Contents:
The presentation of novel stimuli to rats alternated between two lists on successive days. Each list contained six odor stimulus pairs. The order of the stimulus pairs within a list is randomized during each exposure. All the odors in a list are different.
After ACh lesions are applied in entorhinal cortex, about 40 percent of cholinergic neurons survive. Learning of novel stimuli is still possible, but requires significantly more training.
Is that:
This simulation aims to model the memory aspects of the task, not the operant aspects involving association with reward or learning of the DNMS rule.
Figure 2:
Model learning of the stimulus patterns. Approximate the DNMS process, do not learn that explicitly for this simulation.
It is quite alright to use single spikes as the indicators of the specific odor, together with a connection table that creates the 8 neuron patterns for each.
Since this simulation does not involve learning rules in minicolumns, there is no need for spike-pair generating circuitry. A first version of the simulation can use a simple event generator to produce the necessary spikes. There is also no need to wait for many cycles, since the STM buffer does not train the synaptic memory at this time. Instead, three presentations are needed to train the synaptic memory.
Some questions:
Catacomb Bug?: For some reason, activation of a second (separate) pattern on the inputs causes reactivation of the first pattern that is encoded in the router patterns for the synapses of recurrent fibres in PFC of mcgaughy-20040625.ccm. So, I have to temporarily disable that second pattern input at t=760 ms and t=2260 ms for the demonstration of activity in PFC for novel versus familiar item patterns.
Aim (20040801): To achieve reliable activity in PFC for the "familiar" items condition. This must correct the bug in the 20040625 version of the simulation and sustain activity for the initial stimulus of each trail during the trial. This is commenced in mcgaughy-20040801.ccm.gz.
After closer inspection, it seems that the problem is due to confusion about routing through the SpikeProjection link (tested with a routing table and with 1-1 connections). It is likely that spikes intended for different neurons are being sent to different synapses on the same neuron. For instance, with 1-1 connections it would be assumed that each neuron in the first pattern would send recurrent input only to itself and therefore all neurons in that pattern would receive subthreshold input only. Yet the result is that neuron number 0 spikes. Apparently, the recurrent spikes of any of the neurons are sent to neuron 0 in that case. This happens if the output spikes of each neuron are numbered 0. That way, 1-1 connections through a SpikeProjection that recursively distributes spikes to all neurons in PFC send all the spikes with ID 0 to neuron 0. In the case of a routing table that sets up patterns of 8 neurons each, the output spike of all neurons have ID 0 and are then copied to target IDs 0-7. Consequently, neurons 0-7 spike - the first pattern. In previous experiments, learning rules were used to set up connection strengths on all-all recurrent and pathway connections. In the present setup, where populations are created by setting row and column numbers for a captured cell, without automatically adding relays for transmission in and out of a capture box, this is the expected behavior. Nevertheless, it deserves careful documentation and possibly even a warning that appears during model design.
Figure 4: Familiar and novel stimuli maintained in entorhinal and prefrontal cortices of the model. The top two response functions show the membrane potentials of two sampled neurons in prefrontal cortex (PFC) that spike in patterns representing a novel (above) and a familiar (below) stimulus. The bottom two response functions show the membrane potentials of two sampled neurons in entorhinal cortex (EC) that spike in patterns representing a novel (below) and a familiar (above) stimulus. Trials are separated by periods in which no spikes occur. In the first trial, a familiar stimulus is maintained by recurrent reactivation through learned connection strengths in PFC and maintained in EC by intrinsic processes that enable persistent firing. In the second trial, a novel stimulus cannot be maintained in PFC, but is correctly maintained in EC. In the third trial, the familiar stimulus is again maintained in PFC, while EC is lesioned so that the corresponding pattern of spiking does not appear there. Finally, the fourth trial shows the inability to maintain a memory of the novel stimulus when EC is lesioned.
The data for vector graphics presentations of the simulation results are obtained in .cdf output of the "cdfoutput" vector recorder in mcgaughy-20040802.ccm.gz. Figures with vector graphics are produced for the corresponding data with showstacked.sh and showspikesstacked.sh (see mcgaughytask.pdf).
Aspects of the experimental results that are not yet shown in the simulation results above:
Since the task does not require that sequences are learned, dependence on medial temporal function is not absolute. That is apparent in the fact that deficiencies for novel stimuli when EC is lesioned or cholinergic input to EC is suppressed dissipate as a stimulus becomes familiar and maintenance is achieved in PFC.
Aim (20040801): Add a transition of a novel stimulus to familiar stimulus maintenance in PFC. Since Catacomb presently allows us only to either preset synaptic strengths or to train them, not to preset some and train others in the same synaptic population, the most straightforward way to add learning for the novel stimulus to the simulation without breaking the behavior for the familiar stimulus is to add a separate synaptic population. That population of synapses can receive recurrent input from the novel stimulus pattern and experience STDP. The recurrent connections that establish these strengthened synapses must have a synaptic delay greater than zero, but less than about 20 ms.
The delay on synapses in PFC that establish stimulus memory maintenance through recurrent excitation is not as great as the interval between repetitions in EC. Previously, a long delay of 125 ms was used for the familiar stimulus in PFC to achieve similar spike rates for maintenance in PFC and EC. There was no well-formulated presupposition to support the long delay, although circuits that achieve the delay may be devised (e.g. many synaptic relays, recurrent excitation plus rhythmic modulation, etc.). At present, it seems wise to assume the greater spike rate in PFC.
Note that a greater spike rate may appear in EC as well if LTP established in ECIII by repetition in ECII STM is explicitly modeled there.
The transition is implemented in mcgaughy-20040803.ccm.gz.
Figure 5: Familiar and novel stimuli maintained in entorhinal and prefrontal cortices of the model. The top two response functions show the membrane potentials of two sampled neurons in prefrontal cortex (PFC) that spike in patterns representing a novel (above) and a familiar (below) stimulus. The bottom two response functions show the membrane potentials of two sampled neurons in entorhinal cortex (EC) that spike in patterns representing a novel (below) and a familiar (above) stimulus. Trials are separated by periods in which no spikes occur. In the first trial, a familiar stimulus is maintained by recurrent reactivation through learned connection strengths in PFC and maintained in EC by intrinsic processes that enable persistent firing. In the second trial, a novel stimulus cannot be maintained in PFC, but is correctly maintained in EC. In the third trial, the familiar stimulus is again maintained in PFC, while EC is lesioned so that the corresponding pattern of spiking does not appear there. Finally, the fourth trial shows the inability to maintain a memory of the novel stimulus when EC is lesioned. After many additional presentations of the novel stimulus, PFC learns to maintain a representation of that stimulus.
Figure 6: The Catacomb circuitry of the integrate and fire neuron simulation of memory in the delayed non-match to sample task. Pyramidal neurons are simulated in two regions. A prefrontal cortex region encodes associative patterns of 8 spiking neurons for each odor stimulus received by strengthening recurrent synapses. An entorhinal region maintains a pattern of spikes that represent an odor stimulus by persistent firing that is based on intrinsic properties of ECII neurons that experience after-depolarization following a spike. Interneurons provide recurrent inhibitory input that maintains a separation between the repeated firing of successive spike patterns in the buffer and that enable first-in-first-out replacement of spike patterns maintained in the buffer. Our model proposes that novel stimuli require buffering in EC, while known stimuli are encoded in PFC.
Once the theta input was removed, spiking was repetitive with the frequency that pure ADP would achieve. I now change the conductance of the ADP response from 25 nS to 12 nS. There is no longer any repetition (mcgaughy-same-item-dropped-20041014.ccm). The ADP response shows to oddities: (1) It makes a small sudden acceleration in its rise, possibly due to the end of the duration of AHP. (2) It drops off rapidly right thereafter, which is clearly due to the end of the duration of the ADP computation. To make a strong argument, this sudden drop-off must probably be removed.
An important consideration to keep in mind is that the problematic behavior is only observed if the first test odor is same as the sample odor. If another test odor is presented first, then the problem does not occur. Of course, that could simply be a matter of having the memory available during the non-match operation. If the memory is lost before the other odor is presented then the non-match operation cannot be done, which would explain a result in which the rat continues to run back and forth between the test odors.
Given the considerations in TL#200410150923.2, there are two aspects of the model protocol in which the breakdown can occur:
It would be fairly simple to add circuitry that erases the buffer when read, but the circuitry would appear to be created just for that purpose and the assumption that decisions are made immediately, based on the first test odor, may not be valid.
Focusing on the second possibility, we describe the significant difference between the state of the buffer when the first test odor is different than the buffered odor versus the same as the buffered odor:
Even with AHP and ADP responses that reset at each spike, the state at the time of input is not the same in the two cases, since membrane potential does not instantaneously return to the value it would have without AHP and ADP. One characteristic parameter that may therefore make the difference between observing and not observing the effect is the capacitance of the neuron. [MECHANISM 1]
If the membrane potential of the neuron is elevated compared to resting potential (i.e. if the characteristic ADP and AHP responses are such that ADP prevails at the time of input) and if the capacitance is sufficiently large it could affect the phase at which the input causes spiking. If the input spikes earlier then the ADP may not enable reactivation during the theta cycle and the memory could be lost.
In the preceding paragraph, the focus shifted from the capacitance (which is only importance once the preceding AHP and ADP end or are reset) to the way input is delivered. If the input capacitance is not as strong (even if preceded by a well-timed filter), but instead depends on a more gradually rising input response then the modulated membrane potential can affect the phase of input spiking. That phase is significant if reactivation of input is just barely possible during a theta cycle. This is clearly another set of characteristic parameters that I probably changed in the course of the development of the STM buffer. [MECHANISM 2].
I will first explore the second mechanism, since the way in which memory loss occurs then is very clear. In order to make the input phase highly significant, I will first manipulate the input timing so that I determine the earliest phase at which reactivation is achieved. Similarly, I can look for the latest phase at which reactivation is achieved. At phases beyond the successful input interval, the combination of rising ADP and theta do not achieve reactivation within the theta cycle. I will then determine how to modify the input timing and input response function so that the resulting input spike phase is sufficiently affected by the membrane potential:
If the effect of ADP and AHP on the phase of input causes the input to occur earlier, it is difficult to cause memory loss that way, since cyclical reactivation occurs at 125 ms intervals. It is possible, since a combination of ADP and theta is needed and the rapidity of the rise of each plays a roll in the interval within which an input can be reactivated in the same theta cycle. Yet, it may be easier to delay the input timing so that the opportunity for the first reactivation is missed. That is possible if AHP has the upper hand during the input phase for a neuron that is already engaged in persistent spiking. So, characteristic parameters would need to be adjusted so that AHP dominates at that phase and suppresses the membrane potential. Instead of changing the time constants it may be enough to lower the AHP reversal potential.
I should now (a) slow down the input response so that spiking is delayed several milliseconds compared ot the onset of input, (b) set the input timing near the maximum phase at which reactivation is achieved, and (c) give AHP the upper hand at that phase when the neuron is already engaged in persistent spiking.
I've modified the input synapse parameters to conductance 6 nS, rise time constant 5 ms and fall time constant 10 ms. As anticipated, the synaptic response takes 5 ms to reach its peak and so the maximum phase at which that input can cause a spike is before t=145 ms (and t=770 ms). At these phases, the membrane potential at a neuron already engaged in persistent spiking is about -54 mV, that at a neuron not engaged in persistent spiking is about -72 mV and rising.
The desired result would be obtained if the membrane potential at a neuron already engaged in persistent spiking was lower than -72 mV. Two approaches may lead to the required mechanism:
Attempting to find parameters for the AHP and ADP only that lead to dominance of AHP in a large part of the theta cycle did not work, since both the AHP and ADP were already designed with marginal parameter values. Instead, I now focus on the second option in which the theta modulating input response can also be modified.
A successful implementation with same item drop-out is presented in mcgaughy-same-item-dropped-20041018b.ccm.gz. The significant set of parameter values is given with comments in table 1 below:
| parameter | value | comment |
|---|---|---|
| capacitance time constant | 5 ms | A fast time constant allows for steeper ascent following AHP. |
| threshold potential | -50 mV | |
| resting potential | -50 mV | The resting and reset potentials, together with the theta inhibitory input produce a membrane potential modulation between -77.5 mV and -52 mV, so that theta modulation is the deciding factor for the ability to elicit a reactivation spike. |
| reset potential | -50 mV | (Used in combination with the resting potential and theta inhibitory input as described above.) |
| input reversal potential | 0 mV | |
| input conductance | 30 nS | The input conductance is strong enough so that input causes an early spike, while the synaptic response is still rising, if the neuron is not experiencing AHP. If the neuron is experiencing AHP then the synaptic response needs to rise further before an input spike occurs, thereby delaying the input spike. This is the essential difference between input to a neuron that is already engaged in persistent firing and one that is not. |
| input reponse rise time constant | 20 ms | This is used to obtain the necesssary duration of the rising input response for the functional feature described above. |
| input response fall time constant | 20 ms | |
| theta response reversal potential | -90 mV | Theta modulation of membrane potential is caused by GABAergic input that is received from the septum at the frequency of theta rhythm (8Hz). |
| theta response conductance | 30 nS | Strong theta modulation of membrane potential enables dominant phases of input and reactivation. If input is delayed, a sufficient ADP is not achieved before the end phase of the reactivation mode of the theta modulation. A reactivation spike is not possible in the input phase that follows, due to the strong hyperpolarization by the theta modulation. A second input spike on the same neuron therefore does not achieve persistent firing and the memory is lost. |
| theta response rise time constant | 1 ms | The two time constants determine the shape of the theta modulation that causes an asymmetric distribution of neuronal spikes, as seen in the data. |
| theta response fall time constant | 30 ms | |
| AHP reversal potential | -120 mV | A strong AHP insures that the membrane potential is sufficiently inhibited when the same neuron engaged in persistent firing receives another input. This causes the delay of the input spike at a neuron that is already engaged in persistent firing. |
| AHP conductance | 50 nS | |
| AHP rise time constant | 0.0001 ms | The AHP is an exponentially decauing function. |
| AHP fall time constant | 60 ms | |
| ADP reversal potential | -40 mV | The ADP can elicit spike repetition when theta modulation causes little hyperpolarization of the membrane potential. |
| ADP conductance | 43 nS | |
| ADP rise time constant | 125 ms | The ADP response has a rise time that is about the same as the duration of a theta cycle so that reactivation in STM may be achieved once in each theta cycle. The combination of the theta modulation and an ADP that depends on the theta modulation to achieve spiking enables regular persistent firing (as first proposed by Lisman and Idiart (Science, 1995). |
| ADP fall time constant | 125 ms |
Figure 7: The membrane potential of the same two neurons is shown in simulations of short-term memory (STM) for two trials (a) and (b). When characteristic parameters of the persistent firing buffer are such (see table 1) that the synaptic response of input may be delayed by the after-hyperpolarization (AHP) of a preceding reactivation spike at the same neuron then the memory item that is maintained by persistent firing may be lost. In (a), input due to the first test odor targets a different representative neuron in the buffer. Spiking for the sample odor previously encountered is maintained. In (b), the first test odor is the same as the sample odor so that the corresponding input targets the same neuron that is maintaining the STM of the sample odor. The input is delayed by the AHP of that repeated spiking. Due to the delay, the after-depolarization (ADP) of the neuron does not raise the membrane potential to the threshold by the end of the theta cycle. Once theta modulation enters the input mode, hyperpolarizing the membrane potential to suppress reactivation spikes, reactivation is not possible. Without achieving a spike at its peak, the ADP declines and the neuron is no longer engaged in persistent firing. The memory of the sample odor is lost.