Please read the hypothesis (book pdf)/ FAQs before reading this page. Findings in the following published work are examined through the semblance hypothesis with an aim to test whether the hypothesis can explain some of the key findings in these papers.

1.   Kay M. Tye, Garret D. Stuber, Bram de Ridder, Antonello Bonci1, and Patricia H. Janak (2008). Rapid strengthening of thalamo-amygdala synapses mediates cue-reward learning. Nature 453: 1253-7 PubMed

Intra-lateral amygdala (LA) N-methyl D-aspartic acid (NMDA) blockade impaired reward-learning acquisition, indicating that glutamatergic synapses are involved in learning. This study has shown that the miniature excitatory postsynaptic potential (mEPSP) amplitude increases after learning. An increase in mEPSP amplitude indicates an increase in the number or function of alpha-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptors (AMPARs)¹. One possibility is that the increase in mEPSP amplitude could be explained as a function of the additional AMPA channel currents from the functionally LINKed^* postsynapses. It is possible that the oxygenation state-dependency for forming functional postsynaptic LINKs in the LA is less probably due to the formation of (F→S→F) LINKs. This may explain why it was possible to observe the effect through these LINKs during electrophysiological experiments using brain slices. Alternatively, oxygen available in the bath solution could have been sufficient to activate the already-formed functional LINKs during learning.

The AMPAR/NMDAR ratio is increased at the thalamo-amygdala synapses. Based on the semblance hypothesis, during learning, functional LINKs are formed between the postsynapses belonging to the sensory inputs from the cue and the reward. Since thalamic afferents receive sensory afferents from both the cue and the reward sensory inputs, it can be assumed that at the locations where the thalamic afferents synapse on to the dendritic spines of the pyramidal neurons of LA, functional LINKs are formed during learning. When thalamic afferents were stimulated to measure the EPSCs by patch clamping the pyramidal neurons at the LA, an increase in the amplitude of the AMPA current can be observed after learning. It is possible that the measurement included AMPA currents from the neighboring synapses that were functionally LINKed during learning.

* The word LINK is not an abbreviation; it is capitalized to highlight its function.

2.  Adam Kepecs, Naoshige Uchida, Hatim A. Zariwala and Zachary F. Mainen (2008). Neural correlates, computation and behavioural impact of decision confidence. Nature 455: 227-31 PubMed             

In this study, using a mixture of odorant cues at different concentrations, the neuronal firing in the orbito-frontal cortex was measured as a function of decision-making strategies. Some of the observations can be explained in terms of the semblance hypothesis. One, some neurons fired during the correct decision making and some others fired during the wrong decision-making process. Based on the network semblance, it is conceivable that a strong cue activates specific neurons through its functional LINKs as explained in figure 6 in Semblance hypothesis. A weak cue (mixture of odors), definitely has a partial cue for the real item.  It can also activate some neurons through the functional LINKs. However, the network semblance formed by this neuronal activity need not be sufficient to evoke the correct memory. This leads to wrong decision making.

3.  Jeffrey D. Karpicke, Henry L. Roediger III  (2008). The critical importance of retrieval  for learning. Science 319: 966-8 PubMed

This study has shown that repeated retrieval induced through testing (and not repeated encoding during additional study) produces large positive effects on long-term retention of memory. This indicates that retrieval itself can strengthen the functional LINKs made during learning better than re-learning. This means that once initial learning achieves retrieval-efficiency, retrieval sessions alone can activate the partial network of the already formed neuronal network belonging to the learned item through the functional LINKs.

4.  Erika Dahlin, Anna Stigsdotter Neely, Anne Larsson, Lars Bäckman, and Lars Nyberg (2008). Transfer of learning after updating training mediated by the striatum. Science 320: 1510-2 PubMed

Some brain regions may be used for a large number of learning events. There will be limitations in maintaining large numbers of islets of functional LINKs within a given region. In situations where functional LINKs are mainly present in one order of neurons in a particular brain region in a specific learning paradigm, it is likely that some of the established functional LINKs may be shared by new related learning. Specific overlapping of the components in different learning involving the same brain region can thus contribute to an improvement in performance on an untrained task, indicating the possibility that some of the established functional LINKs can be transferable. 

5.  Hagar Gelbard-Sagiv, Roy Mukamel, Michal Harel, Rafael Malach, and Itzhak Fried (2008)Internally generated reactivation of single neurons in human hippocampus during free recall Science 322: 96-101 PubMed

According to the semblance hypothesis, the same cue will activate the same neuronal network that was activated by the learned item to elicit network semblance immediately after learning. Therefore, it is expected that during specific memory recall firing by the same neurons can occur. 

6.  Iori Ito, Rose Chik-ying Ong, Baranidharan Raman and Mark Stopfer (2008). Sparse odor representation and olfactory learning. Nat. Neurosci 11: 1177-84 PubMed  

This study has shown that Hebbian spike-time dependent plasticity in Kenyon cells is not correlated with the odor-elicited reward memories, raising possibilities for still undiscovered neural substrates for learning². It was previously reported that the cue can induce persistent sub-threshold excitatory postsynaptic potentials at the synapses to Kenyon cells³.  Based on the semblance hypothesis, the reward stimuli reaching at the LINKable synapses close to the synapse from the cue stimuli to the Kenyon cells can form functional LINKs at an optimal time following the cue stimulus. The following reward stimulus initiates activity in a network of neurons specific to it. In the presence of the cue, memory can be retrieved using synaptic and network semblance in the neural network using the functional LINKs. The Kenyon cell can be compared to neuron U in figure 6 in the Semblance Hypothesis (see the book pdf).

7. Daniel D. Ben Dayan Rubin and Stefano Fusi (2007). Long memory lifetimes require  complex synapses and limited sparseness. Frontiers in Computational Neuroscience 1:1-14 PubMed  

This paper is used as a prelude to examine the theoretical requirements for any system with efficient memory storage capabilities. The theoretical requirements examined in this paper should hold true for any alternate hypothesis for memory storage. Here, I examine whether the semblance hypothesis can satisfy these theoretical requirements. In addition, I use some of the observations made by the authors of this article (for example, the inferotemporal cortical paradox) to examine the semblance hypothesis. The explanations made here will only hold true when mathematical proofs are derived and biological evidence for functional LINKs between the postsynapses is established.

1.   Provision for unlimited life-times.

Functional LINKs that form between postsynapses are viewed as the fundamental units of the semblance hypothesis. Each postsynapse in an “islet of LINKed postsynapses” is likely to get activated during many events of related or unrelated learning or memory retrieval. This keeps the LINKs functionally active throughout life. Therefore, whenever a specific cue is used to recall a memory, these functional LINKs can be used to evoke the semblance of a specific memory, forming memories with unlimited life-times. It also correlates with the transferability of the LINKed postsynapses in different learning and retrieval processes (see book pdf). If a non-LINKed postsynapse is LINKed to an islet of LINKed postsynapses in a new learning event, it may lose its ability to stay LINKed unless the newly LINKed postsynapse is used continually to keep the newly made functional LINK in a LINKable state.

2.  Absence of overwriting of old memories with new ones

The overwriting of old memories happens if we have a system that requires changing an existing structure to give way to a new structure that is formed by additional learning. Based on the semblance hypothesis, there is an expanding islet of functional LINKs between the postsynapses that do not face this difficulty. Instead of overwriting, new memories can in fact use the existing functional LINKs between the postsynapses through transferability as mentioned earlier (see book pdf). New memories can therefore be easily registered if a large number of pre-made functional LINKs exist from the previous learning events. In a system with a limited number of postsynapses that can be LINKed, the number of items that can be learned may be limited once the combinations of postsynapses that can be used to determine the specific semblances for each item to be learned are exhausted. However, with the human brain’s many orders of neurons and very large numbers of synapses, the possible number of combinations of postsynapses can create very large number of semblances for memories.

3.  Absence of decay of the memory trace by any modifications of the basic units by new learning

According to the semblance hypothesis, new learning uses many existing functional LINKs. In addition, associative learning of new components leads to the expansion of many islets of existing functional LINKs. During retrieval, a specific subset of functionally LINKed postsynapses is activated. Therefore, based on the semblance hypothesis, decay of the memory trace by any modifications of the basic units by new learning is not an issue; rather, the activation of a set of specific postsynapses, some of which are interchangeably used by different cues in different memory retrieval events, is a feature by which net semblances from different synapses can contribute to different memories. The functional LINKs formed by a specific event of learning can be maintained by using those LINKs either by learning or retrieval events by either related/ unrelated learning or retrieval events that use them. Therefore, a cue appearing after many years will be able to induce specific net semblance for memory if the functional LINKs induced by its particular learning has been maintained by related or unrelated learning events. By extension, it can be argued that when a less specific cue is present, the specificity of the item retrieved may be reduced (in part because more than one item is retrieved) in a system that has learned a large amount of information.

4.  Sparseness without reduction in information storage

Sparseness in memory representation has been shown by theoretical studies to have increased memory lifetimes ^4,5,6,7. Sparseness aids in retaining the specificities of the stored memories by preventing any interference between them ⁸. However, reduced amount of information storage by each memory is viewed as a drawback of sparseness ⁸.

Based on the semblance hypothesis, there is a possible alternate solution. While probing with the cue for retrieval of memory, depolarization spreads through the functional LINKs leading to the activation of postsynaptic membranes. Two types of activation were suggested by the semblance hypothesis. 1) A postsynaptic membrane change that is reminiscent of the arrival of the AP at the corresponding presynaptic membrane. 2) Depolarization spread to the postsynaptic membranes through the functional LINKs without activation of their corresponding presynaptic membranes. Both these changes are capable of producing synaptic semblance. The latter can contribute to the generation of action potential in the postsynaptic neuron provided there are sufficient spatial and temporal summations of similar EPSPs. The activated neurons can contribute to the network semblance. Both the changes at the postsynapses marked above as numbers 1) and 2) without leading to an AP (due to inadequate summation) that then leads to the activation of isolated postsynaptic membranes, resulting in only synaptic semblance (without activation of their corresponding presynaptic membranes, which is the basic definition of semblance). It is the cue characteristics that decide the identity of this specific set of postsynapses. This specific set of postsynapses (leading to synaptic semblance) and the specific combinations of all the activated postsynapses (that induce network activity and network semblance) provide an equivalent effect of sparseness as well as the complexity required for specific memory retrieval.

5.  Life-times of memories at the infero-temporal cortex are longer than at the medial temporal cortex

It was noted that in the cortical areas like the infero-temporal cortex memory lifetimes are presumably longer than in the medial temporal lobe and neural representations are less sparse. Infero-temporal cortex is crucial for visual object recognition and contains the last neuronal orders of the ventral cortical visual system. Based on the individual synaptic strength governing the memory storage it was imperative to visualize the longer memory lifetimes in the infero-temporal cortex as a paradox ⁸. Based on the semblance hypothesis, memory depends on the semblances resulting from the LINKed postsynapses. At the infero-temporal cortex, the effect of neurogenesis will be far smaller than that occurring in the medial temporal cortex. Therefore, in this region memories can have comparatively longer life-times.

The limiting factor as well as the specificity determining factor will be the specific cue characteristics. The functionally LINKed large islets of postsynapses will be continuously shared by many different cue stimuli and different learning events so that these LINKs at the infero-temporal cortex will remain for a long period of time. In addition to the isolated functional LINKs activating unique postsynapses by a specific learning and retrieval event at the neuronal orders outside the inferotemporal cortex, combinations of the set of postsynapses within the islets of LINKed postsynapses activated by a specific cue can contribute to the sparseness and specificity.

In summary, based on the semblance hypothesis, the specificity of memory retrieval is  based on two key events: 1) the activation of isolated postsynapses (both the postsynaptic membrane changes reminiscent of the arrival of an AP at the corresponding presynaptic terminal and the EPSP spread through the functional LINKs) leading to synaptic semblance, and 2) the combination of specific postsynapses activated in the islets of LINKed postsynapses. At an advanced level, it is also possible that the permutation of specific postsynapses activated in islets of LINKed postsynapses may be considered for computing memory requiring a temporal order of events. When computing the net semblance, factors like addition and removal of synaptic connections and the semblance effects from the neuronal orders above the level of granule layer of the hippocampus (neurogenesis) need to be taken into account.

8.  Larry F. Abbott (2008). Theoretical neuroscience rising. Neuron 60 (3):489-95 PubMed 

Computational neuroscientist Larry Abbott writes, “..….we more commonly tend to think of synapses as the locus of learning and memory, and neurons as the workhorses of dynamic computation. This may be radically wrong.” Professor Abbott further writes that in order to explain cognitive functions of our brain we need to build models that can provide provisions for the following three functions.

1.  Instant access to very large memory stores. 

2.  The ability to generate hypotheses.

3.  An interaction between internally generated hypotheses and external evidence that   allows sensory data to veto or support internal constructs extremely efficiently.

It appears that the semblance hypothesis has the potential to provide the basic framework to explain these three features. Please note that the following explanations will only hold true after mathematical proofs are derived and biological evidence for functional LINKs between the postsynapses are established.

1. Based on the semblance hypothesis, the net semblance resulting from the transient and permanent functional LINKs between the postsynapses can possibly provide the mechanism for large amounts of associative memory storage and retrieval (please refer to the book pdf for more explanation how this might occur). Based on the characteristic features of the cue, different combinations of the postsynaptic membranes have the potential to get activated through the functional LINKs. When semblances for the sensory inputs from each one of these postsynapses as well as the network semblances are computed, the identity of the sensory stimuli that are memorized can possibly be obtained. Therefore, any cue (internal or external) has the potential to have instant access to very large memory stores.

2. Based on the semblance hypothesis, a system can generate a hypothesis in the following instances. Let us assume that we are going to train an experienced brain, meaning that the brain has already been exposed to large number of diverse associative learning in its life. Based on the semblance hypothesis, this well-experienced brain can have many functional LINKs between the postsynaptic membranes at different neuronal orders. If the brain is exposed to a cue that was used in previous learning sessions, then the semblance of learned sensory input may easily be formed, leading to memory retrieval.

Let us now expose this brain to a completely novel unlearned item N instead of a cue which is part of an already learned item. Based on the semblance hypothesis, sensory stimuli from this novel item N will evoke activation of the postsynaptic membranes of many synapses, without the activation of their corresponding presynaptic membranes (which is the basis of the semblance hypothesis), leading to semblance for some sensory stimuli. These semblances are possible through the functional LINKs that were already established between the postsynaptic membranes during the prior learning events. Please note that many established functional LINKs (transient and structural) between the postsynaptic membranes will be shared during the retrieval process. The combination of the sets of postsynapses that are activated during the retrieval depends on the cue characteristics and will determine the features of the memorized item.

Let us imagine that the novel item N induces semblance for three items X, Y, Z above a certain threshold value resulting in their memory. Here, let X and Y represent items that were learned in the past. Let Z represent an unlearned item. Simultaneous semblances for X, Y and Z in the presence of N allow the system to detect interrelations between these four items (even though the system/individual will require detailed experiments to  explicitly confirm the relationship). In conditions where the brain/individual is appropriately trained to express those semblances that occur during exposure to a novel unlearned item, the system will be able to propose a hypothesis for an association between X, Y, Z and N.

It is likely that semblance for an unlearned item above a threshold occurs only in a nervous system that has undergone large number of prior learning events. In the previous example, semblance for Z in the presence of the novel item N provides the nervous system with memory (actually this not true memory; we can call it imagination) of a novel sensory input related to N. The nervous system may take note of it if the individual gains some advantage from it; otherwise, it gets ignored. The semblances of unlearned items when exposed to a learned cue/cues or the semblance of a learned item when exposed to an unlearned item/s may thus be responsible for possible theoretical inventions/predictions. It appears that the number of prior learning events, exposure to novel cues (that are not part of what was learned in the past) and attention to the semblances occurring from such exposures are key factors in hypothesis development.

Hypothesis building is part of our daily life. For example, we have preexisting functional LINKs that calculates the mathematical operation (9 - 8) =1. However, when exposed to a new operation of 1079 -1078, we get the answer 1. We may have never done this specific mathematical operation before in our life. However, we already have made enough hypotheses in our lives that whenever we minus 8 from 9 we are getting the answer 1 provided other digits are the same and have verified by real life situations to find that our results are correct. Therefore, even though (1079 - 1078) =1 is also a hypothesis, we actually don’t call it as a hypothesis. For the purpose of explanation of semblance hypothesis using functional LINKs it can be taken as a hypothesis. 

3. The internally generated hypothesis that X, Y, Z and N are interrelated will get challenged by the availability of sensory data to veto or support it. A typical example is how our sensory system perceives the apparent movement of the Sun around the Earth. We can clearly sense the movement of Sun during the sunrise and sunset in relation to the stable objects in our visual field. Based on the sensory inputs of the appearance of the Sun in the east and disappearance in the west, we may hypothesize that Sun is rotating round the Earth. However, scientific evidence shows us that it is the Earth that is moving around the Sun. A bit difficult fact to convince our own senses. From examining similar illusions in our daily lives (like the apparent movement of the trees while travelling in a car), we agree to hypothesize that it is the Earth that is revolving around the Sun until we test the scientific evidence in support of by ourselves.

9.  Jin-Hee Han, Steven A. Kushner, Adelaide P. Yiu, Hwa-Lin (Liz) Hsiang, Thorsten Buch, Ari Waisman, Bruno Bontempi, Rachael L. Neve, Paul W. Frankland, and Sheena A. Josselyn (2009). Selective erasure of a fear memory. Science 323: 1492-96 PubMed

It was previously shown that the lateral amygdala neurons with increased levels of the transcription factor cAMP/Ca²⁺ responsive element binding protein (CREB) are more likely to be activated by auditory fear training and testing than their neighbors⁹. It was further shown that selective ablation of this subset of neurons impaired subsequent expression of the fear-conditioning memory. Activation of a partial neuronal network during retrieval is proposed as the basis of network semblance (see figure 6, book pdf). It is possible that the neurons of the partial neuronal network activated during fear memory retrieval (contributing to network semblance) were lost by the selective ablation.

References:

1.   Malenka, R. C., and Nicoll, R. A (1999). Long-term potentiation - a decade of progress? Science 285: 1870-4

2.  Meeks, J. P., and Holy, T. E (2008). Pavlov's moth: olfactory learning and spike timing-dependent plasticity. Nat Neurosci  11: 1126-7

3.  Turner, G. C., Bazhenov, M., and Laurent, G (2008). Olfactory representations by Drosophila mushroom body neurons. J Neurophysiol  99: 734-46

4.  Amit, D.J., and Fusi, S (1994). Learning in neural networks with material synapses Neural Comput. 6: 957-82

5.  Willshaw, D. J., Buneman, O.P., and Longuet-Higgins, H.C. (1969). Non-holographic associative memory. Nature 222: 960-62

6.  Tsodyks, M., and Feigelman, M. (1998). Enhanced storage capacity in neural networks with low level of activity.  Europhys. Lett. 6:101

7.  Treves, A. (1990). Graded-response neurons and information encodings in autoassociative memories. Phys. Rev. A42: 2418-30

8.  Rubin, D. D. B. D., and Fusi, S. (2007). Long memory lifetimes require complex synapses and limited sparseness. Frontiers in Computational Neuroscience 1:1-14

9.  Han, J. H., Kushner, S. A., Yiu, A. P., Cole, C. J., Matynia, A., Brown, R. A., Neve, R. L., Guzowski, J. F., Silva, A. J., and Josselyn, S. A (2007). Neuronal competition and selection during memory formation. Science 316: 457-60