Cognitive Architecture for Direction of Attention Founded on Subliminal Memory Searches, Pseudorandom and Nonstop

Burger 1 Cognitive Architecture for Di rection of Attention Founded on Subliminal Memory Searches, Pseudorandom and Nonstop by J. R. Burger History The term architecture implies a model not onl y for input-output beha vior but for physical structure. Physical structure refers to si ze and weight and also to known electrical and logical properties. Neuroscience aside, the cen tral goal of cognitive arch itecture is to build a practical thinking m achine, that is, an intelligent agent th at ‘reasons’ like a human [1]. Towards this end, it is desired to im plement cognition in general and n ot just the detection of a particular signal from the out side world. Preferably, cognition must occur in a timely self-contained way as it does in humans. Since architecture is supposed to model brai n behavior in logical detail, a cognitive architecture must address information read ily available in psychology texts, including short term memory. A valid architecture m u st have memory that is capable of holding for a brief time encoded signals from the se nses, emotional signals, and signals created by recall from long term m emory. In a cognitive architecture that models phys ical structure, subconscious long term memory must be associative; any inform ati on may be recalled instan tly once proper cues are found in short term memory. Inform a tion is committed to subconsciou s long term memory by a process termed ‘rehearsal’ in short term m emory. We note that gifted individuals have photographic memory that la tches instantly, suggesting that long term memory is unrelated to synaptic growth. Background search capabilities Everyone has experienced trying to remember something, but being unable to do so. Unknown to you, a search proceeds within your subconscious long term m emory for what you are trying to remember. When you least expect it, possibl y at an inconvenient moment, the correct information will pop into short term memory with amazing clarity. This is an indication that the brain has an ab ility to work in the backgroun d without being noticed, adding or subtracting cues fr om a sear ch until the cues are exactly right, as the y must be for recall. Memory searches clearly serve beyond m erely remembering the forgotten. There is a theory that decisions are made not by ‘free will’ but by a search of past sim ilar situations held in memory. Electroencephalography re lating to the timing of finger movements, pursued by Benjamin Libit and others, indicate s that choices are often m ade in the brain before a person realizes it [2, 3]. Surprisingly, the brain seems to retain control. In other words, it is not the other way around, in which a ‘person’ makes a decision, and then tells the brain what to do. Burger 2 Along similar lines, it may be noted that the brain appears to search itself continuously in the background not only for forgotten facts a nd situations, but also for solutions to problems. A problem in this context m ight be a hard problem with no logical solution. Problems like this often use random trial so lutions. In analogy, dreams and daydreams are what we experience as the brain attempts to solve difficult or im possible problems by random search. The qualification, of course , is that many difficult problem s have no good solutions. Randomness has advantages. In order to retr ieve forgotten mem ories, including solutions to hard problems based on past experiences, it is efficient for cues to be selected pseudorandomly. Pseudorandomly chosen initia l values are commonly used in numerical analysis to solve difficult optimization problems. Since random starting points are helpful in computer science, they might also be helpful for brain memory searches too. The importance of search suggests an architectu re in which a cue editor w orks tirelessly to recall random ly related information to a sublim inal level. Here it undergoes analysis unconsciously. It is noted here that subliminal recalls may occur at a rate of tens per second as permitted by neural circuitry. A pers on is aware only of the most ‘im portant’ recalls, the ones permitted into short term memory. A cognitive architecture is now synthesized that includes ps eudorandom memory search a nd sublim inal analysis. Cognitive architecture incl uding subliminal analysis Figure 1 illustrates a system of associativ e memory whose blocks can be related to specific neural circuits that are easily s ynthesized, since neurons achieve arbitrary Boolean logic [4]. Short term memory neurons may be explaine d as having digital outputs like any other logical neuron, but their dendritic pulses ar e a little longer (hundreds of m illiseconds) compared to the typical f ew milliseconds. Short term memory neurons support longer lasting dendritic pulses because of a shortfall of internal potassium or equivalent in their dendrites, causing an extended pul se burst within the axon. Long term memory neurons may be explai ned as digital neurons that transfer neurotransmitters from boutons back into dendri tic receptors. The result is a digital read- only memory (ROM) neuron that can be set in stantly, and will latch indefinitely, as long as the neuron exists, unless cleared b ecause of lack of use. The indefinite cycling of neural signals in a neuron is physically possible b ecause neurons can be modeled as adiabatic, that is, re quiring practically no calories for electrical signaling beyond what is required to sustain any biological cell. Words in memory are assumed wide enough to accommodate every possible feature a human can experience, shape, shade, tone, sm ell, feel, emotional strength and so on for thousands of elemental features. Each feature is defined by its logica l location in a word. Features are encoded from the senses by a large com plex neural network labeled the Sensory Encoder . If a novel combination of feat ures is used repeatedly, the Need to Burger 3 Learn detector calculates that, if possible, the encoder should ‘learn’ new features. The need to learn box may be based on digital filters, as proposed below for m emorization enable, although they are not designed here . As an example of learning, the color chartreuse might be learned as a dig ital AND combination of yellow and green. Combinational le arning like this (no analog para meters) is ak in to implicit or reflexive memory via new synapses and serotonin; it is assumed driven by a need for efficiency in successful species who cannot afford to waste time first recalling y ellow, then green, then a mental definition of chartreuse. Feature 11 Fea tur e 21 Fea tur e L 1 Se ns ory Inp uts Cue Edito r Memorizatio n Ena bl e Se ns or y Enc ode r Long T erm Mem o ry Wo rds Feat ures Index of Importan ce Moto r O utp u ts Pse udo Rand om Sea rc h Ne ur al Stat e Ma ch in e s Subl imin al Ana ly zer Short term mem ory wor d Reco gni ti on . . . . . . . . . . . . . . . . . . Feat ure 1 1 Feat ure 2 1 Feat ure L 1 Feat ure 1 N Feat ure 2 N Feat ure L N Short Term Memory Enable Need to Lear n Detector Figure 1. Cognitive unconscious archit ecture including subliminal analysis. Certain motor outputs are ini tiated under feedback contro l within the brain, giving a sensory response, such as hot coffee on the lips . But there is a type of procedure known as an unconscious procedure that executes automatically. Humans easily execute many such procedures, similar to long poems learne d by ‘heart’ to be perform ed mindlessly. Once initiated, one simply brushes teeth, driv es truck, or adds ch eckbook without a lot of pondering, as an aid to survival of the human species. Burger 4 Unconscious procedures are a major aspect of learning, although, as with combinational learning, we cannot yet synthesize neural circuits that enable such learning in practice. Unconscious procedures are conjectured to be the result of intern eurons that synapse between words of long term memory, forming a neural state machine. Neural state machines are efficient in that procedur al steps avoid passing through the processing associated with short term memory. Long Term memory Undeniably, the most important part of l earning is subconscious long term memory. Long term memory is more influential than m an-made memory with its limited address fields and keywords, in that any subset of features (except a completely empty subset) in short term memory can used as cues to retr ieve old images. Thos e neural paths pointing down in the figure can deliver cues but also se rve for applying features to be m emorized. Those buses pointing up in the figure bring forth previously memorized features for evaluation in the subliminal analyzer, but only if the cues match exactly. Multiple matches are de alt with as in any engineering design via a sim ple neural circuit, so that two separate images cannot be recalled at exactly the same instant. The first recall to reach the subliminal analyzer is assumed to be the one that is evaluated first; other matches are ignored. If the first recall is not correct, meaning its index of importance is relatively low, another subset of cues, slightly different from others in the sequence, is immediately placed on the cue bus as explained below. The subliminal analyzer is assumed to alternate im ages from the sense s with subconscious recalls. Although the circuitry fo r this is not shown, it is simple enough. The sensory encoder is disabled while info rmation is being recalled from long term memory. Sensory information in this model is analyzed same as recalls. Cue Editor The cue editor in this architecture is e nvisi oned as in Figure 2. All cues are assumed called into and taken from short term memory. But these cues are inconsistent when an image cannot be remembered imm ediately. So cues are appropriately masked in a pseudorandom way as shown, using a neural shif t register counter, t ypically fed by neural exclusive OR gates. Counters like this can produce a unique subset of cues. Resulting associative recalls will have some correct features, but not nece ssarily all the right features; recalls can be analy zed many tens per second. Subliminal analyzer The analyzer has the task of determining an index of importance for each subliminal set of features. Digital signals from long term m emory or the senses appear on interneurons, and are re-encoded as suggested in Figure 3. Note that encoders are not necessarily simple and have yet to be synthesized in a re alistic way. Using iden tical neural circuitry, the digital contents of short term memory are re-encoded into an index of importance; we note that as short term memory fades, im portance drops, so new thoughts are expected. At any given time, these encoders assign a dig ital value to recall-rela ted neural signals. A subliminal image whose index app roaches th at of current short term m emory will be Burger 5 permitted to replace the current conte nts of short term memory (assum ed to be the same as consciousness). In this way, short te rm m emory accomplishes what the theorist Richard Semon termed direction of attention . xN xN-1 x2 x1 xN' xN- 1' x2' x1' Pseu do ran dom Co un ter (S elf-t imed Shi ft Regi ster ) Sho r t Term M e mory Wo r d Cu e Sel ector Gat e s To L ong Term Memory W o rd s 1 2 . . . N N Figure 2. Cue editor showin g pseudorandom counter block. Importance is not conflict resolu tion; it is evaluation for: 1) Brightness of sensory images, 2) Magnitude of emotional content, 3 Quantity of matched cues, and 4) Recency of experiences. From Short Ter m Memo ry From Sublimin al Int ern eu rons Encod er Encod er N N Enabl e Replace ment of Sho rt Term M emory Digita l Com p arat or m M M To M emori zat io n Ena bl e Figure 3. Calculation of importance. Trials proceed at many tens per second. Occasionally there is recognition , which is a feeling associated with a particularly cl ear memory with many m atched cues. Memorization enable The availability of blank mem ory words to hold new information is assumed unlim ited. Memorization in this architectur e is triggered by a memorization enable block which is sensitive to recurring images in short term memory, that is, rehearsal. In the example circuit in Figure 4, conditi ons for committing something to m e mory are true if cues are presented but there are no ma tches, or recalls. Additionally, if a given Burger 6 image, as identified by the above importance encoder, appears in short term memory twice, separated by a given delay, it will be com mitted to long term memory. The delay can be implemented by short term neurons in a standard digital filter arrangement. A simple neural multi write cir cuit ensures that only one word is programmed for a given memorization enable. Cue s but No Mat che s Memo rizatio n Enab le Sign al N N Cu es Matche s From Sh ort Term Memory Enco der Digital Equal it y Delay N m M M Figure 4. Proposed memorization en able circuit using digital filter. Conclusions The above cognitive architecture is notew orthy because it in cludes pseudorandom memory searching as an ongoing process. Cu e subsets are selected pseudorandomly, as many as possible per second, so recalls alternate with sensory data, if any, for subliminal analysis. An index of importance is computed for each image by encod ers especially for this purpose. When the index of importance for a subliminal image approaches that of current short term memory, a transfer occurs. A new set of attributes is thus enabled to enter short term memory, with a new set of cues, thus defining a new ‘thought’ and a direction of attention. Neurons are easily specialized to have short as well as long term memory properties, so, since neurons are capable of arbitrary Boolean logic, and since there are trillions of them, sophisticated digital circuits are possible in the brain. An advantage of the above architecture, from the view of designing an intelligent robot, is that it depends not on meaningless symbols and undefined analog para meters, but rather on neural gates that have hardware equivalents. References [1] J. Anderson, The architecture of cognition , Harvard University Press, 1983. [2] Daniel M. Wegner, The illusion of conscious will , MIT Press, 2002. [3] Ray R. Hassin, James S. Uleman and J ohn A. Bargh, Editors, The New Unconscious, Oxford University Press, 2005: Ap Dijkst erhuis, Henk Aarts, Pamela K. Smith, The power of the subliminal: On subliminal pe rsuasion and other potential applications. [4] J. R. Burger, Explaining th e logical nature of electrical solitons in neural circuits, , 2008.