Sunday, April 30, 2017

The Brain, Associative Memory, Resonance

Dante Gabriel Rossetti - The Gate of Memory.jpg
By Dante Gabriel Rossetti - The Gate of Memory.
What Will the Lady Be Reminded Of?


The third type of memory is associative memory. This type of memory is more like the memory or intelligence that we commonly know of. It is also called explicit or declarative memory by some. By contrast, sequential memory is commonly known as long term memory or muscle memory or skill or habit. Some call that implicit or procedural memory. Associative memory is notable in how it is triggered to playback. The triggering can branch out in varying directions or associations, whereas sequential memory is more fixed in its sequence, with minor deviations if there is any.

In terms of IFPs (impulse firing patterns), associative memory can be described as IFPs of one memory circuit (recognition or sequential) being triggered to playback by IFPs of another memory and/or sensory circuits. The triggering and the triggered IFPs are either related or similar. They are related if there are strongly linked synaptic connections between them. They are similar if the IFPs of the two circuits share some similar patterns of firing. These associative triggering can be described in more detail by the F.A.R.E. (Feedback, Adaptation, Resonance, Equivalence) framework.


Associative memories are common experience. For example, reading a recipe of seafood salad in Good Housekeeping magazine, Mary suddenly remembers that two weeks ago she had a tuna fish sandwich for lunch. That is associative memory. The association is fish in the seafood recipe and the lunch tuna fish sandwich. That triggering of recall is based on similarity. John reads a post on a social media app about the total rainfall amount in recent Bay Area storms. One of the comments says that although the Northern California’s drought is over after this rainstorm, the water bill from the utility company is still going up. That comment is also from associative memory. The association is water of the rain storm and the water utility bill. The triggering of recall can be based on similarity (water in both cases) or on relatedness (water and the emotion of reading the charge on his bill).

It can also be a series of associations that leads to the final recall, as in the answer of this example: Q: what happened in your life in Sept., 2009? A: 2009?…That was 8 years ago. I was about … 16, so still in high school, a senior. September was the time when the school semester started up again. So I was applying to go to college and preparing to take the SAT test.

In the discussion of sequential memory, we find it to be specific memories whose circuits are formed by neuronal networks adapting to repetitively similar IFPs. Associative memory, however, is not about memory of specific event or sequence. It is about pre-existing memories being triggered to playback from associative linkages.

Associative linkages in the brain are like hyperlinks in the Internet that allows web surfers to jump from one webpage to another. Learning and understanding are based on associations to a large extent. When we learn a new subject like language or economics, we may use mnemonics, images, analogies, parables, or rhymes to help us remember the contents. These devices build associations in our mind. They link one group of IFPs (unfamiliar, un-memorized perceptions) to another group of IFPs (familiar, memorized perceptions). By such linking or associating, we can recall unfamiliar things by tracing back from the familiar. And with some practice, we can achieve astonishing memory feats, as shown by memory athletes.


Astonishing Feats Achieved by Memory Olympiads.

2-time World Memory Champion Alex Mullen Explains How
He Uses Associations To Train His Memory Skills.


So how do memories get linked together associatively in terms of the F.A.R.E. framework? They are linked together by pathways of feedbacks (F) and/or by action of resonance in neural circuits. Let the symbol -> denote triggering of IFPs. The activation of associative memory is a chained triggering that looks like this: IFPs 1 -> IFPs 2 -> IFPs 3. In this sequence, IFPs 1 equals to a sensory perception or a memory. IFPs 2 is the impulse firing patterns in the pathways between the circuits of IFPs 1 and IFPs 3. And IFPs 3 is a memory being associatively triggered by IFPs 1. This IFPs 1-2-3 propagation linkage may be part of larger loops that connect other groups of IFPs together.

To match with earlier examples, IFPs 1 can be the perception of reading about seafood salad in a magazine, or the report of rain amount in recent storms, and IFPs 3 can be the memory of having tuna fish sandwich weeks ago, or the memory of seeing a high water bill from the utility company. The sequence can go the other way as well: IFPs 3 -> IFPs 2 -> IFPs 1. That is, later on while eating a tuna fish sandwich, one may recall reading about seafood salad in a magazine before. With such linkage between them, IFPs 1 and IFPs 3 are called associative memories to each other.

The triggering from IFPs 1 to IFPs 3 is probabilistic. IFPs 3 can be about a tuna fish sandwich, or a video of penguins fishing, or some party meal planning with seafood, or something else. The triggered selection is not always the same, as can be illustrated by the following story. At one time Confucius asked a precocious child which place was farther away from where they were - the sun or the city of Xian (about 1,000 km from the city of Beijing). The kid answered Xian. Why? Because you can see the sun but not Xian. So Confucius praised the kid for his answer. A few days later he asked the child in front of his students again the same question to show them how the kid reasoned. However this time the kid said it’s the sun that’s farther away. Surprised, Confucius asked the kid why again. The kid then said because he had heard of visitors from Xian but never from the sun before.

The child associated distance to visibility at first, and then to visitors later. His answers hinge on association, and association can vary. This is similar to text and context. When associated context changes, the meaning of a text changes as well. The word “good” means differently in the context of “He is good” than in the context of “Good is an adjective word.”

In neural networks, associative linkages can be physical synaptic connections. In that case an associative memory is triggered by feedbacks through that linkage pathways. So, IFPs 2 is the feedback that IFPs 3 gets from IFPs 1. Alternatively speaking, IFPs 1 feeds forward as IFPs 2 and triggers IFPs 3 to playback. But feedback is a more apt description, since IFPs 3 can pick up signals from IFPs 1 specifically (feedbacks), but IFPs 1 may be just broadcasting in general and not aiming specifically at the IFPs 3 circuits (feedforward). A saying illustrates this: 言者無心 聽者有意 - the speaker does not imply specifically; but the listener infers from it. In many cases, the receiver is the active agent in sensing feedbacks; the sender is not particularly feeding forth to specific receivers.

The associative linkage can also be an action of resonance between memory circuits, instead of feedbacks via synaptic pathways. The sequence then is just IFPs 1 -> IFPs 3. The two memories/perceptions IFPs 1 and IFPs 3 resonate with each other due to some similarity in their firing patterns. By virtue of this similarity, a portion of IFPs 1 may trigger a portion of IFPs 3 to playback, and vice versa. However, if there is a transmission medium between the two resonating IFPs groups, whether it is IFPs 2 or something else (neurochemical?), then that transmission medium must also have some similar movement patterns as those in IFPs 1 and IFPs 3.

Resonance is natural. F.A.R.E. framework uses that to explain other phenomena. It is an explanatory principle that we don’t probe for further explanation. If we do try to explain an explanatory principle (or a black box in engineering terms), then we run into mystery. For example, God is an explanatory principle. What is the answer to who created God if God created everything else? Or, if gravitons cause gravity to exist, then what causes gravitons to exist? Soon or later, all explanations must either stop at some explanatory principle, or continues on in a circular loop that comes back to itself.

The word “resonance” denotes echo, a repeat of the same sound. In the study of memory, we use it to mean a repeat of similar movements of any kind. It can be electrical, mechanical, or chemical, and not just limited to sonic ones. IFPs are movement patterns of electrical impulses fired in and around neurons. If there is no similarity between two IFPs, then there is no resonant triggering between them. However, if there is no resonance, nor feedbacks, yet two separate groups of IFPs fire up spontaneously one after another, then that may be a case of synchronicity, a term coined by psychoanalyst Carl Jung.

Synchronicity or serendipity or coincidence is excluded here as a mechanism for the triggering of associative memory, because it’s mystical and unpredictable. Also, it may be the case that it is the interference patterns (holograms) of IFPs that are actually memories. Because the human memory can be holographically preserved when any given part of the brain tissue is removed. That is not explored here either. Here we simply look at associative linkages between IFPs, and use feedbacks or resonance to explain what they are and how they trigger playback of memories.


Let’s take a closer look at the first type of associative linkage - resonance. There are many examples in nature. Since IFPs are electrical in one respect, we start with resonance of electromagnetic waves.

Crosstalk is a form of electromagnetic resonance. It happens naturally in electronic signal wires. A stereo cable with 2 insulated wires carries the left and right channels signals separately. When signals are running in them, the left-channel wire spontaneously picks up (resonates to) the right-channel signal, and the right-channel wire picks up the left-channel signal. This cross-channel resonance happens with wires in parallel configuration, without physical contact or conduction between them. When engineers need to reduce the resonance effect of crosstalk and achieve greater channel separation, they change the physical wire configuration to twisted pairs.

But, instead of reducing the effect of crosstalk resonance, what would happen if we amplify (tune up) the resonance effect? That would allow signals to transmit in a different way than conduction. And that is what happens in the development of antennas and broadcasting technology.

From a TV station to a TV set, programs in the form of electromagnetic waves travel through different media: metal signal wires, metal broadcasting towers, air, metal antenna, metal coax wires, metal TV tuner wires, and so on. In classical mechanics this is described as movement or propagation of electromagnetic waves from one location to another, driven by differences of electrical voltages. In the F.A.R.E. framework, we describe it as the media of metals and air resonating to electromagnetic waves. That is, the broadcasting transmitter is tuned to resonate to signals from the TV camera and sound equipment. Then air resonates to signals in the broadcasting transmitter. And the antenna is tuned to resonate to the electromagnetic waves in the air. The tuning of transmitters and antennas are done by their shapes and orientations, which can increase or decrease their resonant responses to a particular range of signals.

Resonant responses can go both ways. Antennas can either receive or broadcast signals from or to the air. If it is receiving, it is called an antenna. If it is broadcasting, it is called a transmitter. This is similar to the situation of speaker driver / microphone. They both operate by the same resonance principle. Depending on which direction we look at it, we call them by different names. If we look at the diaphragm of the unit, and take its vibrations as a resonant response to electrical input signals in the voice coil, then we call it a speaker driver. If we look at the voice coil of the unit, and take its electrical vibrations as a resonant response to input vibrations of the diaphragm (resonating to air vibrations), then we call it a microphone. Microphones take input of air vibrations (sound) and produce output of electrical signals, whereas speakers take input of electrical signals and output diaphragm vibrations.

Loudspeaker side en.svg
By Altavoz: Enciclopedia Libre - Side View of A Loudspeaker.
Microphones Have A Similar Structure.

Speakers/microphones are classified as transducers. Transducers are resonators that not only echo movements from one medium to another, but also change the movements from one format to another. So when a speaker/microphone resonates, the movements in the diaphragm and the voice coil wire change between mechanical and electrical vibrations. In this regard, our sense organs are transducers also, except that their resonances are one-way. They resonate to movements/differences in the environment and produce movement of IFPs at nerve endings inside the body, but not the other way around.

Is there evidence that IFPs in different regions (circuits) of the brain actually resonate to each other? So far there are only a few studies done on cortical resonance in terms of oscillation frequencies. But for memory IFPs, we mean a resonance not just in oscillation frequencies, but more broadly in the gestalt movement of branching and sequencing of impulse firings from neurons to neurons.

For a medium to resonate, it needs to be tuned to respond to input movement patterns. For antennas, the tuning is done by the shape and orientation of the antenna body. For microphones, it is the configuration of diaphragm, magnet, and voice coil wiring. For guitars, it is the tension of the strings. For musical water glasses, it is the shape of glass and level of water. If resonance of IFPs do happen in cortical regions, then what conditions tune the circuits there to resonate?

Resonance And Tuning

There is a principle called Occam’s Razor that chooses the simplest theory over a multitude of others as the proper explanation for a phenomenon. For resonances, IFPs or other kind, there must be a similarity of movement patterns between the echo and the original. That is the common denominator for all resonances . So the tuning for resonance is at where similarity of movements can be obtained. In two neural circuits hosting 2 different IFPs, the resonant pickup (crosstalk or otherwise) of the similar portion of two IFPs will be much stronger than the dissimilar portions, and so producing a greater resonant effect there. It is simple as that.

If the resonance is so strong in the area of similarity between two neural circuits, then an actively firing neural circuit can induce a silent neural circuit to fire up at that segment of similar IFPs. Then the silent neural circuit can go on from there and run the course of its native IFPs playback. By this process, different IFPs can be triggered to start around different circuits at different resonant regions. It can jump from one circuit to another. And on and on they go ceaselessly. That explains why during zazen (坐禪, sitting meditation) beginner meditators can easily notice their own stream of thoughts, but find it impossible to stop the streaming.

However, the resonant pickup is not always so strong that the echoing circuit can start its course of IFPs playback from there. Otherwise there will be viral chain reactions. One IFPs can resonate several other IFPs circuits to playback. And each of those IFPs will in turn trigger up many others, resulting in a wildfire of IFPs in the forest of the brain. That does not happen in normal brain activities. Our conscious mind is not normally occupied by a multitude of memories or thoughts at the same time. That would be a feverish or frantic mind. So similarity is a necessary but not sufficient condition for a resonance to be high-powered enough to cross a threshold. To make a resonance strong enough in the resonant circuit so it can precipitate its remaining IFPs to playback, other conditions are needed.

What other conditions can boost the level of IFP resonance? Maybe it is the presence of some neurochemicals/nutrients, and/or multiple other low-intensity ambient IFPs that also contain that similarity, or related to it. Then the resonance can be amplified by quality and quantity. If those conditions are met, then one memory/perception can instigate another memory/thought via amplified resonance, cascading onward to become a stream of high-intensity IFPs playbacks.


By Chrissshe, Emotiv Pure*EEG Software
14 Probes, 14 Traces. 1 Trace = 1 Aspect of an IFP

As shown in EEGs, there are multiple streams of IFPs running in our brain across various regions. Most of them are subconscious as our consciousness is an awareness of only a few thoughts/memories/perceptions at a time. What is consciousness then? Could it be just those IFPs that are the most prominent ones, dominating over other IFPs in terms of firing magnitude/scope/duration? (“scope” here refers to the number of other IFPs in symphonic play with the dominant IFPs)

The idea that consciousness is high-intensity IFPs at play and subconsciousness is IFPs of lower intensity comes from the revelation in an YouTube video. The inventor Tan Le is showing how her remote control headset works. (This headset is now sold under the name Emotiv Epoc+.) A) The headset picks up thought patterns (IFPs) in the form of voltage or some such modulations that the brain is prominently radiating. Upon receiving and identifying an unique modulation pattern, the headset can then issue commands to control the movements of a remote object. B) The user wearing the headset focuses his consciousness on an unique thought pattern that matches the one the headset is programmed to respond to. Or he can turn his consciousness away from the unique thought. So, by A) and B) together, consciousness = dominant IFPs.

Consciousness = Dominant IFPs (Unique Thought).

Thinking and imagining can be described as variations (mutation, combination) of playing back memories. That will be in a future article.

MRI and fMRI

In recent decades, advances in neurology and brain science have been closely tied to advances in MRI and fMRI technologies. MRI, or magnetic resonance imaging, can echo the contours of inner body parts and let us see them without having the body cut open. The imaging is done by resonances of different water molecules in different cellular structures to magnetic signals. If scientists want to see not only the structures of brain but also what it is doing, they use machines called fMRI, or functional MRI. That machine is tuned to echo certain contrasts in blood cells and in water molecules in the brain. From contrasts in resonant responses of water molecules come the images of brain structures, and from contrasts in resonant responses of oxygenated blood cells come brain activities, because the presence of oxygen-rich blood cells in an area indicate more neuronal firings there.

Some of the above examples show that systems without internal energy (passive system) like metal wires and antennas can have resonant responses that allow signals to jump over from one medium to another. However, the magnitude of such resonant response is small. Crosstalk in audio wires amounts to an echo of less than a thousandth of the original signal magnitude. But systems with internal energy (active systems) like MRI machines can have resonant responses that are much greater. By design and by internal energy supply the resonant responses can be amplified to a level that is needed to motivate further responses from other components inside or outside the system.


Let’s look at some examples of resonance in living systems. Living systems have internal energy supply to boost their actions and reactions. Their resonant responses are still either a straight echo (as resonator) or a transformed echo (as transducer).

In the biological world, imitation is a resonant reaction. When we hear a song we “like”, or see a “cool” dance, we sing along or dance with it. When a wolf howls in the wilderness, so do other wolves nearby. The expression for this is “monkey see, monkey do”. Also, we may involuntarily laugh when we see others laugh, or bow when others bow, even when we don’t particularly like to. But we do it anyway, unless there is something that inhibits us from doing so.

Spontaneous camouflage by chameleons and octopuses is a form of imitation. Their skin can change colors or even texture in less than a couple of seconds to resemble the nearby background. How do they do it? One theory is that it happens by the resonance of nervous signals. There are photo sensors and pigment actuators (chromatophores) all over octopus’s body. When the sensors on the near-the-background side of the body pick up the color information of that area, this information becomes IFPs in the creature’s nervous systems. That sensory IFPs are then resonated in the pigment actuator circuits on the away-from-background side of the body. The result is that an octopus’s skin looks like the background and thereby making itself invisible.

Master of Camouflage - Octopus

Camouflage may have an emotional component that is a response to the danger of predators. If emotions (hormonal responses) are involved, then hormones may be a shutter or trigger of the camouflage process, since neurochemicals can affect impulse firings of neurons very much. As humans, we can and do learn to control our emotions and body language as we adapt to cultural conditions. That will be using IFPs to direct the release of hormones. We may be able to artificially laugh or cry in tears by practice willing it. The act of gating emotional expressions itself may be a resonant response to other people’s controlling of their emotions. With repetition, emotional control can become a sequential memory, a habit. And human actors train themselves to excel in this skill of acting out controlled emotions.


Not surprisingly, emotions are expressed outwardly near sense organs: the eyes cry, ears prick up, lips curve, and skin blushes, etc. As such, our five senses are almost like transducers that transfer and transform information bidirectionally. When sense organs portray emotional information from the inside outward, other observers detect our temper and feelings by the look of our skin, eyes, mouth, etc. When sense organs resonate to information from the outside in, the brain picks up changes and variations of the world around us. But the direction of resonance in sensory information really goes just one way - from the outside in. The inside-out passage is not really a resonant response but a reaction.

For eyes, the inward resonance movements start from outside optics, which are contrasts, gradients, differences, or changes of electromagnetic waves in the visible spectrum. The optical nerve endings of retina resonate to those outside electromagnetic differences and produce movements of optical IFPs. After that, other neural circuits in the brain receive or resonate to these optical IFPs. This will result in a perception of sight, and possibly invocation of prior memories of similar or related sight.

With ears, the process is similar: acoustic movements in the environment (air/fluid pressure differences, modulations) -> auditory IFPs in the cochlear nerve endings -> IFPs in the cortex as perception of sound -> IFPs of prior memories involving similar sound -> actions/reactions IFPs motivating other parts of the body -> …

In general, propagation of information (a difference that makes a difference) can go on indefinitely in different media via resonance or feedbacks. Media may transform information in its propagation. And sense organs are media that resonate to information travelling on feedback pathways to the sense organs.

There is a saying that everything is connected. Some call it interconnectedness or interbeing. Others say that all things can be no more than six degrees of separation apart. Such connections are mostly not ropes or hooks that tie objects together physically. Rather, they are informational connections that take place in sense organs and nervous systems. In other words, we can perceive and imagine connections of things whether things are physically connected or not. Without IFPs taking place in the brain and sense organs, the connections of information will be lost. That disconnect will be similar to the state of nirvana, where all connections to suffering are gone. But in nirvana all other connections are gone too - joy, confidence, jealousy, anger, ignorance, knowledge, self, world. So sometimes nirvana is translated to quiescence.

The Buddhist Heart Sutra classifies the mind as a sense organ. It is an organ that senses dharma (Buddhist teaching). It should really be called the Mind Sutra. The text makes the correspondence between sense organs and sensory perceptions as: “...No eye, ears, nose, tongue, body, mind; No color, sound, fragrance, taste, touch, dharma…” (...无眼耳鼻舌身意, 无色声香味触法,...). The author probably meant the brain as the sense organ for dharma, since brain is a physical organ but mind is not. But then the brain’s functions were unknown to people at that time.


In ophthalmology, the study of eyes, eyes are classified as a part or an extension of the brain. That means the boundary of brain is a sensory layer, or at least some part of the boundary is. In a branch of computer science called deep learning, AI (artificial intelligence) machines have been built to accomplish amazingly intelligent identification tasks. These tasks can be computer vision or speech recognition or medical diagnosis or playing chess. The intelligence part of these machines are similar. It is a set of hierarchical computation layers that mimic neural networks of the brain. Each computation layer has sensor elements (artificial neurons) that transform information from the previous layer and passes that on to the next layer. However, the outputs, which are mathematical transformations of the input information, at these layers are not echoes but biases of the input. So they are not high fidelity sensors, but more like biased sensors.

How do such biases accomplish the feat of intelligence that can identify things? It is because those biases “push” the information in the direction towards identification and not somewhere else.

Dr. Li fei-fei describes the state of computer vision.

Each biased sensor layer has a different type of information, starting from the type of raw data at the first layer and ending up the type of labels at the last layer. Take computer vision for example. The input data fed into the first bias layer of the machine is of the logical type pixels, or picture elements. Their values may be halftone (dark-light) or RGB (red-green-blue) or CMYK (cyan-magenta-yellow-black). The output of the last layer is of the logical type labels, or names we use to call objects. Their values may be yes-no, orange/non-orange, or other names. The changes of types and values through the layers are like a metamorphosis prompted by biases rather than by genetic codes, as in biological cases such as caterpillars/butterflies.

While we associate logical types to the information passing through the layers, the layers themselves make no such association. They simply follow rules of computation to transform number-coded information from input to output at each layer. How do these computation rules accomplish the act of intelligent bias, and correctly turn values such as 001011111001…. (halftone dots) into a value of 1 or 0 (yes/orange or no/non-orange)?


Before we get into how this bias computation works, let’s first look at where these biases come from. These biases come from a calculus method that does reverse engineering. Called backpropagation in the AI (artificial intelligence) community, this calculus method compares an actual outcome of the data with the pre-identified correct outcome and uses that difference information to retroactively adjust and refine the biases for the layers. It shifts the study of machine intelligence from looking for features and meanings in the data to setting up biases and computation pathways in the machine. And it worked amazingly well! The door to a new era of AI was opened by it.

What are these biases? They are just some multiplier numbers coupled to the data, which are passed between sensor elements (artificial neurons) in adjacent layers. They are also called weights, coefficients, or bias factors. How the biases work can be shown in the details of how they are set up, which is done by a broader process called supervised learning. Three things are involved in supervised learning. 1) a massively large sets of pre-identified data to train a neural network model how to bias, 2) a neural network model of layers of connected artificial neurons awaiting for bias factors, and 3) run the big data sets through the model and use the backpropagation (also called gradient descent) algorithm to seek out the best bias factors across all layers of that intelligence model.

Multilayer Neural Network.png

By John Salatas - Multilayer Artificial Neural Network For Supervised Learning

The action of backpropagation algorithm is similar to that of negative feedback control. It is an adjustment process that aims at reaching a certain target. Backpropagation uses a difference information, the deviation of an actual outcome from a reference, to calculate small adjustment amounts for the bias factors, which are initially set to some random numbers. That adjustment will produce a reduction on the deviation between the reference and the next outcome. Then the reduced deviation, also called an error function, is again used at the next round to calculated another set of small adjustments on the bias factors, and so on.

Such progressive adjustments are iterated many times until they have reached or are very near the goal. At that point deviations from the correct outcome are forced down to a minimum level. Therefore, the biases can transform input data into the expected outcome because they are mathematically set to take the input data on a minimal-deviation ride to the correct output.

While backpropagation adjusts bias factors in deep learning machines, negative feedback control adjusts something else in its circuit. The car cruise control is an example of negative feedback control. A cruise speed (reference) is set on a smart gadget and it will drive the car at that speed for you. What that gadget adjusts is some mechanical actuators that change how much gas is fed to the engine and/or brake force applied to the wheels. It progressively adjusts the actuators to minimize deviations of the car speed from the set level. The outcome of these adjustments, which are done very quickly, is that the speed of the car will converge to the set speed and stay close there.

A cruise control enables a car to drive at a fixed speed regardless whether the terrain is a flat road or hill or curve. Similarly, bias factors enable an AI machine to produce the correct identification outcome, such as the name of an animal, regardless whether the input data is an image of that animal sitting, jumping, or looking sideways.


While it is not surprising that the bias factors, after being optimized by the supervised learning sessions, can be utilized back to identify correctly the images from those sessions. It is surprising though that the trained machine can also identify new images outside the training data sets. And in some cases it can identify as well as humans can. How do the bias factors accomplish such intelligent task? Where do they get that intelligence from and what is that intelligence?

The hypothesis here is that artificial intelligence is a resonance effect due to similar movement patterns of information propagating through the biasing circuitry.

By Akawikipic, AI Robot Created by AKA Intelligence

1) Suppose that an artificial neural network machine is trained by the backpropagation method with hundreds of similar yet different cat pictures. After the supervised learning sessions it can correctly identifies most if not all of those images as cat(s). What is the relationship between the trained bias factors and those images that were successfully identified?

2) The bias factors (weights) do one thing only to the output data of sensor elements (artificial neurons) passing from one layer to the next: they either emphasize or lessen the data value being passed. In doing so they produce a meta-information, or information about information. Layer 2 produces a meta-information of layer 1, and layer 3 produces a meta-information of layer 2, or meta-meta-information of layer 1, and so on. An example of meta-information is the map-territory relationship. A map is some information about the information of a territory. And a territory is a meta-information of some geological areas and man-made structures. The passage of picture data through a smart identification machine is a metamorphosis of information in different cognitive stages (layers). They go in as the logical type of pixels and come out as the logical type of labels.

3) The bias factors are adjusted till they produce the same id outcome on all those different pictures of cats. With one picture, there are many possible sets of bias factors that will work. With two pictures, there are less choices of proper bias factors that can work for the double amount of data points. With hundreds of pictures, the likelihood of coming up with one set of bias factors that will work for all becomes increasingly impossible. Yet not only is the finding of bias factors not thwarted by the increase of pictures, the quality of these biases factors are actually improved by the increase - the accuracy of id becomes better. This suggests that the training process is not just about finding bias factors that work for all the data points of all the pictures, but it is about finding biases that will draw out what is common among the datasets of pictures, and then bias those commonalities into a common outcome. If commonalities exist among the pictures, then this approach will not be dead-ended by problems associated with the increase of picture data.

4) Each time a training picture is changed, the previously worked-out bias factors must be readjusted for the new set of picture data. This is a different kind of training than the backpropagation calculus. The picture-change training is a meta-adjustment. It adjusts the backpropagation adjustments done on a single picture. From this meta-adjustment comes bias factors that emphasize what is common among the pictures datasets and diminish what is not common. This emphasis/reduction is due to the common-outcome constraint imposed by the different yet similar pictures.

5) Suppose there are N pictures from the supervised learning sessions that result in one set of bias factors identifying all of them correctly. And we make a data chart for each of these picture identification run. Each chart contains the output data numbers from all the neurons in all the layers. This compilation shows the successive identification progression on one picture’s dataset. Having these N data charts from the N training pictures under one set of bias factors, we can compare them side by side. There will be a similar movement pattern on the data numbers going through the layers. That common movement pattern across the charts is the effect of filtering, of meta-adjustment.

6) The data movement through the layers should be a pattern where, out of the initial dissimilar data at the first input layer, some similar data values (meta-information) appear at some later layer(s). This can be deduced backwards from the last layer. Let’s call the last layer the layer L, the before-last layer L-1, and the second-before-last layer L-2, and so on. The outcomes of the L layers of all charts are all 1 (it’s a cat). This common outcome constrains some output numbers at the L-1 layers across the charts to be similar. They need to be similar in values, especially the ones coupled to weightier biases. Otherwise the weights of the bias factors between the L and L-1 layers will magnify their differences to be large enough to change the outcome at the L layer.

7) Just as the common outcomes, viewed across the N charts, at the L layers put a constraint on the L-1 layers, the similar outcomes at the L-1 layers place a looser but same kind of constraint on the L-2 layers. The variations of values at the L-2 layers across N charts will be larger than the range of differences allowed at the L-1 layers. Going back this way to the first layers, the outcomes there will be quite diverse across the charts as they were originally. Then, going forward from layer 1 to layer L, the pattern of the data movement is that the output data values of the layers become progressively more and more similar across the charts.

8) This data movement pattern is enforced by the meta-adjustment of changing pictures. Some call this movement “extraction” of features embedded in the data. But that extracted commonalities, the features, are not in the input data itself, but in the higher-level meta-data of later layer(s). The similar data values across the charts at the L-1 layer (or maybe even earlier layer) are those “commonalities” which represent similarities in the appearances of cats. Maybe they represent features or relationships such as outline/contrast, ratio/proportion, topological order/sequence of connection, or something more abstract. Anyway, if the meta-data at that layer are outside that range of similar values, then they probably don’t reflect the common look of cats.

9) So this provides a context to describe the intelligence of the trained machine. It can identify a cat picture outside the training dataset correctly because its intelligent bias factors were set up to make certain features or meta-data to emerge from the picture data. If the meta-data of the new cat picture are close to the common meta-data of the training cat pictures, the machine then reaches the correct identification by following through (bias further) with these common meta-data to the common outcome. Alternatively, the intelligence for correct identification can be described as a resonance effect. The movement of data through the machine’s biasing circuitry, from data to meta-data to meta-meta-data and so on, are similar for those pictures that have similar features of “cat-ness”. They are similar in terms of the emergence of similar meta-data and outcome. This is of course not physical movement like that of sound or electricity, but rather an abstract movement of information. By this similarity of abstract data movement through the biasing circuits, the machine exhibits intelligence for identification of new cat pictures.

10) The meta-adjustment of changing pictures is like unsupervised learning. The resulting bias factors must move the data of similar cat pictures in a similar way, as well as making that way coming out with the same outcome. The second part is done by supervised fitting of backpropagation. The first part is unsupervised because there is no reference as to what that similar way is. Is it just by a high number of iterations during supervised learning (process of elimination) that the matching set of bias factors emerges? Or do they come by schemes of unsupervised learning? Perhaps advanced computer vision machines incorporate unsupervised learning. Anyway, currently there are a few models of artificial neural networks that can achieve unsupervised learning, among them adaptive resonance theory (ART) and self-organizing map (SOM). These models mimic the interactive activation and competition behaviors of biological neurons that transmit sensory (visual or auditory…) information in separate brain cortex areas.

State of the Art Computer Vision. World Memory Champion Wang Feng vs. Robot
Identify an Adult Out of 30 Elementary School Class Pictures

Anyway, resonant responses exist in nature abundantly. We can engineer it to make musical instruments or medical equipment. It is one of the two ways that can trigger associative memory to playback in the brain. The other way is feedbacks. We will look into that in the next article.