“Necklace” unsupervised learning algorithm


#1

There are some entities which are easily comprehend by almost everybody, yet are very difficult to describe in details or give strict science definition. That is true for almost any term from the field of our intelligence, and term intelligence itself. Thought, feeling, understanding, consciousness, learning, recognition and the interest, the sense of beauty in image, form, melody, and the sense of humor. It is hard for intelligence to formulate these entities, same as it is hard for a snowman to answer “what is snow?” question.

I feel that “Interest” is the one among these entities, that is undeservedly deprived of attention from AI researches. And the one that I am long interested in.

In the 4 pages pdf paper below you can read my thoughts on the interest and how it can help to build unsupervised learning HTM.


#2

Hi Aleksandr

I have been reading your questions and your latest white paper.

May I propose this?
As you I have spend many years trying to understand, describe, and relate many classical concepts from cognitive neuroscience. Most of these old classics like learning, memory, attention, motivation, and as your work on “interest” have their origin from William James, the first professor in psychology in the US. He wrote a famous dissetation and in here he used these words, most of them translated from Aristoteles and other ancient greek scientists. James is famous for writing "everybody knows what attention is…but of course they don´t and “everybody” has had and has the same problems as you have with “interest” in understanding and transforming it to a concept that can be used in matematics= in a calculation using a measured number. Bringing the concepts that explain something we can observe as humans, but have difficulties in trying to measure unambiguiously, is necessary if we want to make thinking machines.
As I have been in these problems of understanding, expressing, measuring etc. for the last 30 years, I ended up reengineering the brain from “one bit yes/no or start/stop” into the entire Human Decision System ™…this system remarkably ended up with an explanation of the function of the six layers in neocortex (= separating information sources in time: Now/start, moment, present, future, past, doubt/stop), and it gave nice interrelated definitions of all William James and others brain functions and learning processes. This logical reengineering is also a logical reconstruction of the information processing necessary to combine sensing and movement based on differences in input and changes in output under risk. The analysis showed that traditional neuroscience in its concepts for modelling are missing several rather important facts about the brain and the nervous system:
The brain and the nervous system is in its essence a measurement system (as is also the core of physics, there is always a measuring agent in physics, that in the end interferes with the object under measurement). This measurement system is removing doubt (=entropy) to lower the risk (=probability of several types of faults, each connected to each information channel=signal producing channels in time). The neocortex is thus a decision system transforming doubt into decisions handling the automated machine of subcortex (it is mylineated contrary to neocortex). This involves probably four states of the neuron: Passive, active, predicting (inference) and in doubt,

In The Human Decision System “interest” is the same as "attention. Doubt is operationalized into six categories following each other when the decision has been made (probably the synchronized firing in the neocortex between the layers), they are: Where,what,which,why,when,how, in that sequence which minimizes entropy…here what=awarenes, which=attention, why=focus, when=fixation and where is the allocentric system and how is the egocentric system.

The Human Decision System ™ thus explains what must go on to explain from where to how and the decision that is performed in behavior (muscles moving)…from that conceptual level, we can then start transforming it to bits moving around…a classic code transformation that Numenta is trying, not that I agree they describe and understand all the necessary functions. One thing is that doubt (=entropy) is not a concept in HTM…though all interneurons are doubt neurons…! The consequences are, that many functions will not be described and represented…but Numenta is the only enterprise in AI taking biology serioulsy and that is what is important…

Hope you would like some kind og dialogue…since keeping the grand overview while detailing must be the road forward…
Regards
Finn Gilling


#3

Greatings Finn.
I never heard about William James, I will read futher.
I see that concepts are pretty complicated.
What it is easy to see is that difference:
I don’t consider “interest” in humans same as “attention”. Attention is more common mechanism.
It is very important to discern simple attention that is present in reptilian behaviour, and intelligent “interest” that controls playing behaviour in mammals. Interest using attention mechanism as a newest evolutionary development using eldest.


#4

Greating Aleksandr

Yes I understand you think that “attention” is more “general” …analysis reveals that the general concept of “attention” has to be dissoluted into the four concepts “awareness as answer to what”, “attention as answer to which”, “focus as answer to why” and “fixation as answer to when” - this means that you must leave the general notion of “attention” and dissolve it in four different types of “general attention”: This is mentally tricky, because thinking of four concurrring “general attention” processes is disturbing, and actually there are six, including the allocentric and egocentric “general attention”: The brain must always be ready to answer “what to do now”, and this has to be done with many competing answers to the questions:

  • where - me in some environment/allocentric
  • what/awareness of possible targets in the environment
  • which/attention to possible goals
  • why/focus on a specific goal/target combination
  • when/fixation on a specifica goal/target combination
  • how - me doing something in the environment/egocentric

If you doubt what I just wrote: It is beyound doubt that the brain online runs the allocentric and ego centric processes. The are run by where and how and produces answers to these two questions.

Then how can you explain that the brain should not produce answers to the other four questions? Especially because they are tied closely to targets and goals, and movements, the three processes that are started when you turn a decsion into action (=the combination of task(goal-target) and activity)…the concept on “interest” has often been used in a famous model called “AIDA”, attentien interest desire action, but these four concepts are ildefined: You awareness to have attention…look at a lion and turn your head and look back with ATTENTION on the lion…are you interested in the lion then? Yes you are…your interest is about if it fundamentally is attracting or repulsing you (like food, people, snakes etc…)…so interest is a function of attention…and in the same doubt/question category of “which” - it is a competition between say a lion, and another lion…(this is shown in neuropsychology in TVA-theory, theory of visual attention by Claus Bundesen)…the “interest” is about making a choice of the next goal: move towards or move away…

Did that answer some questions?

Regards
Finn


#5

Hi again
I can add a crucial question: How does the brain/neocortex separate data from different time periods and still become able to use them online in the “now”?

Anybody have an answer?

Regards
Finn


#6

Finn,
I see you wrote a book on your theory. Do you have a downloadable PDF summary (perhaps an article somewhere, or an interview) that gives the general idea?
Thanks.


#7

Hi gidmeister
Yes I wrote a book titled “The Human Decision System”. The book starts with a case: E. coli eats glucose and lactose. It can uptake glucose and process it immediately for energy. It can also uptake lactose, but needs to produce an enzyme to digest it. Thus invest energy to get energy!. It cannot sense how many lactose molecules that are outside the cell body. So it uptakes app. 300 molecules, and then throw a coin playing 50/50 if it should uptake and take the chance there are more outside or not. The E.coli is confronted with a “decision problem” since it as more that one way to decide about “what to do now”. This “what to do now” i the starting point in the long chain of arguments the book presents. Because this is what the brain is supposed to do, always being able to know what to do now. The finish the E. coli nobody knows if 300 molecules can the energy of starting the enzyme production, but what is sure is that evolution has developed decision algorithms, or better, decision systems, to make a one cell organisms able to make self-selection planning its future. So as the first scientist I coined the concept of “biological decision systems” that makes self-selection for fitness possible. Soon after I was able to describe the human version from the general principles in biological decision systems. Even stem cell scientists have learned that nature and nurture cannot alone explain the differentiation of the stem cell, and in Copenhagen University they have a center for “stem cell decision making” trying figure out how to control stem cell differentiation. So here we are with a general concept of “biological decision systems”. Why do these systems exist? We have the genetic system to transfer information from species to species, we have the epigenetic system to transfer information from generation to generation, and then it is straight that we have the biological decision system to transfer information during the individuals lifetime. Soon I realized that cognitive neuroscience has missed a very important concept at the foundation of neuroscience. Basically neuroscience says that there are sense neurons, interneurons and motor neurons. As you see the sense and the motor neurons are named after their function, but what is the function of the interneurons? Recalling Shannon, interneurons introduce uncertainty. Since a sense neuron connected directly to a motor neuron is a reflex, no decision process there, only 1:1. But the interneuron introduces a time delay (a memory function) and it introduces an uncertainty as to what the response will be (a decision system). Thinking on the E. coli, the number of senses, the number of food items and the number of movements that an organism (or deciding agent) has, defines its total number of degrees of freedom. But the number of algorithms in the decision system determines how to split these degrees of freedom between entropy and negative entropi - between questions without an answer and questions with an answer, at any point in time. So the brain (and nervous system) is an algorithm producing organ, a decision system, that handles how the deciding agent should decide about what to do now, now, now…
This introduces two dimensions in the algorithm: Spacing and timing.
Spacing is about how to move, ie. what sequences of movements that will be the answer to “what to do now, now, now…”.
So what is movement? Well we can split action into “activity” = the degrees of freedom used on moving the muscles. And then on to the task that can be split in the goal and target: The moving agent moves to achieve a goal engaging a target. So when you make a decision about what to do now, this will happen:
Start/stop achieving a goal
Start/stop engaging a target
Start/stop moving
Thus there are three start/stops behind any action.
Where do they come from in the brain? These processes must be handled to decide about the start/stops:

  • you must find out where you are
  • you must find out what targets are available to you
  • you must find out which goals are available to you
  • you must find out why you should pursuit a specific combination of goal-target
  • you must find out when you should pursuit this goal-target
  • you must find out how you will pursuit this goal-target with a movement

The above processes are derived if you reengineer a decision and a movent into its decision parts (degrees of freedom). The brain is also divided into these independent processes that become linked to each other when the agent is moving.
For example the Nobel prize in neuroscience went to the Norweigan Moser couple for describing the “where” process with its place cells and the allocentric system. Same is with the “how” system which is the egocentric system.
The sequence of entropy (questions) that has to be converted intro negative entropy (answers) is:
Where, what, which, why, when, how.

Then about timing the movements to the now, now, now:
To find answers to the questions, the brain needs to be able to process bits into information. But the brain also learns.
What does this mean? Learing is the process from raising doubt (asking questions) to automating answering the questions. This automation will then function as a reflex, since the entropy from the interneuron will be sent to 0.

So if we understand one learning process as the query process (ie. how to raise questions - this is hardcoded in the brain) and the other as automation (when the motor system just makes themuscle move without any interruption) then we have four other learning processes in between, that brings a question into an answer:

  • habituation, when the agent is able to repeat to stay in control
  • sensitization, when the agent is able to differentiate to regulate
  • operant conditioning, when the agent is able to modify based on a assumption
  • conditioning, when the agent is able to restrict based on anticipations

This can be translated into the different time periods from where the information in the learning process comes from. Remember when you burnt your fingers on the hot stove? Now you dont touch the hot stove. You have been conditioned. So you use information from the past to restrict your behavior now based on an anticipation that if you did, the situation would repeat and you would end up in pain. The point here is that information from different time periods can only be used in certain wyas in the now. These are:

  • to start/stop /now (6)
  • to control/moment (5)
  • to regulate/present (4)
  • to modify/future (3)
  • to restrict/past (2)
  • to query /freeze (1)

These basic ways of using information from different time periods in the now, is what the neocortex is able to do. I have out on the layer numbers. As you see layer 5 connects directly into the motor system, and layer 4 connects into the thalumus. This is what is actually taking place, and layer 2 and 3 is about inference (prediction).

As the pyramidal neuron can be in three states according to HTM theory:
Active = firing - this layer 6
Passive = not firing - this is layer 1
Predicting these are layers 2,3,4,5

  • 2 is from the environment
  • 3 is feed forward
  • 4 is feed back
  • 5 is the soma, the moment the neuron is in when it is predicting

This is like ready, aim, fire…

So the trick is to describe the formulars and the specific bit processing that bring the neuron in the different states.

Did that explain something positive to you? Ask me any question…

Regards
Finn


#8

Finn,
Some of the functions you mention are getting integrated into HTM by people outside the company - for instance Ali Kaan Sungur has made a video game with goals as part of HTM, and Daniel Rehman is integrating egocentric and allocentric perspectives in his model.
But it seems to me that you would have to make very specific suggestions on modifications to HTM to get your ideas tested.
As an aside, I remember a journalism student once telling me of the questions every journalist should ask (Who? What? Where? When? Why? How?), or the questions that are referred to as the five Ws and one H.


#9

hi gidmeister
My “ideas” are, as I see it fundamental:
The timing from entropy, past, future, present, moment, now (= automation in sub cortex), are observables with degrees of freedom you can manipulate…ok I am not a logician or network specialist, but these ideas are scalable from the pyramidal neuron, to the six layers to macro behavior.
Yes I read Daniel Rehman´s paper, a very god job done.
Would you be interested in diagramming all functions, because I think/know from presentations, that the model contains al the functions on the fundamental abstract level, and may be we could ask the other two guys to join? We could target a paper with all our names on?
Regards
Finn


#10

hi gidmeister
I have one question I could not find answer to in htm theory:

The feedback to the soma, I think it is the basal they call it…is that through the axon?
Or what is the pathway? I understand they mean it is a synpase on a dendrite but it is pictured as it is on the pyramidal neurons axon…that confuses me…

A feedback on htm? I followed a discussion saying htm did not have an back prop, but a feedback is a kind of back prop…what is your opinion?
Rgds
Finn


#11

Finn,
I don’t think that HTM has feedback yet, though the sensorimotor model allows for optional feedback from the top layer. As far as diagramming, I’m not the right person to do that.
Regards,
Gidmeister


#12

The feedback used in SMI model is initiated by active cells in the output layer, transmitted from their axon to the apical dendrites of cells in the input layer. This biases the cells that represent feature/location pairs for a particular object, allowing multiple sibling input layers to communicate and “vote” on what object is being sensed.

@fine2100 I’m not sure whether that was the type of feedback you were referring to though.