Cholinergic Modulation May Enhance Cortical 
Associative Memory Function 
Michael E. Hasselmo* 
Computation and 
Neural Systems 
Caltech 216-76 
Pasadena, CA 91125 
Brooke P. Andersont 
Computation and 
Neural Systems 
Caltech 139-74 
Pasadena, CA 91125 
James M. Bower 
Computation and 
Neural Systems 
Caltech 216-76 
Pasadena, CA 91125 
Abstract 
Combining neuropharmacological experiments with computational model- 
ing, we have shown that cholinergic modulation may enhance associative 
memory function in pitiform (olfactory) cortex. We have shown that the 
acetylcholine analogue carbachol selectively suppresses synaptic transmis- 
sion between cells within pitiform cortex, while leaving input connections 
unaffected. When tested in a computational model of piriform cortex, 
this selective suppression, applied during learning, enhances associative 
memory performance. 
I INTRODUCTION 
A wide range of behavioral studies support a role for the neurotransmitter acetyl- 
choline in memory function (Kopelman, 1986; Hagan and Morris, 1989). However, 
the role of acetylcholine in memory function has not been linked to the specific 
neuropharmacological effects of this transmitter within cerebral cortical networks. 
For several years, we have explored cerebral cortical associative memory function 
using the pitiform cortex as a model system (Wilson and Bower, 1988, Bower, 1990; 
Hasselmo et al., 1991). The anatomical structure of pitiform cortex (represented 
schematically in figure 1) shows the essential features of more abstract associative 
*e-mail: hasselmosmaug.cns.caltech.edu 
t e-mall: brookehope.caltech.edu 
46 
Cholinergic Modulation May Enhance Cortical Associative Memory Function 47 
matrix memory models (Haberly and Bower, 1989) 1 Afferent fibers in layer la 
provide widely distributed input, while intrinsic fibers in layer lb provide extensive 
excitatory connections between cells within the cortex. Computational models of 
pitiform cortex demonstrate a theoretical capacity for associative memory function 
(Wilson and Bower, 1988; Bower, 1990; Hasselmo et al., 1991). Recently, we have 
investigated differences in the physiological properties of the afferent and intrinsic 
fiber systems, using modeling to test how these differences affect memory function. 
In the experiments described below, we found a selective cholinergic suppression of 
intrinsic fiber synaptic transmission. When tested in a simplified model of pitiform 
cortex, this modulation enhances associative memory performance. 
Afferent fiber 
synapses (layer la) 
Afferent input = A i 
Intrinsic fiber 
synapses (layer lb) = Bid 
- Neuron activation = ai(t ) 
Lateral inhibition 
(via intemeurons) = Hid 
Neuron output = g(ai(t )) 
Figure 1: Schematic representation of piriform cortex, showing afferent input Ai 
(layer la) and intrinsic connections Bij (layer lb) 
2 EXPERIMENTS 
To study differences in the effect of acetylcholine on afferent and intrinsic fiber 
systems, we applied the pharmacological agent carbachol (a chemical analogue 
of acetylcholine) to a brain slice preparation of pitiform cortex while monitoring 
changes in the strength of synaptic transmission associated with each fiber system. 
In these experiments, both extracellular and intracellular recordings demonstrated 
clear differences in the effects of carbachol on synaptic transmission (Hasselmo and 
Bower, 1991). The results in figure 2 show that synaptic potentials evoked by 
activating intrinsic fibers in layer lb were strongly suppressed in the presence of 
 For descriptions of standard associative memory models, see for example (Anderson 
et al., 1977; Kohonen et al., 1977). 
48 Hasselmo, Anderson, and Bower 
100pM carbachol, while at the same concentration, synaptic potentials evoked by 
stimulation of afferent fibers in layer la showed almost no change. 
Control 1 x 10 '4 M Carbachol Washout 
LAYER 1A 
1 2 3 
LAYER lB 
] 2 3 
0.5mV 
-- O,5mV 
20ms 
Figure 2: Synaptic potentials recorded in layer la and layer lb before, during, and 
after perfusion with the acetylcholine analogue carbachol. Carbachol selectively 
suppresses layer lb (intrinsic fiber) synaptic transmission. 
These experiments demonstrate that there is a substantial difference in the neuro- 
chemical modulation of synapses associated with the afferent and intrinsic fiber sys- 
tems within pitiform cortex. Cholinergic agents selectively suppress intrinsic fiber 
synaptic transmission without affecting afferent fiber synaptic transmission. While 
interesting in purely pharmacological terms, these differential effects are even more 
intriguing when considered in the context of our computational models of memory 
function in this region. 
3 MODELING 
To investigate the effects of cholinergic suppression of intrinsic fiber synaptic trans- 
mission on associative memory function, we developed a simplified model of the 
pitiform cortex. This simplified model is shown schematically in figure 1. At each 
time step, a neuron was picked at random, and its activation was updated as 
N 
ai(t -t- 1) = Ai(t) + [(1 - c)Bij - Hij]g(ai(t)). 
j=l 
where N = the number of neurons; t = time  {0, 1,2,...}; c = a parameter 
representing the amount of acetylcholine present. c  [0, 1]; ai = the activation or 
membrane potential of neuron i; g(al) = the output or firing frequency of neuron 
i given ai; Ai = the input to neuron i, representing the afferent input from the 
olfactory bulb; Bid = the weight matrix or the synaptic strength from neuron j 
to neuron i; and Hij = the inhibition matrix or the amount that neuron j inhibits 
neuron i. To account for the local nature of inhibition in the pitiform cortex, Hij = 0 
Cholinergic Modulation May Enhance Cortical Associative Memory Function 49 
for li - Jl > r and Hq = h for li - Jl < r, where r is the inhibition radius. The 
function g(ai) was set to 0 if a < 0a, where 0a = a firing threshold; otherwise, it 
was set to 7otanh(al - 0), where 7 = a firing gain. 
The weights were updated every N time steps according to the following hebbian 
learning rule. 
aj = f(W O) 
aW = W(t + iV)- Wo(t) = (- c)'it(a 
The function f(.) is a saturating function, similar to g(.), used so that the weights 
could not become negative or grow arbitrarily large (representing a restriction on 
how effective synapses could become). 'It is a parameter that adjusts learning 
speed, and 0t is a learning threshold. The weights were updated every N time steps 
to account for the different time scales between synapse modification and neuron 
settling. 
3.1 TRAINING OF THE MODEL 
During learning, the model was presented with various vectors (taken to represent 
odors) at the input (Ai(t)). The network was then allowed to run and the weights 
to adapt. 
The procedure for creating the set ofvectors {A m I rn  {1, ... ,M} } was: set A? -- 
max{O, G(y, a)}, where G = gaussian with average/ and standard deviation a, and 
m 2 2 2 
normalize the whole vector so that liA [I = + ). M = number of memories 
or odors presented to network during training, and A? - the input to neuron i while 
odor rn is present. 
During learning, in the asynchronous update equation, Ai(t) = A for r time steps, 
then Ai(t) = A for the next r time steps, and so on; i.e., the various odors were 
presented cyclically. 
3.2 PERFORMANCE MEASURE FOR THE MODEL 
The pitiform cortex gets inputs from the olfactory bulb and sends outputs to other 
areas of the brain. Assuming that during recall the network receives noisy versions 
of the learned input patterns (or odors), we presume the pitiform cortex performs 
a useful service if it reduces the chance of error in deciding which odor is present 
at the input. One way to quantify this is by using the minimum probability of 
classification error, P (from the field of pattern recognition 
For the case of 2 odors corrupted by gaussian noise, P = the area underneath 
the intersection of the gaussians. For spherically symmetric gaussians with mean 
vectors/q and/a and identical standard deviations , 
where d = II/ - pall. Thus, the important parameter is the amount of overlap as 
quantified by d/o' .., the larger the d/o', the lower the overlap and Pe. 
2See, for example, (Duda and Hart, 1973). 
50 Hasselmo, Anderson, and Bower 
For more than 2 odors and for non-gaussian noise or non-spherically-symmetric 
gaussian noise, the equation for P becomes less tractable. But keeping with the 
above calculations, an analogue of d/o' was developed as follows.  was set = 
(11x- i112), and then f/was defined as 
IIi - II 
where i, j 6 {1,..., M). Here, f/is the analogue of d/o' in the previous paragraph 
and is similar to an average over all odor pairs of d/o'. 
For the model, if f/is larger for the output vectors than for the input vectors, there 
is less overlap in the outputs, classification of the outputs is easier than classification 
of the inputs, and the model is serving a useful purpose. Thus, we use p 
as the performance measure. 
3.3 TESTING THE MODEL 
The model was designed to show whether the presence of acetylcholine has any 
influence on learning performance. To that end, the model was allowed to learn 
for a time with various levels of acetylcholine present, and then acetylcholine was 
turned off and the model was tested. 
For testing, weight adaptation was turned off, acetylcholine influence was turned 
off (c = 0), noisy versions of the various odors presented during learning were 
presented at the input, and the network was allowed to settle. From these noisy 
input/output pairs, 's could be estimated, f/in and fto, t could be calculated, and 
finally p could be calculated. Then, the state of the network could either be reset 
(for a new learning run) or be set to what it was before the test (so that learning 
could continue as if uninterrupted). 
3.4 RESULTS OF TESTING 
A typical example of a test run is shown in figure 3. There, c was varied from 0 (no 
acetylcholine) to .9 (a large concentration), and the various other parameters were: 
N=10, M= 10, r=2, h=.3,7a=l, 0a= 1,7t = 10-3,0t=l, andr=10. In 
the figure, large dark rectangles represent larger values of p. Small or non-existant 
rectangles represent values of p < 1. 
Notice that, for a fixed amount of acetylcholine, the model's performance rises and 
then falls over time. Ideally, the performance should rise and then flatten out, as 
further learning should not degrade performance. The weight adaptation equation 
used in the model was not optimized to preclude overlearning (where all of the 
weights being reinforced have saturated to the largest allowed value). In principle, 
the function f(.) could be used for this, perhaps in conjunction with a weight decay 
term. This was not of great concern since the peak performance is what indicates 
whether or not acetylcholine has a useful effect. Also, the more acetylcholine present, 
the longer the learning took. This is reasonable as, before saturation, ZSW o (1-c). 
Figure 4 shows maximum average performance for various values of acetylcholine. 
Averages were calculated by doing many tests like the one above. This is useful as 
Cholinergic Modulation May Enhance Cortical Associative Memory Function 51 
the odor inputs and the individual tests are stochastic in nature. Obviously, the 
larger values of acetylcholine enhance performance. 
o 
 w,-,,d 
o 
0.9 
................................................................... I .... ,-- I,-.,ll-I-ll, Ill,lllll,I II IIl,l,111ql 
............................. , ........ , ,,,,u,,l',.I,id 
.............. ,..,(flllllldllllildlddlllllll,lllli,.llf i11.11 ,--.[ .............. I-,--fd'-,f 
............... ',"klf'lllllgllllllglll(]iliilll# I"'l,-'-f ...... , I,l-,ll-,l[-I-flll.11'l[-I I. 4l I I,.-, IllfllPilllltlllilll 
............ [,,mla441igllllg[!l,-f ...... I-,[111l1414,r.l-11,14fk,114..I .11,,,, .................................... I .. 
.......... ,,[,i, P'llg,l,illllidl--, "-'1, .fl. dl'11il f I-, f ............................................................................ 
......... Illlllf411,11,[11,,, ...... Illif ........................................................................................................... 
....... ,m,lldll.11,I.-Il--, ................................................................................................................ 
...... ,,fallalii ll-dl ........................................................................................... , ............................. 
..... Iffllllll ............................................................................................................................... I ...... 
 -- II,ll .............................................................................................. 
..... I"fld .......................................................................................................................................... 
.... ,l'f" .............................................................................................................................. I .......... 
.... ,I ..................................................................... I ................................................ I ...................... 
 --I. ...................................................................................................... I ...................................... 
0 Time (t) 4.5x 10 5 
Figure 3: Sample test run, with time on the horizontal axis and acetylcholine level 
on the vertical axis. Larger black rectangles indicate better performance. 
1.7 
1.6 
1.5 
1.4 
1.3 
1.2 
1.1 
0.0 
0.2 0.4 0.6 0.8 
Acetylcholine concentration (c) 
Figure 4: Maximum average performance vs. acetylcholine level. Acetylcholine 
increases the performance level attained. 
52 Hasselmo, Anderson, and Bower 
4 CONCLUSION 
The results from the model show that suppression of connections between cells 
within the pitiform cortex during learning enhances the performance during recall. 
Thus, acetylcholine released in the cortex during learning may enhance associative 
memory function. These results may explain some of the behavioral evidence for 
the role of acetylcholine in memory function and predict that acetylcholine may 
be released in cortical structures preferentially during learning. Further biological 
experiments are necessary to confirm this prediction. 
Acknowledgement s 
This work was supported by ONR contracts N00014-88-K-0513 and N00014-87-K- 
0377 and NIH postdoctoral training grant NS07251. 
References 
J.A. Anderson, J.W. Silverstein, S.A. Ritz and R.S. Jones (1977) Distinctive fea- 
tures, categorical perception, and probability learning: Some applications of a neu- 
ral model. Psychol. Rev. 84: 413-451. 
J.M. Bower (1990) Reverse engineering the nervous system: An anatomical, physi- 
ological and computer based approach. In S. Zornetzer, J. Davis and C. Lau (eds.), 
An Introduction to Neural and Electronic Networks. San Diego: Academic Press. 
R. Duda and P. Hart (1973), Pattern Classification and Scene Analysis, New York: 
Wiley. 
L.B. Haberly and J.M. Bower (1989) Olfactory cortex: Model circuit for study of 
associative memory? Trends Neurosci. 12: 258-264. 
J.J. Hagan and R.G.M. Morris (1989) The cholinergic hypothesis of memory: A 
review of animal experiments. In L.L. Iversen, S.D. Iversen and S.H. Snyder (eds.) 
Handbook of Psychopharmacology Vol. 20 New York: Plenum Press. 
M.E. Hasselmo, M.A. Wilson, B.P. Anderson and J.M. Bower (1991) Associative 
memory function in pitiform (olfactory) cortex: Computational modeling and neu- 
ropharmacology. In: Cold Spring Harbor Symposium on Quantitative Biology: The 
Brain. Cold Spring Harbor: Cold Spring Harbor Laboratory. 
M.E. Hasselmo and J.M. Bower (1991) Cholinergic suppression specific to intrinsic 
not afferent fiber synapses in pitiform (olfactory) cortex. J. Neurophysiol. in press. 
T. Kohonen, P. Lehtio, J. Rovamo, J. Hyvarinen, K. Bry and L. Vainio (1977) A 
principle of neural associative memory. Neurosci. 2:1065-1076. 
M.D. Kopelman (1986) The cholinergic neurotransmitter system in human memory 
and dementia: A review. Quart. J. Exp. Psychol. 38A:535-573. 
M.A. Wilson and J.M. Bower (1988) A computer simulation of olfactory cortex 
with functional implications for storage and retrieval of olfactory information. In 
D. Anderson (ed.) Neural Information Processing Systems. AIP Press: New York. 
Part II 
Neuro-Dynamics 
