| |
||
|
|
|
|
Frequently Asked Questions 1. Why do we need new hypotheses for memory? 2. Summarize the semblance hypothesis and its derivation? 3. What is the nature of the memory code mentioned in the hypothesis? 4. What are the basic differences between this and the existing hypotheses? 5. How can we retrieve memories after many years (even without retrieving it in between)? 6. How can we explain long term potentiation (LTP) in terms of the semblance hypothesis? 7. Can we explain brain function other than memory using the basic units proposed by the semblance hypothesis? 8. Can we explain behavior in terms of the semblance hypothesis? 9. Is there any experimental evidence supporting the present hypothesis? 10. How can we replicate the proposed hypothetical model in a physical system? 11. Is there any direct evidence for the hypothesis? 12. Is memory retrieval an active process? ----------------------------------------------------------------------------------------- 1. Why do we need new hypotheses for memory? Whatever features that we discover in biological system (either structures or functions) have to be used in computational studies to verify their feasibility as the mechanism of memory. We have already found a strong correlation between long term potentiation (LTP) and memory. However, LTP lacks sufficiency to be the mechanism of memory storage (Martin et al., 2000; Piorazi and Mel, 2001). (It becomes necessary that when we discover the true mechanism, it should be able to explain the correlation between LTP and memory). Computational neuroscientists are still looking for mechanisms in biological systems that can substantiate memory storage. Given the above facts, we can say that we have not yet discovered the mechanism of memory storage. A gold standard for its discovery will be to the ability to replicate the mechanism in physical systems (Let us keep this strict criterion before making any search). One method of approach for its discovery is to make hypotheses and test them theoretically, followed by experimental verifications. Hypothesis development is very important since the computations of the features (we don’t know yet what features are to be computed) from nearly 1010 neurons and 1015 synapses are required to understand the mechanism. A single counter example/ proof against a hypothesis can then be used as sufficient reason to modify/ reject it. According to Karl Popper, a philosopher of science, a hypothesis must be falsifiable; i. e. it must at least in principle be possible to make an observation that would disprove the proposition as false, even if one has not actually (yet) made that observation (Popper, 1965). Once such an observation is made, it will lead to the rejection of the hypothesis. However, even with the rejection of a hypothesis, we are likely to make some conclusions that will aid in the development of newer and better hypotheses. A general requirement of a hypothesis of memory is that it should theoretically be able to explain the following features. 1. Retrieval of memory at physiological time-scales 2. Provision for unlimited memory life-times (Rubin and Fusi, 2007) 3. Absence of overwriting of old memories by the new ones (Rubin and Fusi, 2007) 4. Absence of decay of the memory trace by any modifications of the basic units by new learning (Rubin and Fusi, 2007) 5. Instant access to very large memory stores (Abbott, 2008) 6. Ability of the mechanism/ system to generate hypothesis (Abbott, 2008) 7. Provisions for interaction between internally generated hypotheses and external evidence that allows sensory data to veto or support internal constructs extremely efficiently (Abbott, 2008) 8. Ease of learning a related task 9. Disuse reduction in memory 10. Mechanism for retaining specificity of memory retrieval 11. Functional integration and operation of hippocampal new neurons in learning and memory 12. Transfer of the basic units of memory for a different learning and retrieval event (Dahlin et al., 2008) 13. Ability to explain the correlation between LTP and memory 14. Ability to explain how consolidation takes place 15. Ability to explain mirror neuron activity 16. Ability to explain some of the features of psychiatric disorders, if possible 17. Ability to replicate the hypothesized mechanism in physical systems, if possible A hypothesis that can provide a broader framework incorporating the above features needs to be built and tested theoretically followed by experimental approaches to confirm its presence. In summary, an ideal hypothesis should be able to substantiate the molecular, cellular, electrophysiological, systems and behavioral features of the brain functions. 2. Summarize the semblance hypothesis and its derivation? The semblance hypothesis (book pdf) was derived to explain associative learning and memory retrieval through a series of inductive reasoning steps using some key assumptions. The main purpose is to build a broader frame work that can incorporate all the elements that were mentioned above. Once this is achieved, fine-tuning will be required to approach towards the cellular and molecular mechanisms. The derivation of the hypothesis has two major stages. Each stage consists of a series of steps that are numbered. Stage I 1. For the purpose of derivation of the hypothesis, memory is viewed as a virtual sensation of a sensory stimulus since the sensory inputs from the item memorized is not present during the retrieval of memory. 2. We store thousands of memories. A specific internal or external cue stimulus is required to retrieve a specific memory. 3. In the classical Pavlovian experiments, the conditioned and unconditioned stimuli were very distinct items reaching different sensory systems. Stimuli reaching different sensory systems are not mandatory for associative learning and its retrieval. Elements within a single learned item (say for example, different features of a bell) can be used as a cue for memory retrieval. 4. Let us now conduct an imaginary experiment. Let us look at a violet-colored pen. While looking at it, let us assume that a specific set of 105 synapses (out of the total 1015 synapses in our brain) were activated at different orders of neurons (1st order being the order close to the sensory level). After one day, let us insert the same number of (105) electrodes into the brain and stimulate the set of those specific 105 synapses. During this stimulation, it is assumed that we are likely to memorize/imagine/visualize that violet-colored pen. This leads to the possibility that by activating the specific set of 105 synapses among the total 1015 synapses, memory can be achieved (Those synapses in the same order of neurons are likely to get activated simultaneously. Those at continuing orders of neurons get activated in a temporal sequence). 5. How can we activate a specific set of 105 synapses out of the total 1015 synapses? Putting in a different way, how can we selectively activate each of those 105 specific synapses from a set of 1015 synapses for retrieving the memory? If we know how we can activate one of those 105 specific synapses that identify the item to be memorized, then we can extend the same mechanism to all the 105 synapses. 6. Alternatively, we can address the issue in a modified way. What is the minimum requirement that satisfies activation of a synapse? Activation of the postsynapse (postsynaptic membrane) can be taken as the equivalent of activating a synapse since the activation of a postsynapse requires the arrival of an action potential at its presynaptic terminal. 7. Since there is no sensory stimulus available from the item to be memorized, we cannot anticipate any action potential reaching at the presynaptic terminal. Therefore, we need to activate the postsynapses of the synapses that represent the learned item without any action potential reaching their presynaptic terminals during retrieval. Activation of a postsynapse without the arrival of an action potential at its presynaptic terminal (at the synaptic level) can very well represent the idea of evoking a virtual sensation of a sensory stimulus (at the systems/behavioral level). In other words, the activation of a specific set of postsynapses that can evoke virtual sensation of a sensory stimulus is expected during retrieval of memory in the presence of a specific cue. 8. At this point we come across with two key questions. 1) Can we activate the postsynapse of a synapse in the absence of the arrival of an action potential at the presynaptic terminal? 2) How can we choose those 105 specific postsynapses from the total 1015 synapses for specific activation? What we have is a specific cue stimulus that activates a specific set of cue-specific synapses. We can now arrive at a simple question at the synaptic level: “How can we activate a specific set of 105 postsynapses that represents the item to be retrieved in the presence of the activation of the specific set of synapses by the cue stimulus? 10. Let us assume that the cue stimulus evokes the activation (we mean depolarization for convenience; other changes are also considered, (see book pdf)) of the postsynapses of the synapses that identify the learned item. Then it is reasonable to argue that some of the synapses through which activity spreads from the cue stimulus should be physically close enough to some of the postsynapses that identify the item to be memorized (mechanism other than physical proximity may operate; for derivation of the hypothesis, physical proximity is used). In addition, a mechanism should exist that can cause the spread of activity from the synapses of the cue stimulus to the postsynapses of the item to be memorized (Fig.1). Cue Learned item
Figure 1. Illustration of the hypothesized depolarization spread during retrieval. During retrieval, the cue stimulus reaching presynaptic terminal A depolarizes its postsynaptic membrane B, and the depolarization spreads to postsynaptic membrane D. This can only happen provided there is a functional LINK (see explanation of functional LINK in step 11) between the postsynapses B and D. Therefore, we can assume that a functional LINK is required to be formed between postsynapses B and D during learning.
11. Since physical closeness between the postsynapses B and D is required, it can be assumed that for associative learning to occur, the sensory inputs from the cue and the item to be retrieved should inevitably converge at some brain locations. (Note: The hippocampus, identified as a location important for learning and memory, receive inputs arriving from all the sensory systems.) What should be the critical change occurring during learning between the synapses that are activated by the cue and the postsynapses that can later represent the learned item during retrieval? We can at least assume that during learning a functional LINK (the word functional was used to indicate that the LINK is a function of the activity in either one of the postsynapses; the word LINK is written in all caps to indicate that it is the key element of the hypothesis and that we have to explore it further to discover its exact nature) is established between the postsynapses of the cue stimulus and the item to be leaned (Fig.2). The functional LINKs as the building units have the advantage that as the cue features changes the postsynapses activated through the functional LINKs changes and the retrieved memory is also changed. Cue Learned item
Figure 2. Illustration showing the hypothesized functional connection (LINK) formed between the two postsynaptic membranes B and D during learning. 12. During learning, co-activation of synapses from the learned item and the cue stimulus needs to induce specific changes that will later allow the cue stimulus by itself to evoke the activation of the set of postsynaptic membranes that belong to the learned item. This leads to the generation of the semblance hypothesis. A unit of memory, in the presence of an internal or external cue stimulus, results from the ability to induce specific postsynaptic events at the synapses of the neurons from the learned item without the requirement of action potentials reaching their presynaptic sides. 13. The proposed LINKs formed during learning could be of different types: a. Functional LINKs: These will not be visible as physical structural connections. For example, an oxygenation-state-dependent functional LINK can be present only above certain oxygen concentrations (see steps 16 for details). b. Function – Structure LINKs: When the functional LINKs are repeatedly activated over a long period of time (by different learning events) they eventually convert to structural LINKs where depolarization spread may occur even at sub-threshold oxygen levels. c. Structural LINKs: These may be determined genetically and can be responsible for some survival behaviors in newborns without any prior learning (examples include grasp, sucking, and rooting reflexes that are used for survival without any prior learning experiences). 14. Let us examine the effects of depolarization spread through the LINKs. As discussed before, activity arriving at the postsynapses of the synapses activated by the cue stimuli will spread through the functional LINKs to the postsynapses that represent the learned item (Fig.3). When a postsynapse is depolarized in the absence of the arrival of an action potential at its presynaptic terminal, then the postsynapse gets the illusion of an action potential reaching its presynaptic side, resulting in “synaptic semblance”. Cue Learned item
Figure 3. During retrieval, the cue stimulus reaching presynaptic terminal A depolarizes its postsynaptic membrane B, and the depolarization spreads to postsynaptic membrane D evoking the cellular illusion at the postsynapse D of an action potential reaching the presynaptic terminal C. This was named “synaptic semblance”. 15. When the related learning events continue, one of the postsynapses (either B or D in the figure 3) will be used to form functional LINKs with the postsynapses of the neighboring synapses (seen as additional postsynapses on the right side of the postsynapse D in the left panel, Fig.4). As this process continues, it will result in the formation of islets of LINKed (LINKable/ re-establishable during retrieval) postsynapses (right panel, Fig.4).
Figure 4 Left panel: Illustration showing LINKable post synapses. Multiple dendritic spine heads belonging to dendrites of different neurons can become functionally connected to each other upon coincident activation. Only two presynaptic terminals (A and C) and two postsynapses (B and D) are marked. Assume that there are nearly one hundred postsynapses arranged in a horizontal plane. The dotted line shows a cross-section across the postsynapses and used in the right panel. Right panel: A hypothetical cross-sectional view of LINKed postsynapses of the synapses in one horizontal plane in a brain region (see the horizontal dotted line across the postsynaptic membranes in the left panel); in this illustration, we imagine that all the postsynaptic membranes are in the same plane. Postsynaptic membranes are shown in small dark circles (broken arrow). When learning occurs, functional LINKs between simultaneously activated postsynapses can be established. Continued learning using any of those synapses will increase the number of interconnected postsynaptic membranes forming islets of functionally LINKed postsynapses (solid arrow). Multiple LINKs between the postsynapses in an islet can cause the spread of excitatory postsynaptic potential (EPSP) across the islet. The individual islets are expected to be functionally separate from each other. 16. Now let us examine whether the role of proposed functional LINKs match with any known physiological changes. It can be done by asking the question, “Can depolarization spread through the establishment (during learning) or re-establishment (during retrieval) of the functional LINKs between the postsynapses depend on already known factors essential during learning or retrieval?” Let us first examine some of the key factors upon which learning and memory retrieval depend on. Then we can examine whether the proposed functional LINKs can have any relationship with those factors. Examples include source energy and availability of oxygen. Experiments using functional magnetic resonance imaging (fMRI) have demonstrated that there is increased oxygen release at locations within the brain both during learning and memory retrieval. The intensification of the fMRI signal at specific locations were suggested to be proportional to the synaptic activity and not to the spiking activity (Logothetis, 2008). Do the hypothesized functional LINK formation (during learning) and its reactivation (during memory retrieval) oxygenation-state-dependent? During the learning and memory retrieval events, increased fMRI signal intensities are often visible at the hippocampus. One plausible explanation for this could be the fact that hippocampus receives sensory inputs from all the sensory systems. The convergence of the sensory inputs at the hippocampus will result in the formation of increased number of functional LINKs between the postsynapses during learning (and their reactivation during memory retrieval). The higher density of active functional LINKs in a small area will cause increased oxygen release/unit area which in turn result in increased fMRI signal intensity in the hippocampus (and also other regions of sensory convergence like amygdala). On the other hand, if isolated synapses are involved during learning and retrieval, as occurs outside the hippocampus, then the increased oxygen release occurring at a small volume of brain tissue may not be sufficient to be visible as increased fMRI signals. 17. Now let us examine the effect of depolarization spread through the functional LINKs. During memory retrieval, the postsynapses belonging to the leaned item are activated by the cue stimulus by depolarization spread through the functional LINKs. This depolarization at the postsynapses can induce action potential generation in their neurons, provided a sufficient number of postsynapses on the dendritic tree of the neuron are activated to produce sufficient spatial and/or temporal summation of excitatory postsynaptic potentials (EPSPs) at their axon hillocks. These neurons, in fact, represent the learned item meaning that they are normally activated by the learned item. The activation of a partial network of neurons that belong to/ represent the learned item without the need of any sensory stimuli from the learned item can lead to “network semblance” (Fig.5). Network semblance means activation of a specific network occurring at a specific neuronal order without activation of its naturally connected (excluding the functional LINKs) penultimate neuronal orders, providing an illusion that sensory inputs reach those penultimate orders of neurons from the learned item.
[Activated partial neuronal network] Figure 5. Schematic representation of network semblance occurring during associative memory retrieval. A cue stimulus activates neuron M. The action potential reaches a LINKable postsynapse (LPS; shown as a shaded triangular area) and a shared extracellular matrix (SEM). In both these structures, excitatory postsynaptic potentials (EPSPs) can spread to the postsynaptic membranes of the neuronal terminals belonging to the learned item. Two consequences occur. First, at the synapses of the axonal terminals of the neuron R belonging to the learned item, a semblance of activity from the learned item occurs. (Note that semblance at the LPS occurs concurrently with the neuronal activity of T. Similarly, semblance at the SEM occurs concurrent with the neuronal activity of S). Second, the spread of EPSP can contribute to the generation of action potentials at the axon hillocks of the neurons S and T belonging to the learned item. Even though semblance occurs at one of the synapses of the neuron U (at the LPS) the EPSP does not lead to the firing of the neuron U. The partial network activated by the cue and the network semblance formed from this network is an important determinant of the sensory characteristics of what is going to get memorized. Computational studies will be able to determine the characteristic contribution of the network semblance in any given memory. 18. It can be concluded that an optimal combination of synaptic and network semblances can finally provide the net semblance for memory retrieval. Stage II In the stage I, we have derived both synaptic and network semblances. How can these semblances be put together to evoke the virtual sensation of the sensory stimulus from the learned item? This is explained through the following steps. Readers please note that these are written as steps of a computational study using an artificial brain.
Similar to the stage I, a series of steps are used to derive a plausible
mechanism for characterizing the specific nature of the memorized item. Let
us assume that the synaptic semblance occurs at one specific CA3-CA1 synapse
(let us compare it to the synapse C-D in figure 3) in the hippocampus. This
means that there is a semblance of depolarization of the presynaptic
terminal that arrives from a CA3 neuron (let us name it Z). Neuron Z can
only be depolarized by activating a set of axonal terminals of the granule
neurons that synapse to the postsynapses (dendritic spines) on the dendritic
tree of CA3 neuron Z. We really don’t know which of the nearly 4x104
presynaptic terminals of the neuron Z are fired (we know that spatial
summation of EPSP from nearly 40 or temporal summation of EPSP from nearly
10 postsynaspes at the soma can trigger an action potential). Therefore, it
is possible that activity from a multitude of possible combinations of
inputs arriving to Z’s nearly 4x104 dendritic spines from the
granule neurons has the capability to trigger the same action potential.
This set of all the combinations of granule neurons
Figure 6. An illustration of synaptic semblance at one synapse. The semblance of activity is induced from all the possible neurons that can activate it through the connection at the functional LINKs. The bottom circles represent postsynaptic membranes of the learned items D1 and D2 where functional LINKs from postsynaptic membranes B1 and B2 of the cue stimulus induce semblance. Note that the semblance from D1 and D2 in neuronal order 1 overlaps at the neuron N. All other smaller circles represent neurons. Note that the net semblance represents the identity of the sensation likely to reach the overlapping neuron N. The overlapping of semblance from more functional LINKs above a threshold value provides enough strength of semblance for memory. The numbers on the left-hand side represent the order of neurons from the sensory endings towards the higher orders. Summary of the computational model: An artificial brain model can be developed to initiate the computational studies. Sets of neurons from different sensory systems can be assigned at multiple neuronal orders. Postsynapses of the synapses from different preceding neurons are likely to be closely placed at different locations. Based on the closeness of the postsynapses and their simultaneous activation, functional LINKs can be formed during learning. In this model, we can assign values of semblances from each postsynapse that have the potential to establish functional LINK during different learning. We can also assign potential network semblance values at each order of neurons. Both the synaptic and network semblances from each LINKable postsynapse and neuronal network respectively will produce independent sensory maps, sensory locations from which activity can arrive at the CA3 neuron Z discussed in the earlier paragraph. A database of these sensory maps for each synaptic and network semblance can be made for a given nervous system. Learning is carried out by using a cue and an item to be learned. Expected functional LINKs that are likely to be formed between the postsynapses of the neurons belonging to the cue and the item at different neuronal orders (where the synapses belonging to the cue and the learned item come to functionally LINKable distances) are marked. Then, introduce the cue for retrieving the memory. Identify all the specific synaptic and network semblances that occur at the LINKed postsynapses in the presence of the cue. Plot the sensory receptor location map from each of the synaptic and network semblances. Next, overlay these sensory receptor location maps and make a sensory identity map of the item retrieved (Fig.7). By comparing its characteristic features with that of the learned item fine-tune the artificial nervous system to obtain a good match of the sensory density map. This fine-tuning can be carried by changing the: i. Number of neuronal orders ii. Orders at which neuronal axons converge for potential functional LINK formation during learning iii. Number of the functional LINKs at each neuronal order iv. The extent of functional LINKs already present from previous learning v. The extent of the partial neuronal networks vi. Weights for synaptic and network semblance vii.Modifying some neuronal orders by incorporating inputs from all the sensations (similar to that of the hippocampus) viii.Introducing changes for equivalent effects from CA3 recurrent collaterals ix. Introducing smaller islets of functional LINKs in the cortex and larger islets in the neuronal order matching CA3 (based on the experimental findings shown in figure 8. x. Possible permutations of the temporal activation of synapses required for action potential generation xi. Possible combinations of the synaptic activities required for action potential generation xii.Factors like neurogenesis, and addition and deletion of new synapses. These factors can lead to certain level of fluctuation of the identity of the retrieved memory from a given cue over time. Since there are many factors that need to be fine-tuned at a given time, it will require a huge computational effort. It is possible that a novice nervous system (corresponding to a nervous system belonging to a child in the first few years of life) when first designed in computational models needs to be primed with many associative learning events prior to the test learning and retrieval experiment. This is because a nervous system that has undergone many previous learning events will already have functional LINKs that can get activated by a new learning event. The activations of these previously induced LINKs may be essential for reaching a threshold semblance for memory retrieval after a learning event and may depend on the complexity of the nervous system. These assumptions were drawn from the developing brain during the initial years of life, a period during which memory storage and retrieval are at a developing stage. However, the above suggestion may only hold true for complex nervous systems like those of humans. In simple organisms, semblances for memory retrieval achieving features of the items learned may occur even without much prior learning events. However, the features of the memory and the number of items memorized may vary depending on many factors that determine the net semblance.
Figure 7. The formation of specific features of the sensation of memory from semblances occurring at various neuronal orders. The semblance of the identity of the sensory stimulus occurs at: I) Two synapses at an early order II) Two synapses at a later order III) One synapse at a late order IV) Multiple orders of neurons. The area along the horizontal plane (containing the horizontal line) and the darkness of the shade representing the overlapped semblances determines the identity of the sensation. The darkest triangular area represents the identity of the sensory stimuli memorized as it is the net semblance that identifies the sensory input. The black circles symbolize synapses representing the learned item where semblance occurs during retrieval. The numbers at the middle represent the order of neurons from the sensory endings towards higher orders. Note that the nature of semblance occurring at the higher order of neurons will be different than that occurring at the lower levels.
In the next figure (Fig.8), we derive the basic units of semblances
occurring at the postsynapse activated through the functional LINKs.
Activation of neuron N or postsynapse D leads to semblance of activity
coming from neuron Z in the neuronal order 8. Neuron Z is normally
depolarized by activating a set of axonal terminals of the neurons in order
7 (in Fig.8) that synapse to neuron Z’s dendritic spines (postsynapses). The
spatial summation of nearly 40 or the temporal summation of less than 40
EPSPs (from nearly 40 postsynapses (dendritic spines) out of the nearly 5×104
postsynapses of each neuron) triggers an action potential at neuron
Z’s axon hillock. Let the set of all combinations (for the spatial summation
of EPSPs) and permutations (for the temporal summation of EPSPs) of the
neurons in neuronal order 7 whose activity through both normal synaptic
transmission and activity-spread through the functional LINKs ((A) and (B)
respectively (in the inset of Fig.8). It is likely that different neurons in
set
Fig. 8. From Vadakkan K.I (2011) Processing semblances induced through inter-postsynaptic functional LINKs, presumed biological parallels of K-lines proposed for building artificial intelligence. Frontiers in Neuroengineering (2011) 4: 8
Schematic representation of sensory elements induced during the activation
of a synapse or a neuron. The gray circles represent neurons. The numbers on
the left side of the neuronal orders denote their position in relation to
the sensory receptors. Neuron Z is shown in neuronal order 8. During memory
retrieval, a cue-stimulus (marked by asterisk) reaching presynaptic terminal
A depolarizes its postsynaptic membrane B and the resulting EPSP at
postsynapse B re-activates the functional LINK that activates postsynaptic
membrane D (Mechanisms other than depolarization are also considered
(Semblance hypothesis of memory. 3rd edition. iUniverse Publishers,
Bloomington, U.S.A). When postsynaptic membrane D is depolarized, it evokes
the cellular hallucination of an action potential reaching its presynaptic
terminal C. This is called synaptic semblance. Note that presynapse C
belongs to the neuron Z. Either synaptic semblance occurring at the
postsynapse D or random activation of neuron Z produces the hallucination
that it is receiving inputs from the set of neurons
Similar to the above mechanism, the neurons in set
It
is likely that activation of subsets of a minimum number of sensory
receptors from As the functional LINKs get re-activated during memory retrieval, the expected EPSP spread that may occur through some of these functional LINKs can be crucial in adding to the existing sub-threshold EPSP at the axonal hillocks of some neurons that are routinely activated by the oscillatory neuronal activities in the hippocampus and cortex as well as from baseline sensory activities arriving at many neurons. Since the number of functional LINKs continues to change (due to continued associative learning) over the life span of the nervous system, the characteristic features of the semblions are also expected to change gradually. This will lead to gradual changes in the net semblances for memory. Related associative learning can increase the number of LINKed postsynapses and increase semblance. The net semblance can exceed more than the threshold without any effect on the retrieved memory. 3. What is the nature of the memory code mentioned in the hypothesis? During memory retrieval, the cue stimulus activity propagates through different orders of neurons. First, note that there is a temporal order of activation of synapses. So, we can say that permutation of a set of postsynapses that are activated occurs as activity spread from one neuronal order to others. Due to the synaptic addition, deletion and neurogenesis, sequences in the permutation of the sets of postsynapses will keep changing. Some values in the sequence will be lost or replaced or new sequences are added. At each order of neuron, it is likely that the synapses will get activated together. So, we can say that combination of a set of postsynapses activated needs to be calculated at those points. Both the synaptic and network semblances are used to assess the net semblance that qualifies the item to be memorized. Each memory has a labile code consisting of the subset of postsynaptic terminals and neuronal networks belonging to the learned item that are activated during retrieval, out of the set of postsynaptic terminals and neuronal networks that were co-activated with the cue during learning. This code can be used for computing the identity of the retrieved memory. In summary, 1) The semblance code is the set of postsynapses, activated by the cue, representing the learned item 2) This code will be different for different cues 3) This code induced by a given specific cue will keep fluctuating from time to time due to synapse addition, elimination and neurogenesis 4) A typical code may be represented as in figure 9.
Figure 9. 1, 2, 3, 4: orders of neurons activated in a temporal order (permutation); ABC… sets of synapses activated at each order of neurons at different time points (permutation); A1B1C1: the 1 stands for the number of the order of neurons starting from the sensory ending; abcd: synapses at each order of neuron that are activated simultaneously (combination); a1b1c1d1: the 1 here denotes that a,b,c,d,… are synapses of the neuron A of the neuronal order 1 For the retrieval of a specific memory, there will be a minimum cue requirement. Improving the cue further may not change the qualities of the item retrieved for a given nervous system for a given amount of prior learning that it had undergone. 4. What are the basic differences between this hypothesis and the existing hypotheses? a. The semblance hypothesis provides a feasible mechanism for retrieving memories at physiological time-scales. b. Based on the semblance hypothesis, memory is viewed as the virtual sensation of a sensory stimulus in its absence. The mechanism that can induce such a virtual sensation was derived by the semblance hypothesis. c. Functional LINKs (oxygenation-state dependent transient functional LINK is one of the possible types) between the postsynaptic membranes is the fundamental unit proposed by the semblance hypothesis. The concept of functional LINKs between the postsynaptic membranes is different from the strengthening of single synapses by Hebbian learning. At least a pair of synapses whose postsynapses belong to either one (inducing synaptic semblance) or two different neurons (inducing synaptic and effective network semblance) is the necessary building block of semblance hypothesis. d. Semblance starts at the neuronal order where the cue and the learned item stimuli converge. Synaptic and network semblance start occurring from these neuronal orders. The weight of the semblances at higher orders above this level may vary. e. According to the semblance hypothesis, if inputs from two stimuli do not meet at a sufficient number of LINKable postsynapses at their higher orders of neuronal connections, then learning between them (in terms of efficient memory retrieval by evoking sufficient net semblance) is not possible. f. The semblance hypothesis can explain the retrieval of memories at physiological time-scales since semblances occur at the synapses and neuronal network concurrent with the spread of activity from the cue stimulus. g. Overwriting of synaptic strengths formed by a single learning event by new learning has been a concern raised about the previous hypotheses (Fusi and Abbott, 2007). Based on the semblance hypothesis, established functional LINKs can be transferred to use in a different learning and retrieval as shown in experiments (Dahlin et al., 2008). Depending on the specific cue characteristics, the specific characters of the net semblance of the retrieved item can utilize these previously LINKed (and LINKable during retrieval) for memory retrieval. h. The semblance hypothesis has the potential to be extended for building frameworks that can possibly explain many physiological and pathological brain conditions. i. Memory retrieval need not depend on the location of semblance. It is more likely to depend on the net strength of semblance from all the orders of neurons. 5. How can we retrieve memories after many years (even without retrieving it in between)? During learning, functional LINKs form between the postsynapses (Please note that the functional LINKs are transient, likely oxygenation-state dependent LINKs that are a function of postsynaptic activation of synapses belonging to the cue stimulus). As related or unrelated learning takes place over many years, islets of functional LINKs are formed and get expanded (Fig.3). Each postsynapse in an “islet of LINKed postsynapses” (Fig.4) is likely to get activated during many events of related or unrelated learning or memory retrieval. This is possible through the proposed spread of depolarization through the functionally LINKed postsynapses. Let us now consider an important point. Let us assume that a specific learning event had taken place 20 years ago in our brain. We assume that on the next day a cue stimulus can activate a specific set (n) of postsynapses of the learned item (through the functional LINKs) to evoke semblance leading to memory. Let us also assume that continued learning in the following years had caused activation of those (n) postsynapses (few at a time during isolated events of different learning and memory retrieval) through the functional LINKs. As long as sufficient number of LINKs from other postsynapses to the (n) postsynapses under question is maintained (by different event of learning and memory retrieval), we are in a position to activate them using an appropriate cue that was used during the initial learning that took place 20 years ago. Therefore, whenever the specific cue is used to recall a memory, the specific set of postsynapses can be activated to evoke the semblance for the specific memory. This feature can provide the ability to form memories without saturation and with unlimited life-times. It also correlates with the transferability of the functional properties of the LINKed postsynapses in different learning and retrieval events (book pdf). 6. How can we explain long term potentiation (LTP) in terms of the semblance hypothesis? The semblance hypothesis was derived to explain plausible synaptic changes occurring during learning suitable for evoking virtual sensation of a sensory stimulus during memory retrieval. The operational principle of the semblance hypothesis is completely different from that of LTP. However, one general argument is that any new hypothesis of memory should be able to explain the relationship between LTP and memory. A plausible reasoning for the relationship between LTP and memory through the semblance hypothesis is done by explaining the following experimental findings. It has been shown that one-trial inhibitory avoidance learning in rats produced the same changes in hippocampal glutamate receptors as the induction of LTP with high-frequency stimulation (Whitlock et al., 2006). This study showed that learning-induced synaptic potentiation occludes high-frequency stimulation-induced LTP. Based on the semblance hypothesis, these findings can be explained as follows. a. Learning first followed by LTP induction According to the semblance hypothesis, prior learning events in a caged environment would have already made many islets of LINKed postsynapses (dendritic spines) in the hippocampi of the rats as explained in figure 1. Since associative learning opportunities are finite during caged life, we can expect a slow expansion (by LINKing more postsynapses with additional related learning events) of discrete islets of LINKed postsynapses as the rats grow up. When rats undergo avoidance learning (a novel instance of associative learning), we can expect the formation of functional LINKs between two or more islets of functional LINKs that are already present in the animal. Even though this is particularly important in this experimental context, it will also hold true in any novel associative learning. In experiments using inhibitory avoidance testing (Whitlock et al., 2006), not all the recording electrodes recorded an increase in field excitatory postsynaptic potential (fEPSP) slope, indicating that ionic changes at the locations of the tips of these electrodes (CA1 dendritic tree) required to produce an increase in fEPSP slope did not take place. However, among those electrodes that recorded an increase in fEPSP slope after inhibitory avoidance learning, a sufficient number of Shaffer-CA1 synapses were potentiated. Let I and II stand for two islets of functionally LINKed postsynapses that were already present in the animal before the avoidance learning session. During learning, it is likely that LINKs were formed between the islets (islets of LINKed postsynapses) I and II. This will generate a sudden increase in the size of an islet of LINKed postsynapses to nearly two-fold, forming a mega-islet of LINKed postsynapses (Fig.10). Figure 10. Illustration showing potential LINKable site between islets of postsynapses (please see the figure 4 for details of the islets; they are visualized by hypothetical cross-sectional view through functionally LINKed postsynapses) that belong to two different CA1 neurons. During an associative learning, LINK formed between the postsynapses (marked with asterisks) of islets 1 and 2 (large circles) can lead to the formation of a mega-islet that can continue to contribute to the LTP recorded from the recording electrode as explained in the text. Position of the stimulating electrode is at the Schaffer collaterals.
Illustration explaining the basis of long term potentiation (LTP) based on the present hypothesis. Shaffer collateral from the CA3 neurons synapse to the dendritic spines (postsynapses) of the CA1 neurons. Many of these postsynapses are functionally LINKed to form islets in an animal (see figure 5 for details of the islets of functional LINKs). Here two such islets I and II (large circles) are shown. One of the postsynapses from each of the islets I and II is shown to continue towards the soma of the CA1 neurons. Activation of any one of the postsynapse within an islet will result in EPSP spread towards the somas of the CA1 neuron. The islets are formed between postsynapses that are concurrently activated during previous associative learning. During an associative learning of a novel item or during induction of LTP (note the position of the stimulating electrode is at the Schaffer collaterals), a new functional LINK may form between the postsynapses (marked asterisks) of islets I and II. This can lead to the formation of a mega-islet combining the two islets. This can contribute to the LTP recorded from the recording electrode as explained in the text. Activation of a postsynapse of this mega-islet of LINKed postsynapses can cause spread of depolarization between its postsynapses. Since a subset of postsynapses in the mega-islet already LINKed to one of the dendritic spines (postsynaptic membrane) on the dendritic tree of one CA1 neuron, multiple EPSPs from this subset will reach the main dendrite of a CA1 neuron simultaneously. This results in a summated EPSP at this dendritic location sufficient to produce a corresponding increase in current sink in the extracellular matrix. Immediately following the associative learning event, a proportion of sensory inputs reaching the animal for a long duration of time is likely to activate the postsynapses of this mega-islet, leading to prolonged activation of the main dendrites of the above CA1 neuron (until the CA1 neuron begins homeostatic mechanisms to reduce this prolonged and increased EPSP generation). The extracellular signal recorded from the apical dendrites of a population of pyramidal neurons in the stratum radiatum of the CA1 region in response to Schaffer collateral stimulation, namely the field EPSP, will now show an increase in amplitude and contribute to an increase in fEPSP slope for a long duration of time (LTP). This learning-induced LTP can occlude further LTP induction. b. LTP induction first followed by learning The occlusion process explained in the study (Whitlock et al., 2006) can be considered a bidirectional process, meaning that the induction of LTP in a sufficient number of synapses that are involved in inhibitory avoidance learning will prevent consequent avoidance learning. It is likely that hundreds of axons of the CA3 neurons in the Schaffer collateral pathway are activated by high-frequency stimulation (LTP induction), activating the postsynapses (dendritic spines) of a CA1 neuron. During this process, many postsynapses can get functionally LINKed due to the simultaneous activation of closely placed postsynapses by high-frequency stimulation (assuming that sufficient oxygenation state is present during this process). Some of these LINKs will occur between the islets of already LINKed postsynapses, leading to the generation of mega-islets. Following this, the activation of one or more postsynapses by a regular stimulus (not high frequency) can lead to the spread of depolarization between the postsynapses within the mega-islet. Since one or a small subset of postsynapses in the mega-islet originates from the dendritic tree of a single CA1 neuron, multiple EPSPs from these postsynapses can reach one dendrite of a CA1 neuron simultaneously. This results in an increase in the EPSP at these dendritic locations, leading to LTP. This artificially-induced LTP can occlude further learning-induced LTP. If we can artificially induce LTP in a sufficient number of fibers that are critical for the learning, then the animal may not become successful in associative learning using those synapses following the LTP induction. When we say that the animal cannot learn, we mean that the animal cannot retrieve the memories; i. e., when a cue stimulus tries to retrieve a memory using these synapses, the induced depolarization spreads across all those postsynapses that are LINKed by the LTP induction. The retrieval using a specific cue now induces synaptic semblances at all those LINKed postsynapses in the mega-islet, some of which were non-specifically LINKed during the LTP induction. Activation of those non-specific postsynapses will also lead to the activation of non-specific neurons, leading to the induction of non-specific network semblances that are not related to the learned item. In other words, the expected specificity of semblance for the learned item gets diluted by the large amount of non-specific semblances, preventing specific memory retrieval. Note: Please refer to three possible types of islet of LINKed postsynapses explained in the book. 7. Can we explain brain functions other than memory using the basic units proposed by the semblance hypothesis? If the basic units proposed by the hypothesis are correct, it can be extended to examine other brain functions. With this aim in mind, explanation for various required features that are highlighted in the “Why do we need a new hypothesis of memory” section were sought and explained (book pdf). An example is the ability to extend the hypothesis to explain some of the features (delusions, cognitive defects and a possible mechanism for dopamine receptor antagonists relieving the symptoms of delusions) of schizophrenia. 8. Can we explain behavior in terms of the semblance hypothesis? Memory retrieval is often tested in animals by observing the behavioral patterns that the animal exhibits. Behavior is expressed as a combination of different motor activities. The semblance hypothesis may explain how this can occur. In the case of fear conditioning experiments, the conditioned reflex of a shock to the foot induces the animal to withdraw the leg. Therefore, it can be taken that the neuronal activity induced by the foot shock activates a specific set of motor neurons at that particular neuronal order that induce withdrawal of the foot. During the retrieval of memories using the unconditioned stimulus, we need to explain a mechanism by which the foot can be withdrawn. Based on the semblance hypothesis, a cue stimulus (unconditioned stimulus) activates a partial neuronal network that belongs to the learned item. This partial neuronal network can activate the specific set of motor neurons leading to foot withdrawal. This will basically depend on the qualities of the subset of neurons that form the partial neuronal network which in turn depends on the cue characteristics. In short, the cue stimulus activates a partial neuronal network that belongs to/ represents a part of the learned item that stimulates the motor neurons that lead to the foot withdrawal. 9. Is there any experimental evidence supporting the present hypothesis? a) Presence of multiple spine heads on a single spine neck seen in excrescences at CA3 dendritic tree (Amaral and Dent, 1981; Chicurel and Harris, 1992; Frotscher et al., 1991) suggests possible structural evidence for the possibility of synaptic semblance as a mechanism for memory. Development of functional/ structural LINKs between the postsynaptic segments at an excrescence can induce synaptic semblance. However, since the neuron downstream from each of the synapses on an excrescence is the same CA3 neuron, it will activate the same partial neuronal network in the higher orders of neurons (same neuronal network activation will not contribute to specificity in semblance). b) According to the semblance hypothesis, oxygenation-state-dependent functional LINKs between postsynapses are the possible functional units. A critical experiment is to prove the presence of functional LINKs between the postsynapses of neurons through which sensory inputs travel from different elements of the learned item. For this, two important individual steps have to be proved. 1) Learning can induce functional LINKs between individual postsynapses, and 2) The functional LINKs formed during learning can be used for depolarization spread during retrieval of memory. For this, electrode tips of < 2 µm (normal diameter of a postsynapse) are required. Technical advancement for specifically stimulating the postsynapses using such small electrode tips needs to be developed for this experiment to be carried out. However, we have addressed the question of whether an oxygenation-state dependent chemical reaction spread can be shown between the postsynapses using a chemical procedure. We developed a Golgi-staining protocol with an aim to examine oxygenation-state-dependent functional LINKs in a mature animal (that has already undergone plenty of learning in life and therefore would have plenty of functional LINKs established). We observed the spread of Golgi staining from one dendritic spine to the next in an oxidation-state-dependent fashion. In the original Golgi staining, we observe individual dendritic spines. Golgi staining uses a very strong oxidizing agent to oxidize the brain tissue as the first step, followed by exposure to silver nitrate that leads to the formation of a silver chromate stain in individual dendritic spines. We carried out various experiments to modify the oxidation state of the tissue. In one procedure, when we initially reduced the tissue through the vascular route (capillaries -> astrocytic pedocytes -> synaptic locations) and followed by Golgi staining, we observed the chemical spread of sliver chromate formation between the dendritic spines (Fig.11). c) The result of this experiment only suggests the possibility of oxidation-state dependent functional LINKs between the postsynapses. It does not provide any evidence for functional LINKs as suggested by the semblance hypothesis. More experiments are required to prove the presence of functional LINKs. The results of the experiment provide some precedence for carrying out experiments to examine whether an oxygenation-state-dependent functional LINK formation occurs between the postsynapses in physiological conditions. From the nature of the oxidation-state dependent spread of Golgi stain, it can be seen that such spread occurred between the dendritic spines belonging mostly to different neurons. This finding supports the proposed mechanism for network semblance.
Figure 11. A. Dendritic spines seen in a) conventional Golgi staining and b) modified Golgi staining. Note that the spines made visible by the modified Golgi staining have a white area on/around them. It is possible that this marks the location of the presynaptic terminal synapsing on them. Alternatively, it is possible that the dark stain is present around the dendritic spines (Scale bar = 2μm). B. cortical neuron stained by the modified Golgi staining protocol. Notice the presence of many chemically interconnected dendritic spines at the dendritic branch terminals (8 ± 5.6 (S.D); n = 113 clusters). Also note the presence of normal dendritic spines on the dendritic shafts (Scale bar = 6μm). C. Large chemically-interconnected cluster of dendritic spines on the dendrite of a CA3 neuron in the hippocampus. Note that these large clusters of interconnected dendritic spines (482 ± 106 (S.D); n = 106 clusters) have more than thirty times the number of dendritic spines on a single spine neck reported on a dendritic excrescence 10. How can we replicate the proposed hypothetical model in a physical system? At this point it is only possible to speculate. An artificial system similar to that of the nervous system can possibly be created using thin layer chemistry with multiple layers of thin layer chemicals. Oxygenation-state-dependent functional LINKs can then be established between the postsynapses (identified spots on the chemical layers). Once learning is complete, then the cue stimulus can be used to induce both the synaptic and network semblances. The first question here is how to test that the system has memory. In real-life situations when we test animals’ memory retrieval abilities, they show behavioral features like foot withdrawal (in fear-conditioning experiments). This indicates that motor neurons are activated, inducing muscle contraction of a particular group of muscles. According to the semblance hypothesis, a partial neuronal network (belonging to the learned item) is activated by the cue stimulus and is responsible for downstream activation of the motor neurons. We can incorporate a signal system that gets activated whenever the specific partial networks that represent the item to be memorized are activated. The second question is how to test a system that does not have a motor system to manifest memory retrieval. This becomes especially important when we plan to develop intelligent machines. Here, the artificial system will need an extra segment that calculates the semblances resulting from a specific cue and provides us with the net semblance (additional computation can be used for this purpose). As discussed earlier in the discussion part of the derivation of the hypothesis, for a given nervous system all the possible semblances from individual synaptic and network semblances can be calculated initially. From this calculation, the net semblance during memory retrieval by a given cue can be estimated. Equivalent changes in the system for dynamic synapse addition, elimination (computational studies may reveal whether these changes occur by cause or effect) and neurogenesis need to be incorporated. 11. Is there any direct evidence for the hypothesis? At present, there is no direct evidence to prove the hypothesis. We agree that direct evidence both by computational methods and biological lab experiments are required to prove the hypothesis. However, this hypothesis makes a case for the examination of nature of functional LINKs between the post-synaptic membranes especially those belonging to different neurons (that can provide network semblance). While possible oxygenation-state dependent transient functional connections were examined in preliminary experiments (see results, book pdf), the exact nature of the functional connections need to be explored further. 12. Is memory retrieval an active process? Induction of both synaptic and network semblances at each neuronal order that can be computed to form a multi-dimensional net functional semblance of activity in the presence of the cue is viewed as memory semblance (discussed in figure 6, book pdf). This can be viewed as possible in a system that receives background neuronal activity (contributed by both background sensory stimuli and oscillating neuronal activities) as a continuum. Thus based on semblance hypothesis, the events that lead to the activation of specific set of postsynapses without the activation of their corresponding presynapses is an active process; whereas formation of semblances is a virtual internal sensation. References Abbott, L.F. 2008. Theoretical neuroscience rising. Neuron. 60:489-95. Amaral, D.G., and J.A. Dent. 1981. Development of the mossy fibers of the dentate gyrus: I. A light and electron microscopic study of the mossy fibers and their expansions. J Comp Neurol. 195:51-86. Chicurel, M.E., and K.M. Harris. 1992. Three-dimensional analysis of the structure and composition of CA3 branched dendritic spines and their synaptic relationships with mossy fiber boutons in the rat hippocampus. J Comp Neurol. 325:169-82. Dahlin, E., A.S. Neely, A. Larsson, L. Backman, and L. Nyberg. 2008. Transfer of learning after updating training mediated by the striatum. Science. 320:1510-2. Frotscher, M., L. Seress, W.K. Schwerdtfeger, and E. Buhl. 1991. The mossy cells of the fascia dentata: a comparative study of their fine structure and synaptic connections in rodents and primates. J Comp Neurol. 312:145-63 Fusi, S., and L.F. Abbott. 2007. Limits on the memory storage capacity of bounded synapses. Nat Neurosci. 10:485-93. Logothetis, N.K. 2008. What we can do and what we cannot do with fMRI. Nature. 453:869-78. Martin S.J, Grimwood P.D, and Morris R.G (2000) Synaptic plasticity and memory: an evaluation of the hypothesis. Ann Rev Neurosci 23: 649-711 Piorazi P and Mel B.W (2001). Impact of active dendrites and structural plasticity on the memory capacity of neural tissue. Neuron 29, 779-96. Popper, C. 1965. The logic of scientific discovery. Rubin, D.D., and S. Fusi. 2007. Long memory lifetimes require complex synapses and limited sparseness. Front. Comput Neurosci. 1:7. Whitlock, J.R., A.J. Heynen, M.G. Shuler, and M.F. Bear. 2006. Learning induces long-term potentiation in the hippocampus. . |
This site was started on 10/25/2008
