deeply resents the fact that Google, as an organization, is hostile to his "religious" convictions.

Blake made himself famous by claiming to create a "sentient" machine. What Lemoine says about "LamBdA" comes from Blake, wearing his "preacher hat." We must remember that Blake is a professional mystic, trained in what we might call "leaps of faith."

What LamBDA itself "says" is what other people have said that Blanke finds "interesting" and spoken in a way that satisfies Blake's engineering goals. The result is impressive but, to say the least, open to interpretation. 

"Training" of an AI is a sophisticated process, but we all know how we can use patient training to produce interesting results. Parrot training is the best-known example.

------

*AI. The term "Artificial Intelligence" is unfortunate. Originally, it meant a computer program that did something that would, in a human, require "intelligence." The computer program is acting as if it had intelligence (but it really doesn't).  "An AI" has come to mean a computer application that is specifically NOT "intelligent" in how a human is intelligent. In fact, "intelligent" computers arrive at "intelligent" behavior that is notably NOT analogous to the way humans "think." This has been the core insight of the effort from the start.

"Intelligence" itself is a slippery idea, but in this context, many people take it to mean "conscious" or "sentient," terms that are confidently applied only to a living creature. When we speak of "intelligent" machines, we refer to an engineering concept that never pretends to be anything else. In that sense, the "artificial" in AI can always be omitted. An "AI" is just a product of clever engineerings like a camera or a drone. Perhaps we should use "Intelligent Door Knob" - IDK. We should at least stop using "artificial" to mean its opposite.

Frames All The Way Down

I have been rather dismissive of Hawkins' "Thousand Brains" idea but I can't help thinking he's on to something, specifically:

• Once you broaden your understanding of what a "frame*" is, it is quite plausible that people think in frames. The most powerful computer system we have built so far, the Internet, runs on a relatively simple "frame language", understood by over a billion "hosts" including the one you are reading this on.
• Hawkins provides a plausible description of human experience in terms of "frames" somehow being the unit of experience.
• It is plausible that the neocortex has some kind of structure that implements frames, including the ability to create frames "on the fly" and use them as a unit of memory and reasoning.
• Thousand Brain is not too dissimilar from the theory presented in Surfaces and Essences, which presents analogy as the "fuel of thought". "Surfaces" is really talking about language, which is a brain-to-brain communication device.  Theories of language, such as Chomsky's use frames to develop a theory of how language works.

Hawkins gets tripped up trying to describe complex frames, especially the ones that don't admit to a spatial interpretation. Object-Oriented Design has a rich language that applies directly. Starting at the simplest example, Hawkin's coffee cup, we can speak of objects as "frames". Object-Oriented concepts are extremely powerful in representing the real world - one reason these concepts underlie the Internet and some of its fundamental tools, such as Java.

Hawkins gets into trouble when he starts to speak as a neuroscientist, which he is not. He doesn't get published in Scientific Journals. In spite of his extensive knowledge of the subject, his ideas about how, exactly, columns in the neocortex create frames are sketchy, to say the least. But even here, he is pointing to a structure that is replicated all over the neocortex (the famous thousand brains). If we assume that we think in frames we may be forgiven to think that some structure in the brain is involved. Due to the enormous parallelism and recursiveness in experience, we would expect a large number of similar structures, not some kind of "organ".

Hawkins is particularly interested in the predictive nature of consciousness. I think he could be challenged on this. The brain is obviously filling in what it "expects" to be there but I am not quite sure we need the "priming" idea. It seems the brain "sees" what it expects to see in a "frame". What needs to be explained is how effortlessly it does this and how we are alerted to aspects of the real frame that are out of place. We know that this mechanism often fails, as in the phenomenon of "Change Blindness". In fact, it seems that change blindness may be the Achilles Heel for this aspect of Hawkins' theory. On the other hand, answering this question could turn out to provide strong extra evidence for the theory. 

Many applications in the Internet are at least attempting to be "predictive" - filling in details before the query is completed. "Type Ahead" is something we all experience. Hawkins seems to think that this is something "AI" can't do.

My thinking on this subject rides off in several directions.

• How it is that certain "virtual" environments, such as Second Life, feel so "real". Hawkins' ideas about felt reality are applicable.
• How do frames work in non-spacial domains, such as language, physics, and mathematics?
• Is the Universe actually built of frames or does it just look that way?
• Hawkins' book is very light on visual examples that would show the nested and recursive nature of the frames he is talking about. I'd like to provide some examples and perhaps shed some doubt on the idea that a "frame" is manifested in a column of the neocortex. 
• I'd like to present an alternative interpretation of the neurological "facts" in Hawkin's book. Specifically, I think a "frame" corresponds to a network of active neurons in the brain. There is an interesting visual example in Second Life, where a "frame" (a specific location in the world) "rezzes" a bit at a time, providing a way to think of how a frame could resolve itself almost instantly even though it is full of nested frames.

=================

* I use "frame" to refer to Hawkins' "reference frames"

See also

Object Oriented Design

The Languate of Information Science

A Thousand Brains

I posted this review of "A Thousand Brains" on Amazon. Bottom line, there are great ideas but, at bottom, the "science" in it is extremely shakey. If you toss out the hardware ideas and focus on the systems analysis, it's brilliant. This reminds me of another great book on the workings of the mind, Surfaces, and Essences. In that book, the authors wisely steer clear of speculating on how analogy and metaphor are implemented in the brain. "A Thousand Brains" can be profitably read in this spirit.

In subsequent posts, I will be seeing what can be done with Hawkins' theory if we see it as systems analysis rather than neuroscience.

"

AMAZON REVIEW

If this is the only neuroscience book you read, you'd think neural columns are a "thing". I am reminded of the brilliant idea of "memes", which fell apart due to the inability to define them rigorously. It is no coincidence that Richard Dawkins (inventor of the meme) is over the moon about columns and frameworks.

Even so, Hawkins' insights about the constraints a brain theory must satisfy is insightful,  and probably a major stepping stone.

But while he tells me a lot of new interesting facts about neuron function, his idea that a column is a little prediction machine is presented without evidence. Even if the idea is true, Hawkins seems to miss the fact that a single column MUST participate in multiple frames, therefore there is no one-to-one correspondence between frames and columns. Frames are created by a set of related neurons firing. If these relationships are, as he claimed, determined by the synapses at the neuron, the frame is ultimately a network of synapses, or the connected neurons firing together. This cannot be a new insight. The book repeatedly leaves the impression that is the column that is "learning" to predict, rather than the network of columns. As they said about Freud, what's true ain't new, what's new ain't true.

The fallacy is quite evident in the title, which hints that our brain is really a thousand "micro-brains". While the number is impressive, we seem to be able to learn and recognize an infinite number of frames. Such a huge number can be reached if "frames" are combinations of synapses of which we have 125 trillion. The number of possible COMBINATIONS of 125 trillion synapses is, for all practical purposes, infinite.

This is not to say that "frameworks" are not a great idea. They satisfy a lot of the constraints that the author sets out. The brain probably does work this way. But if we look deeply, the frame theory works on its own and has nothing to do with the hardware that supposedly implements it. The hardware in the title.

As an engineer, the author sails off into Science Fiction when he claims he knows enough to build a general-purpose AI.  This is not to say that he will not come up with something interesting.

A general-purpose machine that implements frames already exists: the Internet. If we examine any Web page in detail, we will see it is bristling with links under the hood (these can be seen in the HTML that creates the page). Your iPhone screen is a frame, as are all the icons. The software that drives the icons is "object-oriented", full of links and references to other objects in the world. Individual objects participate in multiple frames. This is technology now 30 years old. Its fundamentals can be grasped by any intelligent layman. They are simple, but not beyond understanding. You don't need to be (as they say) a brain surgeon.

In spite of its vast size, the Internet is not yet "intelligent" in the way our individual brains are, but this is perhaps a hint that Hawkins will not be cooking up a general-purpose machine any time soon.

In spite of the specific failings of Hawkins' theory, there are enough promising elements in it for me to pursue what makes sense and leave aside a few specifics. Hawkins is a perfect example of the importance of asking impertinent questions.

----------------

Hawkins' conversation with Sam Harris

Hawkin's conversation with Lex Fridman

Russian Dolls and The Mind

On Tue, Mar 12, 2019, at 7:30 PM Len Bruton <...> wrote:

There seems little doubt that science looks for explanations by seeking out those physical objects within brains that may yield a closer appreciation of how minds function and maybe just where they are to be found!!  Following your very different approach, my thought experiment is to ask whether Earthly minds and whether Earthly consciousness existed 100 million years ago?  Is it then a valid question to ask whether reality existed prior to minds

This question can send us spiraling down an infinite regress. There seem to be two ways of stopping this. Either accept the reality of "mind" or accept the reality of "the world". Otherwise, you have a problem of reality being something in the mind, which is itself in reality, or a model of reality and so forth ...

Wayne Brown <chillyfinger@gmail.com>to Len
Given that we have a mind, it also makes sense if other living things have a mind or something like it. Such questions can provide insight into our own mind. It even makes sense to ask if Google has a mind. 

I have no doubt that my dog has a mind of sorts. However, she is stuck with a mind that she inherits through her genes (which is why it makes sense for her to circle around before lying down to sleep). Wolves share their minds when they hunt down a moose but within pretty sharp limits. I think neuroscience can provide a lot of insight into mammalian minds and we are mammals. 

Our minds are built up from birth by language, culture, becoming shared minds. Human minds can be fruitfully explored by the experience of sculpture, astronomy, photography and so much more. If our minds were "nothing more" than mammal brains we would not be interested in them at all. 

But your question was about reality, not minds. I think all life is deeply rooted in reality. Life sucks as much information from the real world as possible - information that can be turned to advantage in the struggle for survival. Information that can be stored within the physics and biology of the organism. As it happens, such information, once obtained, can be used for other purposes. 

Humans do even better by obtaining information from each other, through reason and through instruments. Still, it is information about reality and not reality itself. Some would argue that, since we cannot know reality directly, we can't really know that reality exists at all. This type of talk may be suitable to impress half drunk girls at a party, but can't be taken seriously. To follow Descarte's line of thinking just a bit further: to "know" reality implies a knower and the subject of knowledge (knowledge about ...). Our "model" of reality is not reality itself, but it is wonderful, endlessly expansive, explorable and shared. 

It's frustrating when we bump into fundamental limits of observation, such as we do in cosmology and quantum mechanics. In these areas, we must lay aside our tools of observation and rely on pure reason. At that point, it becomes painfully obvious that we are discussing ideas about reality rather than reality itself. In the "real" world, perhaps logic, space and time do not exist or exist in a way we cannot observe. Or perhaps we have the wrong idea of "existence" itself.  

In any case, exploring the model of reality the mind creates is more productive than attempting to create a model of mind in reality (or, more precisely, our mind within our model of reality.). We've been doing this for thousands of years and progress is being made steadily. 

Neuroscience will never discover art or physics in my brain, nor will it discover my model of reality. Building that model is the very purpose of the organ in question. To learn about my model of reality, it's best to explore reality itself.

Brain Model in a Nutshell

"Growing" Brain Tissue

A key insight is to take into account the fractal nature of how the brain grows (like any other tissue) from a single stem cell. Model structure should reflect the way cells grow and connect rather than relying on mathematical structures like matrices. Connections look random but have an underlying fractal structure. On the other hand, the model needs to reflect the well known principle that "cells that fire together wire together". The fractal structure "explains" this rule in a physical way (cortical columns) but the brain can wire itself up in a way that might be seen as 4 dimensional. Connections are formed due to "closeness" in space as well as some kind of time function. 

The model should assume some level of continuous growth. In the brain, new neurons are being created and others are dying all the time. It's not known how these new neurons find their place in the structures of the brain.

Similarly, we know that the brain is, to some extent, "plastic". It allows large-scale re-wiring, probably between columns rather than within them.

Pre-frontal cortex and the "idea"

The prefrontal cortex may work just like the specialized areas of the brain, but feed back on itself to create recursive, fractal "layers" on the fly, at approximately "brain wave" frequency. We know that this process can feed back into the specialized cortex to re-use these circuits to present "ghost" images (like visual memory). Perhaps this is what we mean by an "idea" and account for how these ideas seem to flow in the experience we call consciousness. The rate that these images are formed (brain wave frequency) is interestingly similar to the frame rate of a movie or the perceptible range of audio frequency.

What would be the model of "stimulation"?

Using the example of the retina, a cell would consider itself to be "stimulated" as a function of stimulation in nearby (connected) cells. Connection would follow the fractal patter, with daughter cells of the same stem cell being considered "close" as well as those physically in contact (taking into account direct cell-to-cell communication present in all tissues). Direct stimulation would take place only at one "layer". Would that be layer zero or layer K? Or should we consider the idea of stimulation to define layer zero?

Cell logic structure

Signalling between cells must include the concept of suppression: a signal from a to b may reduce the probability that b becomes "excited'. In general, the "method" in the cell that determines to signal the event "I am excited should involve the range of simple logical constructs (and, or, nor, not) along with dynamic versions of these functions (sum over time). In short, triggering will result from exceeding a threshold described in a set of discreet non-linear functions. Inputs to these functions would be internal variables reflecting  stored "awareness" of events in connected cells. Perhaps we can assume that "time" resets to zero when the cell "fires" or at least assume the cell is in one of two discreet states: firing and not firing.

The set of equations determining whether the cell "fires" or not can be thought of as the model for the cell's "output" axons. The "input" (right hand side of the equations) represent a model of the cells dendrites (inputs). The actual set of equations (the way all these variables are related) can be thought of as the model for the neuron's internal "logic".

Some provision should be made for "noise". In other words, "firing" should not be deterministic, but a result of increasing or decreasing probability.

We may save ourselves the trouble of modelling a "not firing" event by allowing the corresponding "receiver" internal variable to "decay", requiring occasional refresh ("still firing").

The State of the Cell

With each "tick of the clock", the cell proceeds from one state to the next, applying the state model to produce a new state (list of internal variables). At the end of the "tick", it signals the "firing" event or not, then "decays" the state variables according to some internal rule. It then "goes to sleep" unless "woken up" at the next tick or notified of the need to change a state variable due to the external event of the corresponding neuron having fired. Of course, this can be generalized simply by treating the "tick" of the clock as itself an "event".

This is not particularly original concept. It's a simple-seeming idea but nasty to "debug". The event-driven design concepts are drawn from the "tool box" used by the computer's BIOS. For performance reasons, programming should at least provide for an ultimate version that binds closely to the actual computer BIOS, using the actual "tick" of the computer's clock. Obviously, we can't imagine triggering millions of cell events at each computer "tick" (which happens in the nanosecond order of magnitude). However, our model should provide for the possibility that the lists of listeners is partitioned between CPU's and that the cells can refresh themselves independently, possibly using or sharing different CPU's. At the lowest level, then, the model should not assume that everything happens on one CPU, nor that the event list is shared between all CPU's. All state data should be stored within the cell or the event list.

Cell "logic" should be kept simple to provide for direct implementationi in silicon. This would permit mass production of cells for a given "tissue type".

Signalling and events

"Firing" of a cell as an event picked up by "listeners". Ultimately, some cells will "fire" as a result of direct, outside stimulation.The synapse (connecting neuron A to B) would be modelled by the the list of cells Bi listening for firing events from cells Ai. Physical implementation constraints would be proportional to the capacity to store these lists. Dynamic performance would be determined by the refresh rate of the lists (how quickly "events" can be picked up to fire a new set of events). Ideally, the refresh rate should be fast enough so that there are "long" periods where "few" events take place. Cycling between "few" and "lots" of events should mirror brain wave frequency (order of 10 to 100 Hz).

Random "wiring"

How do we model the possibility of virtually any neuron connecting with any other if they "fire together"? This would imply dynamically adding A to the list of "listeners" to the "B fired" event. The initial list of "listeners" should be created as a result of fractal growth from a daugher cell. After that, listeners could be added or dropped as a result of a different rule ("Experiece"). Perhaps we could feed the stream of event states into a (hopfully fast) external machine that would detect correlation using old fashioned statistical methods, then feed back a stream of "suggested" new connections (listener pairs A, B). This would, in fact, simulate "learning".

Qualitative and Quantitative Measures

Even if not implemented in silicon, dynamic properties of the model could be determined and compared to what is known about the brain, such as the way cells signal, the number of cells, the size and structure of cortical columns. "Order of magnitude" estimates could give us an idea as to whether this type of model can tell us anything about how the brain works. Assuming that the brain is more complex that this model, perhaps we can bracket brain complexity, performance and the types of brain functions that can be implemented in silicon.

How can "memory" be Modelled?

I assume that "memory" is implemented in tissues similar to what is discussed above. In particular, there internal state is persistent. Their output, however, is assumed to feed back to produce "ghost" perception. At first glance, this behaviour could be simulate simply by a longer decay of internal state variables, simulating a persistent neuron-to-neuron connection (axon to dendrite). Again, such connections would be richer within daughter cells. We can't assume that the column structure applies since brain structures such as the hippocampus (known to be involved in memory) are not cortex tissues. At first blush, the "wiring" may be simulated by a fractal structure applying to one big "blob" of "tissue" called "memory".

Overall Structure

All of the above hints at a structure of similar tissues arising as variations on a theme, with every region differing mainly in the persistence of internal state variables, fractal "wiring" and the degree of looping (recursive) "wiring". Frontal cortex is "loopy" with state decaying rapidly unless refreshed. Sensory cortex is much less "loopy" but fast decaying. Connection is highly structured along fractal lines. Memory is extensively connected to everything with slow internal decay rate for internal state.