Eutherm Ltd, Poznanska 129/133,
Ozarow Mazowiecki, 05-850 Poland
w.galus@eutherm.eu
Preprint of an article submitted for consideration in International Journal of Machine Consciousness © 2011 copyright World Scientific Publishing Company http://www.worldscinet.com/ijmc/ijmc.shtml
Received Day Month Year
Revised Day Month Year
wersja polska:
http://mindconsciousness.wordpress.com/2011/03/06/pamiec-epizodyczna-%E2%80%93-klucz-do-myslenia/
Abstract
Intelligent machines should be equipped with motivation to broaden their knowledge via their own experiences and to infer on basis of previous experiences. Thus, they should possess self-consciousness that would allow them to express higher mental functions such as reasoning, curiosity, understanding, stubbornness, bravery, etc. This paper presents a flowchart schematic for an artificial brain that would aspire to realize the higher mental functions that are present in the human brain. Also presented is the significance of episodic and association memory plus their interrelations with the inclusion of short-term memory, working memory, associative memory and long-term memory. A structure for the episodic memory is presented alongside its place in the architecture of emergent connectionist network. The paper also shows the inaccessibility and uniqueness of the internal language of neural correlates of mental events and difficulties in creating hybrid architectures.
Keywords: self-consciousness; cognitive architectures; curiosity; machine reasoning; understanding; ; episodic memory, schematic of artificial brain, mind
1. Introduction
An artificial cognitive neural network should realize mental functions similar to those of a natural mind. In specificity, the cognitive networks should be characterized by a rich repertoire of behaviours and thought processes typical for animals or humans. Cognitive networks that aspire to achieve self-consciousness are characterized by motivation – understood as the necessity to understand the reality – as well as by curiosity, imitated by proper mechanisms of environment and memory exploration, as described in Machine Mind [Galus 2011]. This should, however, not be taken as a statement that any artificial network can possess a psyche that is comparable, or beyond, that of a human being.
We need to start the analysis by focusing on the ways in which the mind registers stimuli from the senses. In the quoted work was presented a recognition and categorization method for visual stimuli. The proposed schematic was described and acknowledged by many researchers and appears sufficient for analysis of the issues of motivation and the mechanism that activate the memory in order for the mind to carry out cognitive functions. However, when faced with the activity of a individual specimen we come across the problem of recognizing and registering an actively changing environment. Applying that to visual observation of environment, we see static images that change in a dynamic fashion, episodes composed of series of images that resemble film clips. The higher fields of the network, responsible for creation of qualia, are therefore not excited by a stable configuration of neurons but rather by rapid changes of these configurations occurring within the input nodes, where the dynamic of the changes themselves depends on the rate of the input sensations.
We know, that both, human beings and animals, possess ability to remember the observed episodes in a storage area called the episodic memory. This is an important location on most flowcharts presenting the mind’s operations. However all models for episodic memory, employed until today, have been far removed from the actual operating principles of the episodic memory in the animal brain. Although the importance of episodic memory function in the cognitive processes is well recognized, the actual mechanism of storing the episodes has not been researched sufficiently. This paper presents a hypothesis of how such processes may be realized; how to complement the architecture of artificial neural networks, allowing them to register and recall the required episodes; and what would be the consequences of such model for the psyche of neural networks organized accordingly.
2. Requirements for the cognitive network
It seems obvious that in order to achieve higher mental functions similar to those occurring in the human brain, the network should be connectionist with an emergent architecture [Duch, 2010]. It should meet the requirements set forth in the Machine Mind [Galus, 2011]. A special characteristic of such a network is its multilevel, hierarchical structure, high capacity memory and feedback loops that allow excitation of the descending hierarchy of the neural cell fields. Apart the fields and layers that constitute the hierarchic structure, specialist areas should also be designed – responsible for the functions of: speech, emotions, reasoning, logic, geometric and mathematical transformations, timekeeping and planning as well as motor fields, controlling the body, speech apparatus and sensors. A neural network modified by synaptic reinforcement seems quite inadequate when faced with the need to record a quick, massed, great inflow of data that needs to be stored in the episodic memory. The neuron cells should be accompanied by memory cells capable of permanent change of state under the influence of a singular electro-chemical stimulus provided by the network. Such role could perhaps be realized by astrocytes that are believed to serve an important function in the information processing taking place in the brains of humans and animals. In any case, the nodes of an artificial network should consist of cells capable of introducing permanent changes in the transmission parameters of the network. Sampled temporal segments of each episode, much like movie frames, must be registered in subsequent memory cells. If we remember that each percept and scene is accompanied by a significant tree of cell group excitation that ascends to the higher layers of the processing hierarchy, we come to conclusion that registering subsequent configurations of excitations of these cells will result in the memory being flooded by a giant torrent of unnecessary information. The way in which images are constructed in a hierarchic, connectionist network results in a remarkable ability to compress the information recorded. As the processing pathway passes through fields and groups of cells that are specialized in recognizing singular aspects of the analyzed image, in the subsequent configurations of excitations responding for the sequence of time segments, the only elements to be registered will only be these cells that recognize the changing aspects. The remaining configuration sets, associated with the observed scene will remain unchanged, and when the episode would be recalled, they would create a complete and dynamically changing rerun of the episode in cooperation with cells responsible for registering changes.
All information reach the Central Executive (CE) – the highest field in the semi-hierarchical structure of information processing. In many of the available works the CE is tasked with the role of situation analysis i.e. thinking, associating abstract notions, planning, motivation and goal setting, concentration and control over the effects of undertaken actions [Baddeley 2000, Starzyk & Prasad D.K. 2011, Nuxoll & Laird, 2007, Edelman 1992]. Such wide range of tasks that CE is forced to fulfill, results, that no specific structure was singled out that would effectively realise all these functions. In the works referenced above as well as in other attempts to propose working cognitive architectures, the consciousness functions have been moved to this central field without providing information on how exactly are they to be realized or stating that it impossible to achieve such functions at the current state of the art of technology.
A paper by Starzyk from the University of Ohio [Starzyk 2011], analyzes the mechanisms for concentration of attention and describes systems for effectors and location control. An attention switching mechanism located in the CE, based on competition of excitation configurations by stimuli reaching the association memory, being a part of the CE, would select the winning mental representations on the winner-takes-all (WTA) basis, creating in effect, a stream of consciousness. The episodic memory serving as a basis for the learning processes, would be filled with such pre-selected data provided by the semantic memory. The “Planning and thinking” block would be responsible for planning and thinking, triggering conscious attention focusing, recalling of episodes, controlling effectors and would cooperate with the block responsible for goal definition and motivation. While we can imagine a stream of conscious registration of stimuli that reach the level of comparison and selection, still unknown is the actual aim and way of recalling required episodes by the consciousness that has to be already emerged before it may serve its selective and regulatory functions.
3. Schematic of a brain-type neural network
Armed with the hypothesis regarding the operating principles of the episodic memory and taking into consideration the need to understand the mechanisms of curiosity as described in Machine Mind [Galus, 2011] we can attempt to propose a certain network architecture that would realize functions of the human brain, demonstrating self-consciousness and other complex mental states. A schematic for such a network is presented on Illustration 1.
Fig. 1. The ascending signals are shown as solid lines, descending signals as dashed lines (the schematic shows only the primary excitation pathways between the operative blocks). The Center lines show repetitions of blocks in vertical direction (to determine the level of processing), and in horizontal direction (for denotation of sensory modalities). The separate sensory blocks that constitute levels of processing A1, … A4, … Ak correspond to fields V1, … V4, … Vk in the human brain. The brain interacts with its environment via sensors and effectors. The Body, Internal Parameters block is a collection of sensors that analyze the physical state of the network or agent body in the case of embodied intelligence. The internal structure of blocks is presented on Illustration 2.
The most important part of the presented schematic is the network of connections between the separate areas of the “brain.” An especially important and, sadly, neglected element is the network of backward, descending connections which are responsible for the curiosity mechanism. It should be remembered, however, that the connections presented, show only a chosen set of excitation transmission and that they represent clusters of neural connectors that sometimes consist of millions of neural strands that connect individual processing fields, which allow groups of cells – network nodes – to transmit the excitation states.
Individual areas constitute a semi-hierarchic structure only in functional terms. They could be incorporated into a uniform network structure with similar characteristics. In specificity, the CE area does not have to constitute a uniform whole. It could be divided between other areas of higher mental functions – uniting their actions. The actual localization may be dependent on the requirement to minimize the length of connections and their optimal cooperation.
The division of areas into layers is in itself symbolic. The sensory and perceptive fields have a classic neural network form as they serve cognitive functions. Their multilayer structure is required in order to allow cell groups to realize functions that in the human brain are realized by a singular neuron or a group of a few neurons. The long-term memory cells are not differentiated as a separate layer in artificial neural networks but rather accompany the association and short-term episodic memory cells. Perhaps they should serve a function similar to the function of astrocytes in the animal brains. The association memory cells realize a comparative function, i.e. they calculate the correlations of excitation configurations that arrive to these cells from the lower levels.
A fragment of a cognitive neural network corresponding to a specialized block is presented on Illustration 2.
Fig. 2. Creation of episodic memory in typical memory; time sequences are changed into spatial arrangement.
Excitation at moments t1, t2 ….tn are registered in cells c1, c2,… cn located in the vicinity. They are registered in a sequential order, when a cell of a lower number has changed its state by registering the previous excitation. Strong or recurring stimuli are recorded by the long-term memory cells lm1,… lmn that correspond to association memory cells. Excitations are filtered through the association memory and transmitted out of the block to connected structures in a way presented on Illustration1.
Separate layers on Illustration 2, designated with numbers 1,2,3,4 represent specialized layers that connect “neuron” cells – nodes that serve specific functions. Looking at the functions served by the cell groups in their respective layers, we can differentiate them into:
1 – Input Layers (IL)
2 – Short-term Memory (SM) and Short-term Episodic Memory (SEM)
3 – Association Memory (AM)
4 – Long-term Memory (LM) and Long-term Episodic Memory (LEM)
Short-term episodic memory cells, presented on Illustration 2 as “n”, Long-term Episodic Memory, presented as ln , and Association Memory cells Cn are separate neuron cells localized in artificial cerebral cortex in accordance with the localness hypothesis. They are connected in such a way that an excitation can be forwarded to subsequent, neighbouring cells. The excitation propagates automatically to further long-term and short-term episodic memory cells which results in “displaying” of the recorded sequence. However, for this to occur, the episodes need to be properly registered beforehand in these memories. This takes place thanks to specific operating principles of the association memory cells. Therefore, a hypothesis is proposed that in order to achieve the required effect, the operation of the cell groups must be as follows:
- Stimuli received by the receptors form object patterns that are recognized in the input layer (layer no 1) that creates a cognitive neural network.
- At the highest level of the network, that we shall call operational short-term memory, the recognized patterns responding to subsequent time segments for a given tn moment are recorded. The set of neural cells being a neuron correlate of these changing patterns will undergo changes so it will be different in any time moment tn.
As in a short time period most elements of a scene do not change significantly, the majority of neural cells will remain in their previous state. The changes will affect only the part of cells responsible for recording of the changing elements of the scene. Illustration 2 shows these cell groups as squares denoted tn in the SEM layer (layer 2), their designation matches their corresponding time moments. - All stimuli configurations from the SEM layer should be transferred to the next cell layer characterized by its further specialization. Their operating principle is that if subsequent signals do not change for a longer time, their corresponding excitation pattern is recorded in the long-term memory in the LM layer. Regardless from this process the signal is forwarded from these cells to other areas (blocks) that are placed higher in the processing hierarchy. However, if there is no correlation between the excitation patterns (which means there are changes in the excitation configurations), the new configuration is recorded in new memory cells designated as Cn. Due to their ability to recognize the level of pattern correlation and the ability to forward them on the basis of existence or lack of existence of such correlation, these cells are called Association Memory (AM, layer 3). If the state of neural excitation in AM is maintained for a longer period, no matter the reason, it must be recorded in LM. If a group of neighbouring cells lm records a sequence of subsequent configurations corresponding to a changing scene in a series of time moments, such memory will be called the Long-term Episodic Memory (LEM, layer 4).
A group of cells created on the basis of the above operating principles may realize a pattern-recognition and recording function. Based on the method of correlation recognition, the patterns may be recorded in long-term memory or be transferred to higher fields in the processing hierarchy. It should be noted that in the proposed model the episodic memory cells tn do not stand out from other short-term memory cells and in a similar manner the episodic memory cells do not stand out from other long-term memory cells ln. In this model the episodic memory is created using long-term memory cells typical for a given structure. Moreover, it should be noted that such a network changes the temporal sequences into spatial sequences, recording data from a series of subsequent moments in neighbouring memory cells. The recalling of a sequence transforms the spatial distribution into a sequence of scenes that are played out. Similar method may be utilized for recording mental processes dealing with logical, abstract or mathematical thinking that would constitute specific types of mental episodes. Such recorded information could be accessed by the system, at will, much like episodes from its? individual history.
Processing areas consist of similar blocks – arrangements of cells that in their entirety resemble a cerebral cortex. Just as the CE block, localized among other specialized blocks of an artificial mind, described above, the remaining blocks can, hypothetically, constitute a single layer of cerebral cortex divided into specialized areas. The semi-hierarchical structure presented on Illustration 1 consists of only a system of connections plus the order of transmissions between different blocks. Apart the actual order of transmitting the ascending signals, these areas are differentiated by their joining to sensory organs that deal with both the environment and the state of the body/network itself. This way, the fields that are joined to sensors create sensory fields with their topographic and retinotopic maps. Depending on the type of sensory input defined by the sensors connected there is a range of modalities of sensory inputs (Sensory Input block). Part of the artificial cerebral cortex that is connected to the effectors should, therefore, serve as a basic motor cortex (Motors, Effectors, Actuators block). A specialized part responsible for effectors would be in charge of the speech apparatus and the orientation and tuning of sensory receptors. If the artificial brain is provided with a full body, it should be conscious of the state of its body. This would provide possibility to use well established motivation systems such like the drive to minimize “pain”. If embodied intelligence would be able to take into consideration many parameters defining state of its “body” it should be possible to create a hierarchy of pain, establishing an excitation scale related to the severity of danger signalized by the pain. However, the dominating motivator would not be the physical pain. Pain signals should be transferred to the Central Executive, where they would be compared with other signals and, after being “understood”, a decision would be made on the strategy of avoiding that pain.
Feedbacks covering the fields of emotion (Mental Pain & Emotions), body (Body & Pain) and motor (Motors, Effectors & Actuators) functions would be responsible only for quick reactions to sudden dangers, as a form of a survival instinct, without actually engaging the consciousness. The cognitive functions would be realized mostly at individual layers of sensory data processing areas.
The long-term memory of the system is dispersed and localized both in the CE block, as far as past episodes are concerned, in the Speech & Language block, as far as semantics and syntax is concerned, plus in other specialized blocks in regards to logic, reasoning, geometric and mathematical transformations. The episodic memory of the lower processing fields corresponding to brain fields V1… V4 and higher ones, consists of components of episodic sensations that are utilized when episodes from higher fields are recalled. The long-term memory ensures the permanency of recognized data transfer pathways in layer 2 and permanency of selected data transfer when in accordance with the expected pattern by the association memory in layer 3 to higher layers of processing. The pattern becomes permanent when:
- the stimuli and excitation are very strong. The excitation may result from sensory signals from the sensor fields as well as from internal sources –signals from other processing fields.
- the stimuli are frequent. The repetitions may result from frequent sensory signals or from return signals from higher processing layers – from cells that lay on the pathway for excitation pattern transfer to CE.
- There is a synchronisation of stimuli with signals transmitted intentionally or non-intentionally from the higher layers, which conditions recording in long-term memory which in turn provides additional reinforcement for a given stimuli while limiting other stimuli that fight for the system’s attention i.e. being forwarded to higher processing layers.
There should be a possibility for the ascending path recording process should be complemented by formation of a parallel descending pathway. Long-term memory cells can be fitted with such a function, although there is no research indicating that such a correlation is present in case of animal brains. Discovering relation between the realization of association functions and the long-term memory is of vital importance for neurological research. Much promise is associated with the role of astrocytes in these processes.
Consciousness is emergent within the Central Executive block. The Self & Understanding block is responsible only for registering the state of understanding and self-consciousness that are, in fact, located in CE block. In a similar manner, all the higher mental functions may be recognized in the association memory of the CE block and recorded as “experienced” in the cerebral cortex block named “Self.” The abovementioned intentional signals that coordinate long-term memory activation are governed by a conscious mechanism of “attention.” When analyzed from the viewpoint of selection of different configurations of mental states, this means that the correlated CE cells send strong return, descending, signals to lower fields, generating a strong excitation that is chosen by the selection and transfer by WTA mechanisms back to the higher levels. The correlation between proper CE cells and groups responsible for configuration recording in the lower fields are created in the process of learning and experience collecting which in turn results in connections (pathways) recorded in the long-term memory. The same signals when send spontaneously due to excitation that is not associated with conscious states, would mean a mechanism for unconscious “attention switching.”
4. Episodic memory – operation
The episodic memory allows for recording sequences of events (scenes) or sequences of states. For that purpose, the brain has to divide a given event into shorter time segments and record them in subsequent cells corresponding to the temporal sequence based on the timeframes of these segments. Some of research seems to indicate that the subsequent cells are in fact neighbouring cells, which would be in accordance with the localness hypothesis, presented in the Machine Mind paper [Galus 2011]. In most cases, scenes are differentiated only by sparse changes in their constituting images. The scenes are recorded in neighbouring, subsequent cells. Therefore, these cells have to be joined by a serial connection so that excitation of one of them results in activating of a whole series of subsequent cells. This results in capability to access whole episodes at will, and replaying them in their entirety. There is also possibility of initiating the recall from a chosen scene in the recorded episode – a scene that is most crucial for the aspect that is being analyzed at the conscious level. After such excitation, the subsequent, serial cells would be activated in a configuration corresponding to recorded episodes, forcing intentional or coincidental “remembering” of past sensations. The exciting signals from other blocks would not have to reach the working memory but could excite recall of configuration episodes directly from the long-term memory.
With such assumptions question arises concerning the faithfulness of recording and recalling of a stored episode. This concern both the temporal resolution, connected to sampling frequency, as well as spatial resolution, or to be more general – sensory resolution. To register visual sensation the sampling time for individual episodes must be short enough to create a feeling of continuity of the image. On the other hand its speed cannot be shorter than the time needed to process the image in its dynamically changing aspects. Most research data on this issue is based on research of the human brain. It seems that in case of humans the sampling time is equal to registration from few to a few dozen frames per second. Shorter sampling time is required for the sense of hearing and is connected with the ability to differentiate frequently recurring sounds. These times equal the frequency of a few hundred Hertz. Other senses do not require such great sampling frequencies.
The spatial resolution of the recalled images is remarkably variable. We know that when reminiscing of a landscape we at first see a general shapes, colours and dominating objects. We also direct our attention to dynamically changing elements. This means that CE is reached by signals from the association memory. Taking a well known landscape as an example, we can assume that CE will consciously recognize the beach, water and sun over the waves. Plus a flying seagull. These sensations will be generated in a sequence, as a result of two variables that force changing patterns of stimuli to be transferred to CE. The first variable is discharge the sequence of short-term episodic memory cells. The waves of changing excitations will be recognized as the movement of the seagull over the horizon. The second variable is the switching mechanism of attention, that focuses on the main elements of the landscape, allowing us to see, in turn, sand, water, horizon and sun. This mechanism is described in a paper [Starzyk at all 2011] is in accordance to the postulated competition of excitations [Baddeley 2000]. The order in which the elements are perceived could be regulated by the mechanism of mental saccades that was also analyzed in the previously mentioned work [Starzyk 2011] and that resembles the saccadic eye movements.
If we focus our attention on the details (which means that the competition mechanisms select new objects of attention, focusing on them), we will be able to perceive smaller details, like: the breaking of waves, footprints in the sand or the movement of the seagull’s wings. What mechanism allows for such a change in resolution of the recalled episode? This is possible due to excitation of the deeper memory structures in fields responsible for recognizing scene details. Therefore, we can expect, in relation to the natural brain, a system of return, descending connections, from the higher processing levels to the lower, closer to perception fields. A similar system should be included in designs of artificial systems. This type of backward, descending connections was described earlier [2] and is connected to creation of patterns for comparison within the “curiosity” mechanism. This name was chosen because this mechanism is responsible for penetrating and exploring of the memory, i.e. realizing cognitive functions that allow introspection, reflection and, through association, understanding of new phenomena arriving from the sensory fields. This excitation of episodic memory may be conscious, as the excitation signals that trigger recalling of episodes originate in the highest layer of CE in association with other notions, possibly originating in other areas. This could be an area responsible for processing other modal functions, e.g. the hearing and speech recognition block, which is responsible for the fact that when we hear the words “beautiful summer vacation” certain semantic memory structures are excited and the excitation signal is forwarded through internal vertical and later descending connections to the episodic memory. Similar mechanism can invoke even more detailed elements in images from the episodes recorded in our memory, which we can call sub-episodes. However, we cannot be certain whether given sub-episodes are in fact elements of the episode that evoked them. This is because there is a possibility of exciting sub-episodes from a completely different ”fable” as long as they seem to match our memories and as long as our CE accepts and “recognizes” the final, artificial episode. If our memory has a sub-episode consisting of the sound of seagulls’ cries, it can be used as a sub-episode in different episodic recollections.
Sub-episodes can be excited by descending connections, visible on Illustration 1 and denoted by letters Ak , Bk , Ck where ABC stands for sensory modality and the k index is the level of pattern analysis. This pattern can be correlated with the level of detail of a given scene. This is when an interesting question arises: how to choose the k level? Two hypothetical possibilities need to be addressed. The first assumes that we excite all the hierarchy levels simultaneously. In this case the consciousness is provided with a level of detail chosen by the pattern selection of the competition mechanism (equal to attention switching mechanism presented in [Starzyk, 2011]). The second possibility is choosing the k level by binding the k descendent pathway to its corresponding ascending pathway, as it was described above. Such connection must be recorded in the long-term memory and can be created via learning processes, i.e. recording of events. It is possible that both possibilities are at work in the animal brain. The current understanding of neural processes and the role of astrocytes in these processes does not allow us to infallibly determine the ways in which association and long-term memory cooperate and what is their exact relation to the episodic memory.
We can see that the described operation of recalling episodes with varying level of detail creates different needs for working memory in which it happens. The brain can be said to protect itself from overload. Episodes that we seldom recall are eventually forgotten (except their elements that are frequently used as sub-episodes) and usually, after some time, we remember only static scenes that are animated only by imagination or verbal narration. If these memories are not used, they also eventually disappear.
The mechanism for forgetting episodes is as important as the mechanism for their recalling. The short-term episodic memory of the sensor areas is bombarded by gargantuan amounts of input data. The competition and inhibition mechanisms select only a fraction of these inputs to be processed and recorded in long-term episodic memory. The working memory must be continually emptied to allow for new sensations to be taken into consideration. After the information is recorded in long-term memory, the association memory can be made available for future operations. As one of the determinants of recording information in the long-term memory is repetition of the stimuli in given time periods, the association memory has to record these information for at least the duration of the said periods. The timeframe for recording data in the long-term memory is dependent on the life strategy of the embodied intelligence. If the memory resources are limited, the forgetting mechanism, activated in relation to unused data may be beneficial. The strategy for forgetting, that was an evolutionary achievement for animals, we have to determine on the basis of economical analysis.
5. Communication with the cognitive network
Recognizing these remarkable characteristics is however impossible without first establishing a language for communication. Similar problems are present in approaches trying to discover and assess the level of consciousness of living organisms. We are witnessing an endless discussion on self-consciousness of apes, dolphins, elephants, dogs, cats and other species, although they possess similar to humans: social behaviors, body language, somatic reactions and even facial expressions. This should, theoretically, make conducting appropriate experiments easier. We can but imagine, what would the communication problems be in relation with an entity that would resemble nothing that we have ever seen, if such an entity is not provided with a proper language beforehand.
It seems that the contemporary level of speech analysis and simulation offers a possibility of equipping a neural network with capability to understand speech and to use it consciously in contact with other entities. This capability must be developed as a learning process. The network’s ability to recognize and categorize objects must be used in parallel in two modalities. There must be presentation of objects and simultaneous naming of the said objects. The mental representations of objects and their names, recognized at the level of cognitive functions, will be simultaneously transferred to CE, where created will be a universal pattern for a recurring new excitation configurations related to these objects. A similar situation of repeating sequences of words that coincides with certain operations on the presented objects, may result in creation of patterns corresponding to syntax relations between words. If these sets of symbols expressed as sequences of phonemes will be fed to the speech generator, the network will gain ability to express its own thoughts. This way a new symbolic language, expressing the mental processes within the network, will be created.
Using a complex symbolic system will allow the network to create abstract ideas, with a higher level of generality. This requires a conscious operation on sentences of that language, sentences that also are episodes recorded in the episodic memory of the speech centre. Activating these lingual episodes may be achieved by exciting the input layers of the speech centre by signals originating in CE on the basis of associations with ideas being processed in CE structures. A different possibility is a conscious focusing of attention on uttered sentences or following a verbalized logical reasoning. In such a case, lingual episodes will be fed to CE and possibly connected to logical symbols or additional emotional motivators (like ambition, stubbornness, fear, etc.). This will force the network to search for complex configurations (via “curiosity”) connected with ideas defined by the speech centre. This way a conscious fulfillment of curiosity will be demonstrated, which will result in conscious cognitive actions.
6. Internal communication – language of mental states.
External, symbolic knowledge must be translated into the language of neural excitations. The language of excitation configurations is individual for each and every network, as it stems from that network’s experiences: the set of experiences or even the order in which they were gained. Moreover, natural neural networks are characterized by unique configuration of neural connections. Due to this, every network deconstructs and processes an image or episode in a different way on the intermediary levels. In consequence, there is no universal system for recognizing and categorizing of notions or ideas by the network. Every network creates its own individual language of mental configurations, which, due to high complexity of networks and individual history of gaining experiences, self-teaching and teaching, makes learning such internal language impossible.
This is why an automatic translating system for external, encyclopaedic knowledge remains an impossibility. Knowledge may only be translated via a process of presenting the said knowledge via the senses. It must be “lived through” by the system and connected to previous experiences. The process of such translation will therefore be similar to the processes of learning exhibited by the human brain. Despite the artificiality of the network and seemingly easy access to its mental states, the learning process will be just as slow.
This will be true especially in relation to new, complex ideas that will require individual approach and patience in order to find a form that would be accessible for every individual network. Lack of universal internal language makes it impossible for the entities of the external world to communicate using this language. However, it is possible for the networks to create a universal symbolic language that will allow them to express their mental states, just as it happens with humans. This language will allow communication, coordination, learning or social action.
Individual language of mental state configuration serves as a barrier against intruding into a human personality. Every conscious brain will be forced to create its own model of the world and its own viewpoint, stemming from the personal experiences, learning and self-learning, pondering and internal, verbalized discussion.
Even if we copied the exact same neural networks, filled with preconceived knowledge, such approach does not guarantee creating uniform personalities. Intelligent entities, just as their independent “life” starts, will begin to gather and categorize differentiated experiences which will determine their form and function. On the other hand the infallibility and processing speed of machine memory, no limitations as to its capacity and development of knowledge generalization levels may result in emerging of an intelligence and self-consciousness far suppressing the human ones.
7. Conclusions
The proposed architecture of the cognitive neural network will be capable of categorizing the recognized objects, recording and recalling episodes from its own existence/past, logical thinking and understanding its environment. What is most important, the network will possess motivation to act, to actively care for the comfort of its operation and to gather and record knowledge due to its own curiosity. The ability to act intelligently will result in self-consciousness and many other higher mental functions. These spectacular effects result from the ability of the connectionist network to recognize and categorize, from its ability to record its own experiences and sensations in the episodic memory and from the ability to recall these memories at will. It seems that these singular abilities may be well aided by complementing them by symbolic knowledge that humankind has gathered in available databases.
However, the issue is not as simple as it may appear. Networks with a symbolic architecture are able to express words and ideas recorded in their memory. The operation however resembles re-playing of stored data. Humans are also capable or reading a given text. The ability to mechanically speak the words with or without understanding is not a great feat. Human brain, equipped with connectionist architecture is able to, at any given moment, activate the mechanism for recognizing and understanding of the words being read. This ability is unknown to a network with a symbolic architecture in which word-symbols are not processed into an individual, internal language of neural configuration excitation. Association memory has in such a case no basis for correlation. The “Self” block receives no signals of understanding. There is no self-consciousness.
The example of the human brain proves that there is a possibility of a hybrid architecture. In such a system, the symbolic layer is analyzed by the cognitive network and further operations are carried out on the mental representations of the input signals. This can be replicated in artificial brains without difficulties. The downside of this characteristic is that we cannot simply copy, into an artificial brain, a piece of software containing symbolic knowledge as in case of e.g. algorythimc databases. This is because we cannot foresee the method in which a given connectionist network will categorize information. With a large number of elements, this can be a chaotic process and its effects can be unique. What is more, even if the networks are identical, the categorization depends on previous experiences. This is why the process of “translating” the symbolic knowledge into the language of internal mental states must be realized via a strenuous process of individual learning. The effects of this learning could be different and depend on “capabilities” of a given network, its motivation and hierarchy of “values.” This way we know that each such network will possess unique personality. In every case we will have to deal with an individual Person.
References:
- Duch W. (2010) Architektury kognitywne, Neurocybernetyka teoretyczna, Ch. 14, pp. 329-361, editor R. Tadeusiewicz.,
- Galus W.L. (2011) Machine Mind, International Journal of Machine Consciousness (in print)
- Baddeley A.(2000) The episodic buffer in working memory, Trends in Cognitive Sciences, Vol. 4, No. 11,
- Starzyk J.A., Graham J.T. Raif P., Tan A-H., (2011). Motivated Learning for the Development of Autonomous Systems, Cognitive Science Research (In print).
- Starzyk J.A., Prasad D.K. (2011) A Computational Model of Machine Consciousness, International Journal of Machine Consciousness (In print).
- Starzyk J.A. (2011) Mental Saccades in Control of Cognitive Process (In print)
- Edelman G.M., (1992) Bright Air, Brilliant Fire, On the matter of the Mind (1999), tłum. Przenikliwe powietrze,Starzyk J.A. jasny ogień. O materii umysłu. Warszawa: PIW 1999.
- Nuxoll, A.M., & Laird, J.E. (2007). Extending cognitive architecture with episodic memory. In Proceedings of the twenty-second AAAI conference on artificial intelligence. Vancouver, BC.