COGSYS 274 Page Proofs

7/29/2019 COGSYS 274 Page Proofs

http://slidepdf.com/reader/full/cogsys-274-page-proofs 1/11

U N C O R R E

C T E D

P R

O O F

1

2 A computational account of dreaming: Learning

3 and memory consolidation

4 Action editor: Rajiv Khosla

5 Qi Zhang *

6 Sensor System, 8406 Blackwolf Drive, Madison, WI 53717, USA

7 Received 19 February 2008; accepted 20 June 20088

9 Abstract

0 A number of studies have concluded that dreaming is mostly caused by random signals because ‘‘dream contents are random1 impulses”, and argued that dream sleep is unlikely to play an important part in our intellectual capacity. On the other hand, numerous2 functional studies have suggested that dream sleep does play an important role in our learning and other intellectual functions. Specif-3 ically, recent studies have suggested the importance of dream sleep in memory consolidation, following the findings of neural replaying of 4 recent waking patterns in the hippocampus. This study presents a cognitive and computational model of dream process that involves5 episodic learning and random activation of stored experiences. This model is simulated to perform the functions of learning and memory6 consolidation, which are two most popular dream functions that have been proposed. The simulations demonstrate that random signals7 may result in learning and memory consolidation. The characteristics of the model are discussed and found in agreement with many8 characteristics concluded from various empirical studies.9 Ó 2008 Published by Elsevier B.V.

0 Keywords: Computational model; Dream; Random activation; Learning; Memory consolidation1

2 1. Introduction

3 Dreaming refers to the subjective conscious experiences4 we have during sleep. The experience is vivid, intense,5 bizarre, and is hard to recall. Various studies have con-6 cluded that dream sleep may help us in learning (e.g.,7 Greenberg & Pearlman, 1974; Hennevin, Hars, Maho, &

8 Bloch, 1995; LaBerge, 1985; Smith, 1995), and may be a9 perceptible embodiment of a dreamer’s conceptions (Hall,0 1953). Findings of the correlation between REM (rapid1 eye movement) sleep and waking learning have suggested2 that dream sleep may play an important role in learning3 and memory consolidation (e.g., Bloch, Hennevinm, &4 Leconte, 1979; Fishbein, 1970; Pearlman, 1971; Winson,

1985). Similar findings have been also concluded in morerecent psychophysiological studies, although only these ini-tiating studies are cited.

Long-term memory can be fractionated into declarative(explicit) memory and nondeclarative (implicit) memory,and declarative memory can be further divided into epi-sodic memory and semantic memory. Episodic memory is

the memory of past experiences, while semantic memoryis about factual and generic knowledge (Tulving, 1972).The hippocampal complex, including the hippocampusand its surrounding areas, is considered a critical regionin retaining recent episodic memory or its traces. On theother hand, the general neocortex is considered the placewhere semantic memory is stored. Memory consolidationis considered a neural process by which episodic memorybecomes independent of the hippocampal complex and isconsolidated into the neocortex (Squire & Alvarez, 1995).Findings from neural recording, which reveal the replaying

1389-0417/$ - see front matter Ó 2008 Published by Elsevier B.V.

doi:10.1016/j.cogsys.2008.06.002

* Tel.: +1 608 217 0742; fax: +1 608 821 0068.E-mail address: [email protected]

www.elsevier.com/locate/cogsys

Available online at www.sciencedirect.com

Cognitive Systems Research xxx (2008) xxx–xxx

COGSYS 274 No. of Pages 11, Model 5+

1 September 2008 Disk UsedARTICLE IN PRESS

Please cite this article in press as: Zhang, Q., A computational account of dreaming: Learning ..., Cognitive Systems Research (2008),doi:10.1016/j.cogsys.2008.06.002

mailto:[email protected]

mailto:[email protected]



U N C O R R E

C T E D

P R

O O F

of recent waking patterns of neuronal activity within thehippocampus during sleep, have further reinforced the viewthat dreaming may play an important part in memory con-solidation. This reactivation of hippocampal cells has beenrecorded in rats (e.g., Pavlides & Winson, 1989; Wilson &McNaughton, 1994) and in humans (Staba, Wilson, Fried,

& Engel, 2002); in SWS (slow wave sleep) that dominatesNon-REM sleep (Pavlides & Winson, 1989; Wilson &McNaughton, 1994) and in REM sleep (Louie & Wilson,2001; Poe, Nitz, McNaughton, & Barnes, 2000). Further-more, synchronized activity in the hippocampus and neo-cortex in sleep is also reported and attributed to memoryconsolidation into the neocortex (Battaglia1, Sutherland,& McNaughton, 2004).

The significance of dreaming on learning and memoryconsolidation, suggested by psychophysiological studies,is also supported by the findings of hippocampal firing insleep. In some studies (e.g., Fosse, Fosse, Hobson, & Stick-gold, 2003; Pavlides & Winson, 1989; Schwartz, 2003),

dream sleep has been directly associated with the hippo-campal firings for the possible link between the cognitiveactivity of brain and the activation of stored episodic mem-ory. This association is reasonable when we consider thefact that both vivid dreams and thought-like experiencescan be recalled in >70% of REM awakenings (Hobson,1988), and >48% of Non-REM awakenings (Nielsen,2000), respectively.

However, studies in dream reports may lead to a differ-ent conclusion. Carefully looking into dream contents, it isoften concluded that dreams are more or less randomthoughts (Foulkes, 1985; Hobson & McCarley, 1977; Wolf,

1994). A recent study by Fosse et al. (2003), which isfocused on the correlation between daily experiences anddream contents, again confirms the randomness nature of dreams. This study found that daily experiences arereplayed in the form of segments, rather than entire epi-sodes, during REM sleep. In other words, daily experienceis replayed more in random fashion and less in sequentialfashion in dreams. The randomness has led to the proposalof the activation-synthesis model (Hobson, 1988; Hobson& McCarley, 1977; Hobson, Pace-Schott, & Stickgold,2000). The model states: dreams are caused by random sig-nals arising from the pontine brainstem during REM sleep;the forebrain then synthesizes the dream and tries its bestto make sense (i.e., dream images) out of the nonsense(i.e., random impulses) it is presented with. In short, thedream randomness has been used against some proposedintellectual functions of dreaming, and has divided dreamtheories into functional and functionless. Therefore,whether or not dreaming has intellectual functions dependson whether ‘‘random impulses” can lead to intellectual con-sequences, e.g., learning and memory consolidation.

Compared to numerous psychophysiological and neuro-biological studies that are associated with dream mecha-nisms and functions, few computational studies havebeen reported on the same aspects. These reported studies

are based on connectionist modelings. In the most cited

1study by Crick and Mitchison (1983), it is concluded that1dreaming is the ‘‘reverse-learning” process to remove so-1called spurious memories (i.e., useless and old memories)1in order to avoid overload of the brain. A similar simula-1tion, however, indicates that the useless and old memory1is actually increased after the ‘‘reverse-learning” of the sim-

1ulated dreaming process, and suggests that we dream to1roughen up our ‘‘memory space” (Christos, 2003). In either1case, it is said (Botman & Crovitz, 1989; Domhoff, 1996)1that the conclusions are generally disassociated with what1has been found about dreaming, and are typically inter-1preted to contradict the psychoanalytic account of dreams.1In this study, a cognitive and computational model of 1dreaming is presented. This model is developed from a pre-1vious construct (Zhang, 2005) of a learning system. In this1present study, dreams of the computational system are per-1formed. The outcomes of dreaming are examined in terms1of ‘‘naming” and ‘‘picture drawing,” which are typical1tasks in the tests for semantic memory. The characteristics1of the model are discussed and found in agreement with1many empirical findings from dream studies.

12. The construct of an artificial intelligence dreamer

1 2.1. A brief revisit of the learning system

1Knowledge can be learned from experience. How learn-1ing occurs and how new knowledge is associated with prior1knowledge, are questions yet to be answered. A previous1study (Zhang, 2005) presents a cognitive learning system,1namely AI counter, which can learn to count. That system

1is constructed based on two rules: (1) a concept (repre-1sented by a common feature) is learned when the common14feature is abstracted and generalized; and (2) new learning14has to rely on prior learning if the newer knowledge is an14extension of the prior knowledge. The learning system14was built with a multi-level structure of information pro-14cessing, which is in fact the system shown in Fig. 1, except14for the ‘‘hippocampal memory”.14In the system, the base level of cognition is called ‘‘single14memory” that stores and reacts to one piece of an entire14external input. A single memory has three inputs (excita-14tion input Iexc, signal input Isig, and interlock input Iint)1and four outputs (excitation output Oexc, signal output1Osig , coordination output Ocor, and interlock output1Oint). The function states of a single memory on how these1signals interact are summarized in Table 1. These states1reflect the two functions of a single memory. One function1is to store an Io/Iexco pair in two steps. In step 1, when1Isig = ‘‘Io” and Iint = ‘‘ yes”, a single memory fires an Ocor

1(=‘‘yes”), which is then transported to the ‘‘bundle of inter-1subsystem signals” (see Fig. 1). Only when the ‘‘bundle”1receives a coordination signal from both the symbol and1representation subsystems, it generates a unique excitation1signal (Iexco) and sends it back to all single memories in

1both subsystems. In step 2, the single memory that fired

1Ocor stores both ‘‘Io” and ‘‘Iexco” together after receiving

2 Q. Zhang / Cognitive Systems Research xxx (2008) xxx–xxx






U N C O R R E

C T E D

P R

O O F

4 the unique ‘‘Iexco”. The other function is to fire the stored

5 information accordingly, i.e., an ‘‘Iexco” can excite the

6 stored ‘‘Io”, and an ‘‘Io” can excite the stored ‘‘Iexco”.

7 The cognitive level above the single memory is called8 ‘‘memory triangle”, which consists of three single memo-9 ries, as shown in Fig. 2. In a triangle, the Oint of a single

0 memory is the Iint of next single memory. As a result, all

three single memories form a loop with the interlock sig-nals. The purpose of a triangle is to learn a unique datapoint (‘‘Io”) three times, based on an assumption that, if a data point can represent a common feature of external

stimuli, it should appear in all external stimuli. After amemory triangle has stored the ‘‘common data point” forthree times into each of its single memories in the orderfrom 1 through 3, a common feature is considered learnedand generalized because of the existence of the loop. A sub-system with several interlocked memory triangles (e.g., theMTr1 and MTr2 in Fig. 1) is constructed to learn morecommon features. Since a memory triangle is the base fora common feature, knowledge is locally stored. The local-ized characteristic has the advantage of self-assembling alogically interrelated structure among acquired knowledge.

A concept is an abstract idea or a common feature, anda word is the symbol for a concept. A conceptual knowl-edge is considered learned only when a symbol is associatedto a common feature. This association is realized by pairingthe two subsystems into a semantic system. Such pairing isinspired by the finding of split-brain (Myers & Sperry,1953; Sperry, 1982) that indicates each brain half appearsto ‘‘have its own learning processes and its own separatechain of memories”. In his Nobel lecture, Sperry (1982)further noted that our left hemisphere is capable of com-prehending printed and spoken word, and our right hemi-sphere is word-deaf and word-blind, but capable of comprehending spatial and imagistic information. Basedon these characteristics, the symbol subsystem (dedicated

for symbolic learning) and representation subsystem (ded-

Fig. 1. The cognitive framework of a computational dreamer. This system has two cognitive sections: the learning center (i.e., the combination of thesymbol and representation subsystems) that acquires conceptual knowledge, and the hippocampal memory that stores past experiences. The S in and Rin

are input ports of a symbol and representation, respectively; while S out and Rout are output ports. The mode selector can set the system to either wakingmode (for the learning center to receive external signals) or dreaming mode (for the learning center to receive randomly activated past experiences comingfrom the hippocampal memory). The pathways and kinds of signals are indicated under the ‘‘Signal keys ”. For clarity, coordination and excitation signalsare combined. Detailed signal kinds are given in Table 1.

Table 1Three important states of a single memory

State Input Output

Learning Step 1: firing Ocor signal Isig = ‘‘Io” Osig = null

Iexc = null Oexc = nullIint = ‘‘yes” Oint = ‘‘no”

Ocor = ‘‘yes”Step 2: storing ‘‘Io” and‘‘Iexco” permanently

Isig = ‘‘Io” Osig = null

Iexc = ‘‘Iexco” Oexc = nullIint = ‘‘yes” Oint = ‘‘yes”

Ocor = ‘‘yes”Firing stored ‘‘Io” upon receiving

‘‘Iexco” after the single memoryhas learned

Isig = any Osig = ‘‘Io”

Iexc = ‘‘Iexco” Oexc = ‘‘Iexco”

or null*

Iint = ‘‘yes” Oint = ‘‘yes”Ocor = ‘‘no”

Firing stored ‘‘Iexco” upon receiving‘‘Io” after the single memory has

learned

Isig = ‘‘Io” Osi g = ‘‘Io” ornull**

Iexc = any Oexc = ‘‘Iexco”

Iint = ‘‘yes” Oint = ‘‘yes”Ocor = ‘‘no”

* Depending on Isig .** Depending on Iexc.

Q. Zhang / Cognitive Systems Research xxx (2008) xxx–xxx 3






U N C O R R E

C T E D

P R

O O F

icated for common feature learning) can be correspondentto our left and right hemispheres, respectively. And, thesubsystem pair can be correspondent to the neocortex sinceboth have the functions of the learning and storage of semantic knowledge.

The underlying mechanism of subsystem pairing is done

via the unique excitation signals generated by the bundle of inter-subsystem signals whenever learning occurs. As men-tioned earlier, the ‘‘bundle” generates a unique excitationsignal, ‘‘Iexco”, only when it receives a coordination signal(Ocor) from each subsystem, and the ‘‘Iexco” is storedtogether with ‘‘Io” in the two single memories that firedthe Ocor in both subsystems (see also Table 1). Thus, aknown common feature can excite a stored ‘‘Iexco” inthe representation subsystem, which can excite the storedsymbol in the symbol subsystem. Similarly, a known sym-bol can excite a stored ‘‘Iexco” in the symbol subsystem,which can excite the stored common feature in the repre-

sentation subsystem.According to Immanuel Kant, a concept is a common

feature or characteristic, and concepts are abstracts in thatthey omit the differences of the things in their extension,treating them as if they were identical. In order for theAI counter to be able to tally, it has to learn three conceptsthat are associated with three common features. The threefeatures are about the meaning of ‘‘zero”, ‘‘one” and‘‘tally”, and are carried by those eight pairs of examplesin either Table 2 or Table 3. For example, the 4th example,in Table 2, is an input pair of a symbol ‘‘I”, and a single-peaked representation. This example ‘‘tells” the learningsystem that the symbol ‘‘I” is the sign for a one-peaked rep-resentation. Further, the 6th example is an input pair of thesymbol ‘‘I”, and another single-peaked representation.However, this single peak looks different from that of the4th representation input. Thus, this example ‘‘tells” the sys-tem that the symbol ‘‘I” is also the sign for a different one-peaked representation. The learning system has to ‘‘omitthe differences of the things in their extension” and toabstract what is in common from given examples. In here,the differences are the widths of different single-peaks,while the common feature is ‘‘one” peak.

In Table 2, the input pairs of example 1 through 3 areabout the concept of ‘‘zero” or ‘‘nothing”, whose represen-

tation input is a flat line with no peak in it. The purpose of

24the three examples is to teach the concept that a flat line is a24background and it is like the surface of a desk with no24‘‘object” on it. The pairs of example 4 through 6 are about24the concept of ‘‘one”, and whose presentation input has a24single peak. The pairs of example 6 through 8 are about2the pattern of increment in both peaks and symbols. After2the system has learned these three concepts, it can recog-2nize external stimuli and count them, even if the stimuli2have never been experienced before, as long as they are2consistent with the concepts that have been learned. When2receiving a symbol input, the system answers it with a rep-2resentation of the symbol; when receiving a representation2input, it responds with the symbol of the representation.2The former operation is equivalent to ‘‘object drawing”2and the latter to ‘‘object naming” that are terms of typical

2tasks for semantic memory.

Fig. 2. A memory triangle that is made up of three single memories (afterZhang, 2005). All signals are correspondent to those in Table 1.

Table 2Sequenced input pairs

Table 3Non-sequenced input pairs







U N C O R R E

C T E D

P R

O O F

1 However, the learning rate of the system is dependant of 2 training examples. It only needs one round of training with3 the examples given in Table 2, which has been demon-4 strated in the previous study. However, it has to be trained5 for six rounds with the examples given in Table 3, or for6 eight rounds if examples in Table 2 were in reverse order.

7 Both Tables 2 and 3 contain the same eight examples.8 The only difference is that the examples in Table 2 are listed9 in sequence for the meanings of ‘‘zero”, ‘‘one” and ‘‘tally”,0 and are listed from simple to complex. Comparably, the1 examples in Table 3 are not listed in the sequence from sim-2 ple to complex. Based on the proposed rule of learning3 (Zhang, 2005), new learning has to rely on prior learning4 if the newer knowledge is an extension of the prior knowl-5 edge. Thus, the learning system cannot learn extended6 knowledge if its base knowledge has not already been7 learned, which is the reason why it learns slower with the8 examples in Table 3. When it cannot learn from an exam-9 ple, it simply let the external experience pass through like0 nothing has happened. Therefore, it has to experience those1 examples many more times before learning is finally real-2 ized. What if the learning system had a subsystem that3 could store external experience as its ‘‘episodic memory”4 for later reprocessing? Comparably, humans, and most5 mammals, do have the capability to remember rapidly6 what has happened and to retain this as episodic memory7 after a single experience.

8 2.2. An AI dreamer

9 In humans and almost all mammals, there is such a

0 memory system, namely the hippocampal complex, which1 has been revealed to have the function of immediate stor-2 age (e.g., McClelland, McNaughton, & O’Reilly, 1995;3 Squire, 1992; Tulving, 1972). When the complex is dam-4 aged, the capacity of retaining episodic memory is impaired5 (Scoville & Milner, 1957) and patients become amnesic6 (e.g., Squire & Alvarez, 1995). Deterioration of this com-7 plex has been commonly found in patients with Alzhei-8 mer’s disease for whom poor retention of episodic9 memory is the first sign.0 Based on the function of the biological hippocampal com-1 plex, a ‘‘hippocampal memory” subsystem is added to the2 learning system of AI counter. As shown in Fig. 1, the sub-3 system has a number of ‘‘memory cells”, and eachcell is des-4 ignated to store a segment of an external experience. All of 5 the cells are interlocked so that the sequenced segments of 6 an external experience are stored in their original order of 7 occurrence. The interlocking mechanism has similar func-8 tions as the hippocampus in terms of sequential learning9 (e.g., Levy, 1996) when the variable is time, and spatial nav-0 igation (e.g., Muller, Kubie, & Ranck, 1987) when the vari-1 able is distance. Thereby, all stored segments of an external2 experience can be re-accessed and replayed in their originally3 occurring sequences (to be demonstrated). The segments can4 also be fired randomly, which becomes the source of random

5 signals in dream simulations of this study.

The system, shown in Fig. 1, is named AI dreamer,which is made up of two cognitive sections. One sectionis the AI counter of the previous study, which is also calledthe ‘‘learning center” in later discussions for convenience.And, the other is the ‘‘hippocampal memory” that storesits experience and allows the experience to be re-accessed.

Here, the ‘‘conceptual knowledge”

and ‘‘stored experience”

are compared with the concepts of semantic and episodicmemories defined by Tulving (1972), respectively. Thistwo-sectioned construct also agrees with an understandingthat the brain has a complementary memory system of knowledge and experience, each of which uses differentcomputational strategies for storing information (e.g.,McClelland, McNaughton, & O’Reilly, 1995; O’Keefe &Nadel, 1978).

Besides the added hippocampal memory, a ‘‘mode selec-tor” is also added that can set the system either to wakingmode in order to receive external stimuli, or set the systemto dreaming mode to process internal signals that are fired

randomly from the hippocampal memory. Similar as forthe learning system of the AI counter, the cognitive system(namely AI dreamer) shown in Fig. 1 is also built in theLabView language (e.g., Bishop, 2001), although the samecognitive construct can be built in any commonly usedcomputer language.

3. Simulations of the AI dreamer

The external experience for the system to learn is a seriesof input pairs given in Table 3. Each pair of the input consistsof a symbol and its representation. When the system learns,

it takes in an input pair via separated channels, S in and Rin,shown in Fig. 1. The representation input carries the mean-ing of the symbol input, while the symbol is a ‘‘name” thatsymbolizes the representation. The term ‘‘meaning” in heremeans that the input carries the conceptual knowledge thatcan be perceived and understood in a real world. We cansee from Table 3, that a representation input may be a flatline or may come with a number of peaks. Every peak repre-sents one ‘‘object”, while a flat linemeans‘‘no” object. Whenwe see an input has three peaks, the meaning of three peaks isto be perceived by the learning center. Those input pairs ineither Tables 2 or 3 carry the same knowledge (i.e., ‘‘zero”,‘‘one” and ‘‘tally”) necessary for the system to learn howto tally. The only difference between the two tables is inthe given sequence of the input pairs.

Firstly, in training phase, the system is set to wakingmode to receive all given pairs in Table 3 in their givensequence. This system only receives those pairs once, butit must ‘‘remember” the entire experience thereafter. Thisis comparable with the fact that human can remember whathas happened after a single experience, and this property of rapid remembering is demonstrated through the ‘‘serialrecall” (that is also a standard test in memory study) shownin Table 4.

When an input pair of symbol and representation

arrives during the training, the symbol is forwarded to







U N C O R R E

C T E D

P R

O O F

the symbol subsystem, while the representation is for-warded to the representation subsystem. In turn, the inputsare delivered to the memory triangles and single memory atlower cognitive levels. The learning center may learn from,react to, or reject the input pair. What the learning centereventually does is a result of interactions among all itsenclosed single memories, each of which it also has the

potential of learning from, reacting to or rejecting. If thecenter cannot learn from an input pair, it rejects and for-wards the pair to the hippocampal memory via the twopathways as indicated in Fig. 1. The pair is then storedin one of the memory cells as a part of ‘‘episodic memory”.

After the training phase, the learning center is only ableto learn from the 6th pair, and it has to store the rest of thepairs into the hippocampal memory. This is because the 6thpair is the starting point of the most basic knowledge of tally – a flat baseline that means there is ‘‘no” object.Learning from the 6th pair is only the beginning to learnthe baseline. Thereafter, the learning center has to learnfrom the 5th pair and then the 4th pair before it can fullygeneralize the meaning of a baseline (i.e., the meaning of ‘‘zero”). Only when it knows what is a baseline, it is ableto subtract one ‘‘object” (i.e., one peak) from the baselinesof the 3rd, 8th, and 2nd pairs, and to learn the property of the subtracted ‘‘one object”. After it has learned the mean-ing of ‘‘one”, it then can subtract more peaks and abstractthe meaning of tally in subsequent learnings of the 7th and1st pairs. This progressive learning is the result of themechanism proposed in the previous study that new learn-ing has to rely on prior learning when the newer knowledgeis an extension of the prior knowledge. All these subse-quent learning steps, however, does not occur in the train-

ing phase, and only occur in the dream phase in which the

4stored examples are randomly fired and processed by the4learning center.4After the initial external experience, the system is set to4dream mode to allow the learning center to reprocess those4stored experiences fired from the hippocampal memory.4During dreaming, the learning center acts in the same4way as it is ‘‘awake”. It learns from an input pair if the

4input’s prior knowledge has already been learned; it4responds to an input if the meaning of the input has4already been learned; and it rejects an input pair if the4pair’s prior knowledge has not been learned.4During dreaming, the stored segments of experience4are randomly fired, which is intended to simulate the ran-4dom nature of dreaming that has been concluded by a4number of dream studies. The randomness has strongly4challenged the view that dreaming has any intellectual4function. For example, dreaming is thought only a neural4process by which our brain produces dream imagery from4noisy signals (Hobson, 1988; Hobson & McCarley, 1977).4The central point of the challenge is based on the4assumption that randomly assembled information hardly4contributes to cognition and knowledge. But, the assump-4tion may not be valid. From Russell’s viewpoint (1913),4knowledge can be learned through experience of a direct4causal (experience-based) interaction between a person4and the object that the person is perceiving. On the other4hand, it has also been argued that true knowledge (or at4least the most important knowledge) is essentially inde-4pendent of sensory experience (Locke, 1690). If both4statements are true, it may be said that knowledge can4be acquired through experience, but the acquisition pro-4cess is not necessarily dependent on an exact experience.

4Random firing may be sufficient enough to provide an

Table 4Recorded streams of firings from the hippocampal memory







U N C O R R E

C T E D

P R

O O F

7 alternative to the exact replay of an actual experience in8 knowledge acquisition.9 Table 4 lists some recorded streams of representation0 flows that are intercepted along the pathway from the1 mode selector to the representation subsystem. The first2 simulation is a ‘‘serial recall” recorded in waking mode.

3 If we simply attach all representation signals, given in4 Table 3 except the 6th input, into a ‘‘long representation5 signal”, we will see that the ‘‘long representation signal”6 is same as the first recorded stream in Table 4. For exam-7 ple, one may notice that the combination of the first three8 peaks in this serial recall is the first representation input in9 Table 3, and the last peak is the eighth representation0 input. Other streams are recorded in six different dream1 simulations of randomly activated ‘‘experiences”. Again,2 each stream is made up of activated experiences of the eight3 examples in Table 3. The only difference is that the experi-4 ences are randomly activated. The streams are compressed5 in order to fill in the limited space in Table 4, thus the coor-6 dination between the streams and the eight examples are7 not clearly shown in Table 4. But some of them may still8 be seen, e.g., the first two peaks in the last stream are the9 two peaks of the seventh example in Table 3. The results0 of learning, after the learning center has experienced the1 recorded dream streams, have also been indicated in the2 same table. As noted before, the input pairs in Table 33 carry all conceptual knowledge (including ‘‘zero”, ‘‘one”4 and ‘‘tally”) that are sufficient for the system to perform5 tallying after full learning. These concepts may be partially6 or fully learned in dreams under a given number of total7 firings. The efficiency of learning in a specific dream varies

8 depending on how soon newer knowledge arrives after its9 required prior knowledge has been learned. The learning0 center may learn all of the conceptual knowledge in one1 dream (e.g., the first and second dreams in Table 4) or in2 many dreams (e.g., the series of four dreams in Table 4).3 Furthermore, with additional cognitive structure, this4 system can be extended to learn other counting systems,5 e.g., Roman numerals or Arabic numerals. The latter has6 been demonstrated in the previous study (Zhang, 2005).7 However, since the present study is centered on semantic8 learning from stored experiences, the extension is not9 discussed.

0 4. Discussion

1 4.1. Phenomenal and neurobiological properties of dreaming

2 As we go to sleep, we slowly sink down into deeper3 stages of sleep (i.e., Non-REM or NREM sleep). After4 an hour or two, the first REM period begins and lasts a5 few minutes. Then, we sink back into NREM sleep again.6 This cycle occurs about every 90 min. Towards the end of 7 the night, the REM periods get longer. NREM sleep alter-8 nates with REM sleep and includes all sleep apart from9 REM. Each NREM sleep has four stages corresponding

0 to increasing depth of sleep as indicated by progressive

dominance of the electroencephalography (EEG) by high-voltage, low frequency wave activity (Rechtschaffen &Kales, 1968). This low frequency wave is also called ‘‘slowwave”, which dominates the deepest stages of NREM.Almost all mammals have the NREM–REM cyclic alterna-tion in sleep, which suggests not only a shared mechanism

across species but also a universal functional significance(Hobson et al., 2000).Dreams can occur in both REM and NREM sleep. In

70–95% of awakenings from the REM state, normal subjects report that they have been dreaming whereas only5–10% of NREM awakenings produce equivalent reports(Hobson, 1988). If there is no clear dream reported, 43– 50% of NREM awakenings can elicit reports of thought-like or mentation recall (Foulkes, 1962; Nielsen, 2000). If the mentation recall can also be accounted as the resultof dreaming, it means the mind is mostly in dream statesduring sleep. There are more subtle differences betweenREM and NREM dreams. Reports from REM sleep awak-

enings are relatively longer, more perceptually vivid, moremotorically animated, more emotionally charged, and lessrelated to waking life than NREM reports (e.g., Antrobus,1987; Foulkes, 1962). In contrast to REM reports, NREMreports are more thought-like and contain representationsof current concerns more often than do REM sleep reports(Foulkes, 1962).

It has been noted earlier in the simulation of dreamsthat, in dream mode, the learning center is as active as itis in awaking mode. A similar characteristic has been con-cluded from neurobiological studies. It was stated (Hobsonet al., 2000) that, in REM sleep, the brain is almost as

active as when awake, except for the sensory input andmotor output, which are blocked, and for certain areas of the dorsolateral prefrontal cortex and the primary visualcenters which are selectively deactivated. A number of functional imaging studies in humans have revealed a fairlyconsistent pattern of activities in REM sleep that appear toreflect dream processes (Braun et al., 1997; Maquet et al.,1996; Nofzinger, Mintun, Wiseman, Kupfer, & Moore,1997). Some important findings from the studies are: inREM sleep, the brainstem reticular formation is highlyactive; primary sensory areas (e.g., striate cortex for thevisual system) are inactive; by contrast, extrastriate (visual)regions (as well as other sensory association sites) are veryactive; limbic and paralimbic regions (including the hippo-campus, amygdala and anterior cingulate) are intenselyactivated; and widespread regions of the frontal cortexincluding the lateral orbital and dorsolateral prefrontalcortices show marked reductions in activity. These studiesalso show that the following areas are relatively less activein NREM than in REM sleep; they are the brainstem, mid-brain, anterior hypothalamus, hippocampus, caudate, andmedial prefrontal, caudal orbital, anterior cingulate, para-hippocampal and inferior temporal cortices (Braun et al.,1997).

Besides the fact that the general neocortex in dream

sleep is almost as active as when awake, the highly acti-







U N C O R R E

C T E D

P R

O O F

vated limbic system, especially the hippocampus, is of spe-cial interest. The activated hippocampus coincides with thereactivation of recent waking patterns of neuronal activitywithin the hippocampus. The reactivation has beenrevealed in human and other mammals, and in REM andNREM sleeps, while the patterns of neuronal activity are

generally thought to be associated with daily experienceor episodic memory (Louie & Wilson, 2001; Pavlides &Winson, 1989; Poe et al., 2000; Staba et al., 2002; Wilson& McNaughton, 1994).

What could be the common functioning areas in bothREM and NREM sleep? Functional imaging studies areable to indicate the activity intensities of various brainareas, but are not adequate to rule out the involvementsof the less active areas. For example, bursts of firings havebeen recorded from neurons in the hippocampus duringNREM sleep (Wilson & McNaughton, 1994), althoughbrain imaging has indicated that this area is less active inNREM sleep. On the other hand, the findings that patients

with large pontine lesions can still dream suggest that thebrainstem is not the critical component for all dreams tooccur, but rather it may only be important for REMdreaming (Solms, 1997). This study of pontine lesions isconsistent with the finding from functional brain imagingthat the pontine brainstem and anterior cingulate cortexis deactivated during NREM sleep (Braun et al., 1997).

If only considering those activated areas in both REMand NREM sleep, it may be concluded that the commonlyactivated areas necessary for dreams to occur, at least,include: most of the general neocortex and of the limbicsystem (especially the hippocampus). This simplified acti-

vation pattern is compatible with the cognitive structureof the computational dreamer, which has a learning center(mimicking the conceptual learning function of the generalneocortex) and a hippocampal memory (mimicking thestorage and reactivation functions of episodic memory inthe hippocampus). However, it must be pointed out thatthis system is only a simplified approach to explain somecommon properties of dreams in both REM and NREMsleep. By no means, is it thought to be able to enclose alldream properties of a brain.

4.2. Psychophysiological characteristics of dreaming and

dream mechanism

The characteristics of dreaming that are concluded frommany empirical studies can be employed to examine theproposed theory of dreaming and its simulations. The firstis the characteristic of the randomness. This characteristicis concluded from or supported by empirical findings con-ducted by, e.g., Hobson and McCarley (1977), McCarleyand Hoffman (1981), Hobson (1988), Reinsel, Antrobus,and Wollman (1992), Williams, Merritt, Rittenhouse, andHobson (1992), Revonsuo and Salmivalli (1995). Thesefindings have been summarized by Hobson et al. (2000)as followings: (1) dream imagery can change rapidly, and

is often bizarre in nature; (2) relative to waking and, when

6present, dreams often involve weak, post-hoc, and logically6flawed explanations of improbable or impossible plots; and6(3) dreams lack orientational stability; persons, times, and6places are fused, plastic, incongruous and discontinuous. In6Hobson’s view (e.g., 2000), the bizarreness or impossible6plots in dreams are simply resulted from the randomness.

6The second one is the repetitive characteristic of dreams6 – repeated themes and repeated dreams (including recur-6rent dreams) are common among dreams (Domhoff,61993). The proposed mechanism of conceptual learning6states that new knowledge cannot be learned if its prior6knowledge is not already learned. As a result, a segment6of stored experience may by chance have to be fired multi-6ple times before the learning center is able to learn from it.6This multi-appearance of the same segment of stored expe-6rience can be commonly observed in those dream-source6streams shown in Table 4. When the learning center is6not able to fully abstract and generalize the knowledge6from the random signals in one dream, more dreams6become necessary before dream learning is thoroughly real-6ized. That is the scenario of the multi-dream learning,6shown also in Table 4, in which repeated ‘‘themes” and6similar ‘‘dreams” are expected.6The lack of self-reflection and self-control are also the6typical characteristics of dreaming. For example, self-6reflection in dreams is generally found to be absent ( Rec-6htschaffen, 1978) or greatly reduced (Bradley, Hollifield,6& Foulkes, 1992); the dreamer rarely considers the possibil-6ity of actually controlling the flow of dream events (Purcell,6Mullington, Moffitt, Hoffman, & Pigea, 1986); volitional6control is greatly attenuated in dreams (Hartmann, 1966).

6Similarly, this computational model of AI dreamer does6not have a necessity of a central controller in regulating6its learning process.

64.3. The dream function of memory consolidation and

6learning

6The dream function of memory consolidation has long64been suggested in various studies. Newer and relatively64direct evidence comes from the neuronal activity of seem-64ingly replay of waking-experience in the hippocampus in64both REM and NREM sleep. Many researchers agree that64such reactivation is experience dependent, and conclude64that the reactivation is associated with the consolidation64of episodic memory stored in the hippocampus (Pavlides64& Winson, 1989; Poe et al., 2000; Skaggs & McNaughton,641996; Staba et al., 2002; Wilson & McNaughton, 1994).64Others who studied amnesia and hippocampal lesion have6also expected the consolidation function in dream sleep.6For example, Squire and Alvarez (1995) have argued that6if episodic memory is revived constantly, dreaming may6be the best answer to explain why the memory consolida-6tion process does not regularly intrude into our6consciousness.6Neuronal recordings have suggested that daily experi-

6ences stored in the hippocampal complex can be reacti-







U N C O R R E

C T E D

P R

O O F

8 vated or replayed during REM and NREM sleeps, but9 there are no indications about whether the replaying is0 sequential or random. Study of dream contents and their1 connections with daily experience may provide this detail.2 A recent study (Fosse et al., 2003) concluded that only3 1.4% of dream reports contained the reproductions of

4 experience ensembles; while 65% of dream elements are5 linked to fragments and features of waking events. This6 apparent contradiction between dreaming’s lack of a fully7 episodic structure and its proposed role of the consolida-8 tion, has led to the suggestion (Fosse et al., 2003; Stickgold,9 Hobson, Fosse, & Fosse, 2001) that sleep only has its role0 in the consolidations of associated semantic memory and1 procedural memory, not episodic memory. In interpreting2 the data, however, some researchers have realized and sug-3 gested that the key is to understand why memories appear4 in dreams as fragments or partial episodes but also occa-5 sionally as complete replays (Nielsen & Stenstrom, 2005;6 Schwartz, 2003). The understanding offered by this study7 is that, when random firing is the only fashion of the reac-8 tivation, daily experiences should only appear in the form9 of segments rather than the sequential replay of entire past0 experience.1 Memory consolidation is defined as the process by2 which memory becomes independent of the hippocampal3 region (Squire & Alvarez, 1995). The presented computa-4 tional system can demonstrate that past experiences stored5 in the hippocampal memory have been consolidated into6 the learning center in tests called ‘‘naming”. Naming test7 is a standard test in memory studies in which the tested8 subject calls out the name of a presented item. During

9 the naming tests, the hippocampal memory is removed0 after dream consolidation, in order to demonstrate that1 the memory has become independent of the ‘‘hippocampal2 memory”. The first three rows in Table 5 are summaries of 3 three naming tests. In these tests, different ‘‘objects” (or4 pictures) are shown and the system answers with the cor-5 rect names to each of the given objects. When compared

with Table 3, one can see that the objects tested and thenames answered are part of the inputs given in the tablethat have been experienced by the system previously. Sincethe tests are done after the ‘‘hippocampal memory” wasremoved, it indicates that, after dreaming, the ‘‘episodicmemory” has been consolidated into the learning center

and become free of the hippocampal memory.It may be noticed, from those examples in Table 5, thatthe memory consolidation in the computational model isnot a simple relocation of stored experiences, rather alearning process that gradually incorporates facts andevents into an already existing framework of knowledge(e.g., McClelland, McNaughton, & O’Reilly, 1995). Thelearning by consolidation is the natural outcome of thecomputational system that coincides with the understand-ing that mechanisms of the hippocampus and neocortexinvolved in the consolidation process, are different (e.g.,O’Keefe & Nadel, 1978) however complementary (e.g.,McClelland, McNaughton, & O’Reilly, 1995). In other

words, learning and memory consolidation can be seen asdifferent sides of the same ‘‘coin”.

The function of dream learning is the most populardream function that has been observed and concluded.The impact of dream learning is broad and can be oftenobserved after dreams. For example, unprepared learning,more slowly mastered and difficult, is especially dependenton the quality of REM sleep (Greenberg & Pearlman,1974); learning tasks requiring significant concentrationor acquisition of unfamiliar skills are followed by increasedREM sleep (LaBerge, 1985); material learned during theday and consolidated over a night of sleep is recalled better

the next morning (Hennevin et al., 1995). The dream learn-ing function can also be demonstrated with the presentedsystem through both naming tests and ‘‘picture drawing”tests (also one of the standard memory tests) after theremoval of the hippocampal memory. The last five rowsin Table 5 are summaries of two naming tests and threedrawing tests. One can see that the inputs in four out of the five tests are ‘‘new” to the system because they werenot included in Table 3. In other words, the system hasnever had the chance of ‘‘knowing” them. The systemwould have made mistakes or made no response to theinputs if memory consolidation was a simple relocationof experience. Instead, it names those presented ‘‘pictures”without mistake, and counts the peaks of the representa-tion inputs, and draws ‘‘pictures” correctly by tallying thesymbol and drawing the peaks out accordingly. Althoughthose four inputs seem ‘‘new” to the system, they are notnew at all in meaning, or in the sense of common featuresand conceptual knowledge. The same common features(and conceptual knowledge) have already been carried bythose input pairs in Table 3 and have already beenabstracted and generalized into the learning center duringthe system’s dreams. This demonstrated property of acquired knowledge is called flexibility, which is consideredone of the key characteristics of semantic memory that is

associated with the general neocortex.

Table 5Naming and picture drawing tests







U N C O R R E

C T E D

P R

O O F

5. Conclusion

Few computational models (Christos, 2003; Crick &Mitchison, 1983) have been reported in exploring theunderlying mechanisms and functions of dreaming. And,the conclusions from these models are generally not associ-

ated with what have been revealed in empirical studies of dreams (Domhoff, 1996). Memory consolidation has beensimulated in a number of studies based on connectionistmodeling (Alvarez & Squire, 1994; McClelland et al.,1995; Murre, 1996) to understand the underlying mecha-nisms of memory consolidation and retrograde amnesia.In these studies, the consolidation processes were notdirectly associated with dreaming process and its character-istics, although the authors suggested a connection betweendreaming and memory consolidation. Due to the nature of parallel distributed processing of connectionist modeling,one neuron in one subsystem may have to connect withmany neurons in a different subsystem, thus a single stream

of information flow that is compatible with dreams wouldnot exist during the simulations of consolidation process.Furthermore, these studies were unable to demonstratethe flexibility of the acquired knowledge from memory con-solidation, whereas flexibility is one of the key characteris-tics of semantic memory that is generally considered as aresult of memory consolidation.

The presented cognitive model of dreaming confirmscomputationally that learning and memory consolidationof episodic memory can be realized with randomly arrivingsignals that are stored segments of experience or episodicmemory. This model offers a positive answer to Freud’s

inquiry (1900): ‘‘Is the dream capable of teaching us some-thing new concerning our internal psychic processes andcan its content correct opinions which we have held duringthe day?” It also demonstrates that dreaming is a cognitiveprocess that deals with the ‘‘daily residual” (Freud, 1900),‘‘unfinished business” (Hall & Van De Castle, 1966) andcorrespondences with waking thoughts (Domhoff, 1996;Foulkes, 1985). Although this system is only able to dem-onstrate the dream functions of learning and memory con-solidation, there is no doubt that dreaming functionsshould cover many other cognitive aspects and psycholog-ical properties as many studies have concluded.

References

Antrobus, J. (1987). Cortical hemisphere asymmetry and sleep mentation.

Physiological Review, 94, 359–368.Alvarez, P., & Squire, L. (1994). Memory consolidation and the medial

temporal lobe: A simple network model. Proceedings of the National

Academy of Sciences of USA, 91, 7041–7045.Battaglia1, F. P., Sutherland, G. R., & McNaughton, B. L. (2004).

Hippocampal sharp wave bursts coincide with neocortical ‘‘up-state”transitions. Learning and Memory, 11, 697–704.

Bishop, R. H. (2001). Learning with LabVIEW 6i . New York: AddisonWesley.

Bloch, V., Hennevinm, E., & Leconte, P. (1979). Relationship between

paradoxical sleep and memory processes. In M. A. Brazier (Ed.), Brain

8mechanisms in memory and learning: From the single neuron to man.8New York: Raven Press (pp. 329–343).8Botman, H. I., & Crovitz, H. F. (1989). Dream reports and autobio-8graphical memory. Imagination. Cognition and Personality, 9,8213–214.8Bradley, L., Hollifield, M., & Foulkes, D. (1992). Reflection during REM8dreaming. Dreaming, 2, 161–166.8Braun, A. R., Balkin, T. J., Wesensten, N. J., Carson, R. E., Varga, M.,8Baldwin, P., et al. (1997). Regional cerebral blood flow throughout the8sleep–wake cycle. Brain, 120, 1173–1197.8Christos, G. (2003). Memory and dreams: The creative human mind . New8Jersey: Rutgers University Press.8Crick, F., & Mitchison, G. (1983). The function of dream sleep. Nature,8304, 111–115.8Domhoff, G. W. (1996). Finding meaning in dreams: A quantitative8approach. New York: Plenum.8Domhoff, G. W. (1993). The repetition of dreams and dream elements. In8A. Moffitt, M. Kramer, & R. Hoffmann (Eds.), The functions of 8dreaming up (pp. 293–320). Albany: SUNY Press.8Fishbein, W. (1970). Interference with conversion of memory from short-8term to long-term storage by partial sleep deprivation. Communica-8tions in Behavioral Biology, 5, 171–175.8Fosse, M. J., Fosse, R., Hobson, J. A., & Stickgold, R. J. (2003).8Dreaming and episodic memory: A functional dissociation? Journal of 8Cognition Neuroscience, 15, 1–9.8Foulkes, D. (1985). Dreaming: A cognitive-psychological analysis. Hills-8dale: Erlbaum.8Foulkes, D. (1962). Dream reports from different states of sleep. Journal of 8Abnormal and Social Psychology, 65, 14–25.8Freud, S. (1900). In J. Strachey (Ed.), The interpretation of dreams. New8York: Basic Books.8Greenberg, R., & Pearlman, C. (1974). Cutting the REM nerve: An84approach to the adaptive role of REM sleep. Perspectives in Biology84and Medicine, 17 , 513–521.84Hall, C. S. (1953). A cognitive theory of dream symbols. Journal of 84General Psychology, 48, 169–186.84Hall, C. S., & Van de Castle, R. L. (1966). The content analysis of dreams.84New York: Appleton-Century-Crofts.84Hartmann, E. (1966). The psychophysiology of free will. In E. Lowenstein84(Ed.), Psychoanalysis: A general psychology (pp. 521–536). New York:84International University Press.84Hennevin, E., Hars, B., Maho, C., & Bloch, V. (1995). Processing of 8learned information in paradoxical sleep: Relevance for memory.8Behavioral Brain Research, 69, 125–135.8Hobson, J. A. (1988). The dreaming brain. New York: Basic Books.8Hobson, J. A., & McCarley, R. W. (1977). The brain as a dream-state8generator: An activation-synthesis hypothesis of the dream process.8American Journal of Psychiatry, 134, 1335–1348.8Hobson, J. A., Pace-Schott, E., & Stickgold, R. (2000). Dreaming and the8brain: Towards a cognitive neuroscience of conscious states. Behavioral 8and Brain Sciences, 23, 793–842.8LaBerge, S. (1985). Lucid dreaming . New York: Ballantine.

8Levy, W. B. (1996). A sequence predicting CA3 is a flexible associator that 8learns and uses context to solve hippocampal-like tasks. Hippocampus,86 , 579–590.8Locke, J. (1690). An essay on human understanding . London: Penguin8Books, Republished in 1997.8Louie, K., & Wilson, M. A. (2001). Temporally structured replay of awake8hippocampal ensemble activity during rapid eye movement sleep.8Neuron, 29, 145–156.8Maquet, P., Peters, J. M., Aerts, J., Delfiore, G., Degueldre, C., Luxen, A.,8et al. (1996). Functional neuroanatomy of human rapid-eye-movement8sleep and dreaming. Nature, 383, 163–166.8McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. (1995). Why8there are complementary learning systems in the hippocampus and8neocortex: Insights from the successes and failures of connectionist8models of learning and memory. Psychological Review, 102,

8419–457.







U N C O R R E

C T E D

P R

O O F

6 McCarley, R. W., & Hoffman, E. (1981). REM sleep dreams and the7 activation-synthesis hypothesis. American Journal of Psychiatry, 138,8 904–912.9 Muller, R. U., Kubie, J. L., & Ranck, J. B. (1987). Spatial firing patterns0 of hippocampal complex spike cells in a fixed environment. Journal of 1 Neuroscience, 7 , 1935–1950.2 Murre, J. M. (1996). TraceLink: A model of amnesia and consolidation of 3 memory. Hippocampus, 6 , 675–684.4 Myers, R. E., & Sperry, R. W. (1953). Interocular transfer of a visual5 forma discrimination habit in cats after section of the optic chiasm and6 corpus callosum. Anatomical Record, 115, 351–352.7 Nielsen, T. A. (2000). A review of mentation in REM and NREM sleep:8 ‘‘Covert” REM sleep as a possible reconciliation of two models.9 Behavioral and Brain Sciences, 23, 851–866.0 Nielsen, T. A., & Stenstrom, P. (2005). What are the memory sources of 1 dreaming? Nature, 437 , 1286–1289.2 Nofzinger, E. A., Mintun, M. A., Wiseman, M. B., Kupfer, D. J., &3 Moore, R. Y. (1997). Forebrain activation in REM sleep: An FDG4 PET study. Brain Research, 770, 192–201.5 O’Keefe, J., & Nadel, L. (1978). The hippocampus as a cognitive map.6 Oxford: The Clarendon Press.7 Pavlides, C., & Winson, J. (1989). Influences of hippocampal place cell8 firing in the awake state on the activity of these cells during subsequent9 sleep episodes. Journal of Neuroscience, 9, 2907–2918.0 Pearlman, C. (1971). Latent learning impaired by REM sleep deprivation.1 Psychonomic Science, 25, 135–136.2 Poe, G. R., Nitz, D. A., McNaughton, B. L., & Barnes, C. A. (2000).3 Experience dependent phase-reversal of hippocampal neuron firing4 during REM sleep. Brain Research, 855, 176–180.5 Purcell, S., Mullington, J., Moffitt, A., Hoffman, R., & Pigea, R. (1986).6 Dream self-reflectiveness as a learned cognitive skill. Sleep, 9, 423–437.7 Rechtschaffen, A. (1978). The single-mindedness and isolation of dreams.8 Sleep, 1, 97–109.9 Rechtschaffen, A., & Kales, A. (1968). A manual of standardized 0 terminology, techniques and scoring system for sleep stages of human1 subjects. Washington: Public Health Service.2 Reinsel, R., Antrobus, J., & Wollman, M. (1992). Bizarreness in dreams3 and waking fantasy. In J. S. Antrobus & M. Bertini (Eds.), The4 neuropsychology of sleep and dreaming (pp. 157–184). Hillsdale:5 Lawrence Erlbaum Associates.6 Revonsuo, A., & Salmivalli, C. (1995). A content analysis of bizarre7 elements in dreams. Dreaming, 5, 169–187.

Russell, B. (1913). Theory of knowledge: The 1913 manuscript. London:Allen and Unwin, Republished in 1984.

Schwartz, S. (2003). Are life episodes replayed during dreaming? Trends in

Cognitive Sciences, 7 , 325–327.Scoville, W. B., & Milner, B. (1957). Loss of recent memory after bilateral

hippocampal lesions. Journal of Neurology. Neurosurgery and Psychi-

atry, 20, 11–21.Skaggs, W. E., & McNaughton, B. L. (1996). Replay of neuronal firing

sequences in rat hippocampus during sleep following spatial experi-ence. Science, 271, 1870–1873.

Smith, C. (1995). Sleep states and memory processes. Behavioral Brain

Research, 69, 137–145.Solms, M. (1997). The neuropsychology of dreams. Mahwah: Lawrence

Erlbaum.Sperry, R. (1982). Some effects of disconnecting the cerebral hemispheres.

Science, 217 , 1223–1226.Squire, L. R. (1992). Memory and the hippocampus: A synthesis from

findings with rats, monkeys, and humans. Psychological Review, 99,195–231.

Squire, L. R., & Alvarez, P. (1995). Retrograde amnesia and memoryconsolidation: A neurobiological perspective. Current Opinion in

Neurobiology, 5, 169–177.Staba, R. J., Wilson, C. L., Fried, I., & Engel, J. J. (2002). Single neuron

burst firing in the human hippocampus during sleep. Hippocampus, 12,724–734.

Stickgold, R., Hobson, J. A., Fosse, R., & Fosse, M. (2001). Sleep,learning, and dreams: Off-line memory reprocessing. Science, 294,1052–1057.

Tulving, E. (1972). Episodic and semantic memory. In E. Tulving & W.Donaldson (Eds.), Organization of memory (pp. 381–403). New York:Academic Press.

Williams, J., Merritt, J., Rittenhouse, C., & Hobson, J. A. (1992).Bizarreness in dreams and fantasies: Implications for the activation-synthesis hypothesis. Consciousness and Cognition, 1, 172–185.

Wilson, M. A., & McNaughton, B. L. (1994). Reactivation of hippocam-pal ensemble memories during sleep. Science, 265, 676–679.

Winson, J. (1985). Brain and psyche: The biology of the unconscious. NewYork: Doubleday/Anchor Press.

Wolf, F. A. (1994). The dreaming Universe. New York: Simon & Schuster.Zhang, Q. (2005). An artificial intelligent counter. Cognitive Systems

Research, 6 , 320–332.




Please cite this article in press as: Zhang, Q., A computational account of dreaming: Learning ..., Cognitive Systems Research (2008),

Documents

COGSYS 274 Page Proofs