Natesh Ganesh asks, "Can minimal dissipation alone be a sufficient condition for learning?"

Two biological systems might fit this description.

First, in Charles Gallistel's book The Organization of Learning, Chapter 11 tells of foraging experiments. Consider fish (are neurons like fish?). Given probabilistic feeding stations, the school divides itself proportionally— which from the...

view entire post

Natesh Ganesh asks, "Can minimal dissipation alone be a sufficient condition for learning?"

Two biological systems might fit this description.

First, in Charles Gallistel's book The Organization of Learning, Chapter 11 tells of foraging experiments. Consider fish (are neurons like fish?). Given probabilistic feeding stations, the school divides itself proportionally— which from the perspective of game theory looks like a Nash equilibrium: no single fish can increase its payoff by leaving one station to feed at another. In this situation, no food energy is "dissipated" because every scrap is eaten. So according to the hypothesis, learning must be involved. Indeed, if this were a probability learning experiment with just one fish and food energy given per play of the game, we would see probability learning (also as in Chapter 11). In probability learning food is "dissipated." Perhaps because of fear of leaving one group and joining another, the visiting behavior that's seen in probability learning is attenuated when the school forages as above; and then no food energy is dissipated.

Second, in Goranson and Cardier's A two-sorted logic for structurally modeling systems, there is a review of apoptosis (programmed cell death, which according to the paper happens in the human body about a million times a second). In the olfactory system, apoptosis of neurons amounts to learning new smells, which involves consciousness.

In Barwise and Seligman's Information Flow: The Logic of Distributed Systems, the example of a flashlight is given. The system— a flashlight in this case— supports different "local logics" and therefore different languages between which there may be perfect or imperfect translations. The language in this essay could be one such language. But there are others, as the example of the flashlight shows (including the specification language that specifies the purpose of the flashlight). Goranson describes software by which many different languages, local logics, or situations like this can be organized.

view post as summary

Thank you for your patience with me, Natesh. I see you are at a nanotechnology lab, so I imagine that "I squared R" dissipates a lot of heat that is of concern. How much dissipation in your hypothesis is from I^(2) R ?

From my work experience in software, implementing a FSM by an array with two indexes, one for current state, one for current input signal, storing at those indexes next...

view entire post

Thank you for your patience with me, Natesh. I see you are at a nanotechnology lab, so I imagine that "I squared R" dissipates a lot of heat that is of concern. How much dissipation in your hypothesis is from I^(2) R ?

From my work experience in software, implementing a FSM by an array with two indexes, one for current state, one for current input signal, storing at those indexes next state and output signal, is much quicker and easier to debug than leaving the FSM in a lot of "if-then" statements-- I.e. using data instead of code increases the speed of execution and reduces debugging significantly. Then if the rest of the code is needed for the full Turing machine, most of the speed loss and complexity in my coding experience would come from the full Turing machine, not the FSM. That's why I ask. And, in your line of work, it seems to me that the vonNeumann architecture would be more relevant for I^(2) R dissipation than more abstract models like co-algebras, streams, or Chu spaces for modeling computing. For example, as in Samson Abramsky's Big Toy Models.

In engineering terms, apoptosis is a quality control in animals (it never occurs in plants) that seems like a way to minimize energy loss and therefore may be relevant to your dissipation hypothesis. In apoptosis, defective cells that would be energy-wise inefficient are destroyed and their components re-used to build other cells, thus holding onto the potential energy in the sub-assemblies and again reducing energy loss. But I don't see how I^(2) R heat loss plays a role in apoptosis.

If you had an equation for dissipation rather than a verbal explanation, It might give me something more general than I^(2) R to think about-- especially regarding apoptosis as an example of minimizing dissipation of energy. What are your thoughts?

Regarding your references-- I will try to find them online. Are they in arxiv? The closest big universities where I could photocopy these papers are hours away from me.

view post as summary

Natesh. All I can get on the second paper is the abstract and a bit of the intro. The rest is behind a paywall.

You wrote, "When you talk about I^2*R expression for dissipation, you are thinking about wires and charges moving through those wires. So that expression will not hold for spin based architectures." It's Joule heating, as in this...

view entire post

Natesh. All I can get on the second paper is the abstract and a bit of the intro. The rest is behind a paywall.

You wrote, "When you talk about I^2*R expression for dissipation, you are thinking about wires and charges moving through those wires. So that expression will not hold for spin based architectures." It's Joule heating, as in this article:

http://poplab.stanford.edu/pdfs/Grosse-GrapheneConta
ctSJEM-nnano11.pdf

And, you wrote about "entropy":

"I want to point out again that I use the dissipation bound as a (good) approximation of the actual dissipation in the processes that we are interested in. The dissipation bound expression as entropy \delta S and mutual information terms I."

But the abstract talks about "energy dissipation":

****

Abstract:

A hierarchical methodology for the determination of fundamental lower bounds on energy dissipation in nanoprocessors is described. The methodology aims to bridge computational description of nanoprocessors at the instruction-set-architecture level to their physical description at the level of dynamical laws and entropic inequalities. The ultimate objective is hierarchical sets of energy dissipation bounds for nanoprocessors that have the character and predictive force of thermodynamic laws and can be used to understand and evaluate the ultimate performance limits and resource requirements of future nanocomputing systems. The methodology is applied to a simple processor to demonstrate instruction- and architecture-level dissipation analyses.

…I. Introduction

Heat dissipation threatens to limit performance gains achievable from post-CMOS nanocomputing technologies, regardless of future success in nanofabrication. Simple analyses suggest that the component of dissipation resulting solely from logical irreversibility, inherent in most computing paradigms, may be sufficient to challenge heat removal capabilities at the circuit densities and computational throughputs that will be required to supersede ultimate CMOS.1

For 1010devices/cm3 each switching at 1013s−1 and dissipating at the Landauer limit Emin≈kBT have Pdiss=414 W/cm2 at T=300K Comprehensive lower bounds on this fundamental component of heat dissipation, if obtainable for specified nanocomputer implementations in concrete nanocomputing paradigms, will thus be useful for determination of the ultimate performance capabilities of nanocomputing systems under various assumptions regarding circuit density and heat removal capabilities.

****

Are you saying the heat dissipation in post-CMOS devices is just due to spin, and none of that heat is due to Joule heating?

Best Regards,

Lee

view post as summary

Natesh,

Great essay, well informed and organised analysis of a very interesting hypothesis.

I also like most and agree with much, certainly with the addage that; 'our aims and goals are shaped by our history', and the importance of; efficiency, neuronal avalanches, branching parameters, critical regions and that a 'hierarchical predictive coding' model is possible.

I'm not...

view entire post

Natesh,

Great essay, well informed and organised analysis of a very interesting hypothesis.

I also like most and agree with much, certainly with the addage that; 'our aims and goals are shaped by our history', and the importance of; efficiency, neuronal avalanches, branching parameters, critical regions and that a 'hierarchical predictive coding' model is possible.

I'm not yet convinced that minimal dissipation itself is a precondition of learning and 'inference dynamics' and didn't feel you proved that. Do you not think it might be more a ubiquitous characteristic than a driver - in a similar way to 'sum over paths'? If 3 lifeguards head off to save a drowning girl 100m down the beach, one heads straight for her, one to the shore point opposite her position, and one at the right angle to allow for swimming slower than running; Of course the 3rd gets there with least energy, but I feel that underlying that may be a still greater and more useful truth and meaning.

Also might the error/feedback mechanism be better described as iterative value judgement comparisons. Perhaps no 'right or wrong' just different outcomes, which we can't value till compared with previous runs.

Say the first 'run' of consequences is imaginative drawn from input history. Say Do I want a PhD Y/N? gives an 'AIM' (Y1). We run a scenario to imagine it and implications. We then keep running it as more data comes in. Subsequent lower level/consequential Y/N decisions are the same, taking Y1 into the loop, an on hierarchically. If it turns out not to be as envisaged, or we win millions and want to be a playboy instead, we change the aim to N and form new ones.

May it be that you're a little too enamored by the formula and suggest conclusions from those rather than the deeper meaning they're abstracted from.

i.e. Might Lifeguard 3 have said 'how do I get there fastest' rather than '..by using least energy'? (or just done that due to it's successful outcome in past re-runs of the scenario).

Lastly on 'Agency', which I see as a semantic 'shroud'. If all subsequent Y/N decisions in a cascade are simply consequential on the first Y/N decision, and the branches lead to a mechanical motor neuron or chemical response, all repeated in iterative loops, might the concept 'agency' not fade away with most of the mystery?

All 'leading edge' questions I think and thank you for a brilliant essay leading to them. Top mark coming.

Best

Peter

view post as summary

Hi Natesh,

sorry about the names! I am a slave of rhymes, and tend to map together all what sounds similar. Actually it's even worse, I also cluster faces together. By all means, I must learn to represent information more injectively...

Yes, sure, as I see it, it may well happen in the brain of the observer. There are many possible settings, which I discuss below. But in all...

view entire post

Hi Natesh,

sorry about the names! I am a slave of rhymes, and tend to map together all what sounds similar. Actually it's even worse, I also cluster faces together. By all means, I must learn to represent information more injectively...

Yes, sure, as I see it, it may well happen in the brain of the observer. There are many possible settings, which I discuss below. But in all settings, ascribing agency (as I understand it) requires an entropy-reducing mapping. If the mapping is not injective, it may still be interesting or useful for the observer, and he or she may still make a valuable acquisition in their life by learning the mapping. But I can hardly relate this operation with perceiving a system that seeks to achieve a "goal". For me, a goal is something that tends to be reached irrespective of factors that tend do interfere with its accomplishment. That is why I require the non-injectivity: the putative obstacles are a collection of input conditions, that are supposed to be overcome by the goal-seeking agent.

Now for the variable settings.

One option is: I am the observer, and I receive information of the behavior of some external system (say, a replicating DNA molecule outside me), and by defining the system under study in some specific way, I conclude that there is some goal-oriented behavior in the process.

Another option is: I am given some information which I represent in my head, and I perform a certain computation with that information, for example, the AND gate you mention. The computation happens inside my head. But I also have observers inside my head, that monitor what other parts of my head are doing. For one such observer, what the other end of the brain is doing, acts as "external" (in all computations except for self-awareness, which is the last part of my essay). One such observer can assign agency (if it wants to!) to the computation, and conclude that the AND gate (inside my head!) "tends" to reduce the numbers 1, 2, 3 to the digit 0 (or whichever implementation we choose). A weird goal for an agent, but why not.

All I am claiming is: goal-directed behavior does not exist without an observer that tends to see things in a very particular way: as non-injective mappings. I am not claiming that this is the only thing that can be learned by a plastic system (one-to-one mappings can also be learned). I am not claiming that the only thing that can be done with a non-injective mapping is to arrogate it with agency and goals. There are many more things happening in the world that may be good to learn, as well as in the brain of the observer. And there are many more computations to do, other arrogating agency. My only point is: if we see a goal (inside or outside us), then we have trained ourselves to interpret the system in the right manner for the goal to emerge. The goal is not intrinsic to nature, it is a way of being seen.

Not much, just one tiny point. Or perhaps, just a definition. If you look through the essays, there are as many definitions of "goal", "agent" and "intention" as authors...

> my main contribution is tying it all together in a physical sense

Yes, and that is precisely why it is so great! And perhaps you are right, within your framework, what so far has been presented as a mere description of the critical brain, now can be seen as the natural consequence when certain physical conditions are assumed. I do appreciate that.

Anyhow, time to rest! buenas noches!

inés.

view post as summary

Dear Natesh,

I very much enjoyed reading your most impressive essay. Since you have read mine and commented, I will look at possible correlations, based on the assumption that one of us is actually representative of reality. In fact, even if my essay is correct about the universal nature of awareness, your model may well 'simulate' awareness and may describe realistic governing...

view entire post

Dear Natesh,

I very much enjoyed reading your most impressive essay. Since you have read mine and commented, I will look at possible correlations, based on the assumption that one of us is actually representative of reality. In fact, even if my essay is correct about the universal nature of awareness, your model may well 'simulate' awareness and may describe realistic governing constraints on dynamics of learning. For example you model a sense of agency as "the awareness of an action being performed as it is being performed." This is compatible with your definition of agency as "pre-reflective subjective awareness that one is initiating, executing, and controlling one's own volitional actions in the world." The key word is of course 'subjective', and that is the great question underlying whether or not the singularity is possible.

Let me first say that the qualms I have about quantum mechanics are based on common interpretations of physical reality. I have no problem at all with your usage of QM in your essay. It is interesting however that Crooks fluctuation theorem of non-equilibrium thermodynamics is essentially a classical, not a quantum analysis. Part of this, I believe, is that work is not an observable in quantum mechanics, and the relevant work averages are given by time-ordered correlation functions of the exponentiated Hamiltonian rather than by expectation values of an operator representing the work as a pretended observable. [Talkner, Lutz, and Hanggi] I'm not familiar enough with England's approach but from what you present of it it appears to be essentially classical.

Although I did not elaborate in my essay, I have in response to questions on my page noted that the field as I hypothesize it senses (and affects) momentum density, and this is very relevant. One could say to me: "You claim that the consciousness field interacts with ions in axons and vesicles flowing across synaptic gaps. Why then would not the same field interact with electrons flowing in circuitry, since the momentum density of an electron is greater than that of an ion or a vesicle?"

An excellent question. Part of the answer is that the charge-to-mass ratio of ions and vesicles makes them less susceptible to EM fields. But the key answer is that momentum density flow in the brain (and even the blood) is in 3-D and the consciousness field exists in 3-D, and our subjective awareness of 3-D is very strongly linked to these facts. Current circuitry (see my paper FPGA Programming: step-by-step) is 2-D, and even the 2-D arrangements of circuits are designed to optimize timing. There is no spatial aspect to computer circuitry, of the sort we find in the brain. If (and it's a big if) we ever reach the point where circuitry (say a nanotube network) could span the 3-D volume (with suitable I/O: see FPGA Design from the Outside In) then I would think it might be possible that a 'super brain' could be built, but this is contingent on the existence of the consciousness field as the seat of awareness! Doing without the field and without 3-D (as opposed to computations of 3-D) is one heck of a task.

In addition to the work I've done on pattern recognition and learning (hinted at in my endnotes) I also covered Steven Grossberg's mathematical model of neural circuits [The Automatic Theory of Physics (my ref.5)] I hope you are so close to finishing your PhD that you have no use for any of this information, but, given your familiarity with my microprocessor systems design you would at least find the info readable, and perhaps even a source of ideas. I hope this discussion stimulates useful thoughts for you.

I would be very surprised if your essay does not win one of the prizes. It is an exceptional essay, and I wish you well in this field.

My very best regards,

Edwin Eugene Klingman

view post as summary

show all replies (3 not shown)

hide replies

Hi Natesh,

The posts in this blog are as interesting a conversation as are in this contest. In particular your conversation with Ines Samengo is most interesting. More on that in a moment.

The wording of FQXi.org's contest is nebulous, unless you realize it is about the MUH of Tegmark. Tegmark's emphasis is about Mathematics. Landauer's emphasis is about Information. Your emphasis is...

view entire post

Hi Natesh,

The posts in this blog are as interesting a conversation as are in this contest. In particular your conversation with Ines Samengo is most interesting. More on that in a moment.

The wording of FQXi.org's contest is nebulous, unless you realize it is about the MUH of Tegmark. Tegmark's emphasis is about Mathematics. Landauer's emphasis is about Information. Your emphasis is about Intention. My emphasis is about how we choose. I would make a hierarchy as shown below:

"Mathematics is Physical"...........Tegmark

"Information is Physical"..............Landauer

"Intention is Physical"..................Ganesh

"Choice (intention from a personal viewpoint) is Physical (maybe), but we can never know it"......Limuti

I did read your essay and honestly I had trouble following it (I did however spot the insulated gate mosfet structures :))

The image your essay triggered in me was Valentino Braitenber's book "Vehicles, Experiments in Synthetic Psychology". It is easy to make the vehicles look as if they had "emergent" goals.

Non equilibrium thermodynamics as treated by you and Ines was interesting. Ines brought out the memory clearing needed by Maxwell's demon to control the entropy (I think I got that right). Perhaps this memory clearing is why we can't know how we choose. For example, move your finger. How did you do that? Do not point to MRIs or brain function. I maintain that you have no direct experiential record (knowledge or memory) of how you moved your finger. I believe the answer is that you moved your finger, but you do not know directly how you did that. Was Maxwell's demon involved? I know this is a bit esoteric, but would like to know what you think.

In my essay I hoped to get across how convoluted the language of determinism and freewill is. Don and Lexi each took a side. However, each also used Unconsciously the other viewpoint during the conversation.

You forced me to think....minor miracle. Therefore this is a super essay!

Thanks,

Don Limuti

view post as summary

show all replies (3 not shown)

hide replies

Dear Ganesh,

I suppose, you like critical examination of your essay. I must confess that I really could not follow the mathematical derivation entirely, may be due to my own limitation. But, I will grant the concluding remarks by you based on those mathematical expressions. I read this essay twice over a fortnight.

I take the following statement as your motivation. "Open physical...

view entire post

Dear Ganesh,

I suppose, you like critical examination of your essay. I must confess that I really could not follow the mathematical derivation entirely, may be due to my own limitation. But, I will grant the concluding remarks by you based on those mathematical expressions. I read this essay twice over a fortnight.

I take the following statement as your motivation. "Open physical systems with constraints on their finite complexity, that dissipate minimally when driven by external fields, will necessarily exhibit learning and inference dynamics."

In Fig.1b, at the first stage we see the external input coming, which is mixed with the prediction of the same coming from higher level, and up goes the 'prediction error'. This is OK, but from the next stage onwards, we see that the predictive estimator (processor/comparator) receives only the prediction error from lower level, and feedback of prediction from the next higher level. A prediction from higher level cannot be compared with the prediction error that took place at the lower level, it would make no sense. A predictive estimator must receive appropriate modular value derived (or predicted) from observation from lower level in order to be able to compare or generate prediction error. I suppose the direction of flow is incorrect. In fact, a predictive estimator should generate a prediction error internally from the predictions coming from both sides, and use the error to predict for the next higher level as well as for the lower level, which must be appropriate for both sides independently. Natesh, in cases of processing systems, always take the limiting cases to test the hypothesis. For example, when the system makes first observation, at the lowest level there is no prediction coming back from the higher level to compute the error. Similarly, at the highest level no prior action to make a correction with only incoming prediction error. Furthermore, this is also to be noted, in any realistic system, a module may receive input from multiple modules and send its output to multiple modules.

"The joint system SA is a quantum system with two components." From this I also gather a classical system might not be able to achieve what a quantum system does, otherwise, there was no need to classify it as quantum. But then, later on I notice that you identify neo-cortex as S, and A as motor-cortex. I trust, you are equating neo-cortex and motor-cortex as quantum systems, a hard to gulp inference.

"Agency is the capacity of a system/entity/agent/organism to act on it’s environment." And if all physical entities satisfy this definition of agency, then I do not see the need of a separate definition taking the attention of some readers on the side of psychology. Being a part of environment, any reaction to the physical context is equivalent to altering the environment. But when you say, "(I am imbuing system A with agency, but not with a specific goal or purpose)", it is as if there could be a system without agency. As you defined earlier, all physical entities are natural agents. So, by stating this, you are priming a reader with certain preconceived notion of agency. Again when you say, "The optimal encoding of R0 in SA is a trade-off between exploiting known information and exploration", where does the exploration come from? I understood that A would simply react physically as per the input from S. But this reaction is aimless. The term 'exploration' also achieves the same goal of priming the reader with certain kind of agency, reinforcing the sense. "While the state of A depends upon balancing exploration with prediction", further enhances this sense.

Even in cases where the system SA is evolving to predict the incoming input correctly, it is just a prediction of the system R, where is the purposeful goals for self sustenance or whatever coming into picture? Therefore, I suppose, one has to design an extra element in SA system such that S tries to optimize on certain parameter, to signal A to act in a particular manner. Otherwise why would S set the task of throwing a ball in any manner, let alone trying to dunk. The purpose also has to be artificially coded.

"Due to these past inputs, let the state of the system A (motor cortex) that is most likely given the prediction-exploration trade off, corresponds to the action "throw the ball." How did the first input come, and what would be any reason to throw the ball at all?

"We will define sense of agency (SA) as the pre-reflective subjective awareness that one is initiating, executing, and controlling one’s own volitional actions in the world."

"Thus the joint state of SA=("see ball being thrown","throw ball") as the ball is thrown will explain the sense of agency, the awareness of an action being performed as it being performed." The association of an awareness of an action being performed to the system SA is in your/our mind. I do not see where and how exactly this sense of awareness is represented in SA. Then you rain statements like, "For example, in the case of visual perception of a face, the higher levels make predictions corresponding to say, 'seeing a face'." I can accept that the system may have representation of all the parts described, but I do not see how 'seeing a face' is represented.

The masterpiece of all statements is, "Similarly predictions made in the higher levels of the hierarchical model in SA, under the minimal dissipation hypothesis, would correspond to the higher level intention of the action-sense of agency (like say "win game" in our example)...."

As I said about your system that goals and purposes would not arise unless especially coded in the system, the same applies to all systems. In a system like brain, such a coding is achieved by the process of natural evolution in the Darwinian sense. You may quote me on any statement here.

Then comes the attribution of ownership, "Crucial to this process though, is a sense of ownership that the system will learn over time about what is within the system’s control and what is beyond that." Natesh, what you could see as a logical extension from your own perspective of a relation of an actor and its acts, you assign that to the system.

"... arguments have been made for inherent intentionality in every perception event [12]. We can view the upper levels of the hierarchical model in the brain as the source of only intentions and make a strong case that intention is physical."

Imagine if we say, "intention is a specification of an information represented in a physical system", then it does not remain physical, yet it has origins in physical systems. But then, if we insist on the paradigm of 'intention is physical', then there must be a way to measure it. Though I trust what you may have meant is 'intention' arises from physical function of the universe, it does not require or depend on any non-physical phenomena.

As a concluding remark, I am going to consider a stone as a system S embedded in surrounding heat bath, the air in thermal equilibrium. A puff of wind blows that applies certain force on the stone, but the stone remains undisplaced, and no exchange of heat (energy) takes place, i.e., the stone dissipated minimum energy. In such a scenario, what learning has taken place in S that it can predict about wind? So, any development from minimum dissipation hypothesis must conform to this limiting case. I suppose, you may require some other constraint in addition.

In an exchange with Ellis, you wrote, "Thanks for a delightful exchange. I am enjoying myself!!" I consider you a system like SA, so which component of S and A is referring to itself as an enjoyer, and which component is being enjoyed? And why would both be claimed to be as oneself?

I feel favorable to consider reasonably well rating for clever usage of the terms so wisely that the reader might end up with the notion of emergence of goals from minimal dissipative hypothesis. Mr Natesh, you are a magician too.

Rajiv

view post as summary

show all replies (1 not shown)

hide replies