Synaptic Noise Improves Detection of Subthreshold Signals in Hippocampal CA1 Neurons

William C. Stacey, Dominique M. Durand


Stochastic resonance (SR) is a phenomenon whereby the detection of a low-level signal is enhanced in a nonlinear system by the introduction of noise. Studies of the effects of SR in neurons have suggested that noise could play a prominent role in improving detection of small signals. Most experimental SR research has focused on the role of noise in sensory neurons using physiological stimuli. Computer simulations show that signal detection in hippocampal neurons is improved by the addition of physiological levels of noise applied extracellularly to synaptic inputs. These results were confirmed experimentally. We now report that endogenous noise sources can also improve signal detection. The noise source was generated by modulating the random synaptic activity on the apical dendrites of CA1 cells in rat hippocampal slices using subthreshold cathodic current. Intracellular recordings of CA1 cells showed that even small increases of synaptic noise are able to greatly improve the detection of an independent, synaptic, subthreshold stimulus as predicted by the simulations. The noise variance in the CA1 cell was compared with the resting variance and with variance changes caused by several endogenous noise sources. In all cases, the increased noise variance was well within the physiological range. These results were supplemented and analyzed with a CA1 computer model. The improved signal detection with small amounts of endogenous noise suggests that the diverse inputs to CA1 are able to improve detection of subthreshold synaptic signals and could provide a means to modulate detection of specific inputs in the hippocampus.


Noise is often seen as detrimental to signal detection. However, in certain systems, noise can enhance detection of some signals. The effect is called stochastic resonance (SR) and predicts that noise can enhance the detection of subthreshold signals in nonlinear, threshold-detecting systems (Benzi et al. 1981; Fauve and Heslot 1983; McNamara et al. 1988; Wiesenfeld and Moss 1995). The theory has been applied to several neural systems (Braun et al. 1994; Bulsara et al. 1991; Collins et al. 1996; Douglass et al. 1993; Gluckman et al. 1996; Levin and Miller 1996; Pei et al. 1996a,b; Stacey and Durand 2000) and has been shown to have significant effects in vivo (Russell et al. 1999). The implications of SR are several, including the paradigm shift that noise present in the neurons may actually serve to enhance signal detection.

SR, as shown in Eq. 1, describes a proportional relationship between the signal-to-noise ratio (SNR) of the output to the noise intensity (D, noise intensity; ε, signal strength; ΔU, threshold barrier height)SNRεΔUD2e(ΔU/D) Equation 1 Equation 1 gives the SNR for a periodic input to a monostable system such as a neuron (Stocks et al. 1993;Wiesenfeld et al. 1994) and is derived from the original SR equation dealing with bistable phenomena (Dykman et al. 1995; Moss et al. 1994; Wiesenfeld and Moss 1995). As noise intensity is increased, the curve produces a sharp rise from zero to a peak SNR followed by a gradual return to one (an example of the curve is shown in Fig. 5 B).

The importance of studying SR in the CNS arises from a consideration of the nature of signal detection in a typical pyramidal cell such as CA1. Possessing extensive dendritic trees, CA1 cells receive tens of thousands of synaptic inputs and require up to 300 simultaneous synaptic events to produce an action potential (Andersen 1990). The effects of attenuation also greatly affect the detection dynamics (Spruston et al. 1993). There is therefore a high probability of receiving subthreshold signals. Because random synaptic events have the same quantal size as single evoked excitatory postsynaptic potentials (EPSPs) (Larkman et al. 1997), the effect of noise on signal detection is significant. CA1 cells thus function in an environment well suited for SR.

The physiological relevance of SR was demonstrated most clearly in sensory systems, where both signal and noise sources are readily apparent (Braun et al. 1994; Douglass et al. 1993; Levin and Miller 1996). In the crayfish mechanoreceptor, for instance, it is logical to describe the noisy ripples in a stream as noise and the approach of a predator as a signal. Central neurons also exhibit SR activity (Gluckman et al. 1996; Stacey and Durand 2000); however, because the inputs are more complicated, it is difficult to design an experiment that directly can test SR using the endogenous sources of signal and noise.

Recently it has been shown that hippocampal CA1 cells are capable of producing SR behavior (Stacey and Durand 2000). A computer model of a CA1 cell was used to determine the role of physiological noise sources in signal detection. Experimental verification of this model utilized applied current pulses to create noise events. The pulses created random, global EPSPs. While the experiment did demonstrate that detection of subthreshold signals could be improved by increasing the noise, the applied noise was only an approximate representation of physiological noise sources. Noise in the CNS involves multiple, independent sources and locally uncorrelated noise events (Lindner et al. 1995). The goal of the present paper is to test the hypothesis that physiological levels of endogenous noise within the hippocampal slice can improve signal detection. The noise variance needed to evoke SR was measured and compared with baseline physiological noise and the augmentation in variance produced by several different noise sources. The improvement in signal detection was then evaluated at the different noise levels.


Preparation of slices

Sprague-Dawley rats (20–30 days old) were anesthetized with ethyl ether and decapitated according to protocol approved by the University Animal Resource Center. The brains were quickly removed and placed in aerated (95% O2-5% CO2), iced sucrose—artificial cerebrospinal fluid (ACSF, in mM: 220 sucrose, 3 KCl, 1.25 NaH2PO4, 2 CaCl2, 2 MgSO4, 26 NaHCO3, and 10 dextrose). Hippocampal slices 400 μm thick were cut in this solution using a Campden 752 M Vibroslice. Slices were then transferred to a submersion chamber and bathed for over 1 h in room temperature ACSF (in mM: 124 NaCl, 3.75 KCl, 1.25 KH2PO4, 2 CaCl2, 2 MgSO4, 26 NaHCO3, and 10 dextrose) and aerated with 95% O2-5% CO2. Each slice was then transferred to a perfusion chamber and incubated to 35°C for implementation.

Electrode placement and signal generation

A tungsten microelectrode was used to generate periodic signal events in the Schaffer collateral layer (see Fig.1 A). The tip of another tungsten electrode was bent 90° and positioned parallel to the slice across the entire CA3 layer. A sharp (60–120 MΩ) glass microelectrode filled with 2 M potassium acetate was used for intracellular recording of CA1 cells. The electrode was mounted on a Burleigh piezoelectric motor (Inchworm) with a 6000 controller. Data were recorded using a Warner IE-210 amplifier and recorded onto digital audio tapes.

Fig. 1.

Methods. A: one electrode (“signal”) was placed in the stratum radiatum, and another (“noise”) was bent parallel to the slice and lain across the top of the CA3 region (bar: duration of current pulse). An intracellular recording electrode was then placed in a CA1 neuron. B: data analysis. Five copies of intracellular data were spliced to allow for power spectral measurement. A detected output time series was created showing onset of action potentials. The power spectrum of this output series was computed. C: the signal-to-noise ratio (SNR) was measured by dividing the power at the input signal frequency (shown by the head of the arrow) by the average baseband power on either side of the signal (the tail of the arrow). D: schematic showing layout of CA1 model and position of synapses. Data for cylinders are in Table 1.

The “signal” consisted of a periodic train of 250-μs current pulses applied at 1 Hz. By injecting the signal pulses into the Schaffer collaterals, the signal reached CA1 cells as synaptic events on the apical dendrites. The amplitude of the signal was adjusted to produce a subthreshold EPSP (3–8 mV, threshold ≅ 10 mV) in the CA1 cell being recorded.

The “noise” was induced by injecting a long (10–15 s), low-amplitude (20–100 μA) cathodic current pulse on the electrode near the CA3 region. This pulse raised most of the neurons in the CA3 region closer to threshold and induced increased random synaptic activity in the CA1 layer. However, it also produced a large (∼10 mV), uniform, negative baseline shift artifact across the entire preparation. This artifact was measured extracellularly and subtracted from the intracellular recordings to determine the change in transmembrane voltage. None of the cells experienced any change in resting voltage within an accuracy of 1 mV. To determine whether noise variance had changed in CA1 cells due to a DC pulse, F-tests were performed on variance samples before and during the pulse (Microsoft Excel 97). Samples were significantly different ifP < 0.01. Finally, the thermal noise from the electrode in the extracellular space (∼20,000–60,000 μV2) was subtracted from intracellular variance to determine the true noise variance in the cell (Wahl et al. 1997).

Experiments were not performed unless the CA1 cell had a stable resting membrane voltage below −60 mV. In some cases, a holding potential (<10 mV) was necessary to maintain resting voltage (holding current <1 nA). Cells were not used if evoked action potentials were <60 mV or >5 ms at half-amplitude. Recorded voltage data were digitized at 4,000 Hz.

Evaluation of detection

Raw voltage data were analyzed by determining the occurrence of action potentials and generating an output time series (Douglass et al. 1993; Gluckman et al. 1996;Wiesenfeld and Moss 1995). This time series was used to compute the power spectrum of the output. The signal artifact was used as a marker for time index for the splicing (Fig. 1 B; also see next paragraph). Signal-to-noise ratio (SNR) of the data was computed by dividing the power at stimulus frequency by the baseband power in the vicinity of the stimulus spike (Fig. 1 C), as previously described for SR (Stacey and Durand 2000;Wiesenfeld and Moss 1995). To normalize the increase in noise variance among different cells, the variance during the DC pulse was divided by the variance at rest. This value is called the normalized noise variance and used throughout the paper.

Since noise pulse duration was <15 s to avoid tissue damage, it was difficult to obtain sufficient frequency resolution in the spectrum near 1 Hz for a measure of background noise. The data windows were spliced assuming that the output signals were stationary for at least 50 s. This is a valid assumption because the data were wide-sense stationary to both the signal input and to the noise pulse. Stationarity was determined by noting that the cellular response to the periodic signal was unchanged for >2 min (all tested cells: failing cells were rejected as unstable) and that both the mean and covariance of the response to the noise pulse were independent of the time index (3/3 cells tested) (Leon-Garcia 1994). The splicing did introduce another periodic component (<0.1 Hz) that did not affect the SNR calculations, as tested in both experimental and simulated data.

Noise modulation

To analyze the effect of physiological noise on signal detection in neurons, endogenous noise sources must be modulated without affecting synaptic transmission. This was accomplished by changing the excitability of CA3 cells without affecting CA1 cells or any synaptic properties. Negative extracellular electric fields or cathodic currents are known to depolarize cells (Jefferys 1995). It has long been known that current injection can increase activity of CA3 cells (Wong et al. 1979). When applied to the CA3 region, electric fields should therefore increase the neuronal firing of CA3 cells and thus also the synaptic noise on the apical dendrites of CA1.

Computer simulation

A simulation was performed of a single CA1 cell with distributed noise sources and a subthreshold periodic signal. The CA1 cell was implemented in NEURON software (Hines 1993) by using the CA1 model in Stacey and Durand (2000) (Table1). The model includes the following: an active soma with one sodium, one calcium, and four active potassium channels (Warman et al. 1994); passive dendrites; and α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) synapses (Destexhe et al. 1998), some of which fired at random intervals (Stacey and Durand 2000) (Fig.1 D). The amplitude of synaptic events was modulated by changing the maximum conductance (g max) of the AMPA current. A periodic synaptic signal was introduced on the apical dendrite. The amplitude was adjusted so that it was subthreshold (g max = 5 nS) for a baseline somatic noise of 12,000 μV2, the minimum baseline in CA1 cells in the cell (Turner 1988; Wahl et al. 1997). Noise was added (g max = 0.22 nS) at varying mean frequency. Noise events were scaled by a Poisson distributed number (mean = 0.3). Synaptic noise variance was computed by recording the current at the synapses (pA2). Somatic noise variance, which is the value recorded in the hippocampal slices, was evaluated by computing the voltage variance (μV2) at the soma. The synaptic current variance and somatic voltage variance were found to be directly proportional. Simulated data were analyzed exactly as explained above for the brain slice experiments, with the exception that the simulations contained 50 s of data and did not need to be spliced.

View this table:
Table 1.

CA1 model specifications


Brain slice experiment


The effect of the application of the current pulse on the intracellular noise is shown in Fig. 2 A. The noise was characterized by an increase in the frequency of depolarizing events resembling excitatory postsynaptic potentials (1.5–3 mV peak, 5–15 ms duration). During the application of a 50-μA pulse, the noise variance increased from 90,000 μV2 to 270,000 μV2. The increase was characterized by higher frequency and/or higher amplitude EPSPs. Occasionally multiple noise events superposed, causing longer, higher amplitude events (Fig.2 A, insets). The pulse does not cause any measurable change in resting membrane voltage. Pulses >100 μA evolved hydrogen bubbles and could not be used.

Fig. 2.

Noise modulation. A: raw data from a cell showing increased noise during the DC pulse. The mean variance during the pulse is triple the baseline. Top insets: expanded data froma and b, showing size of synaptic events.Bottom insets: inset data further expanded to show single events, some of which may contain multiple, superimposed events.B: plot of the noise data for 3 input currents, showing the relationship between pulse amplitude and variance. The variance for each cell was normalized by dividing by its resting variance. As indicated, current >50 μA significantly increased the noise above baseline.

Response to the DC pulse varied greatly among different cells. On average, three pulse amplitudes (20, 50, and 80 μA) increased the variance by 26% (n = 4), 42% (n = 6), and 219% (n = 6) over baseline, respectively (Fig.2 B). The maximum increase was six times the baseline noise. Paired t-tests showed that 50- and 80-μA current pulses were able to generate noise variance significantly different from control, while 20 μA could not (0 μA, P < 0.05). Since the DC pulse did not change the membrane voltage in CA1 cells and the noise variance increased accompanied by EPSPs, the application of DC current could modulate the noise variance by increased firing of CA3 somata and/or axons.


The effect of noise on detection of subthreshold signals was measured with intracellular recordings in CA1 neurons. A subthreshold, periodic input stimulus was applied to the stratum radiatum, generating an EPSP between 3 and 8 mV in the recorded cell. Addition of noise via the DC pulse raised the response to the signal above threshold in 17 cells from 12 different slices. Recordings were performed for 55 cells, but in 38 of them there was no increase in either noise or detection during the DC pulse and the data were discarded. The large number of failures was attributed in large part to the poor connectivity that exists in a slice between CA3 and CA1 (Bernard and Wheal 1994). In a slice with poor connectivity, it is impossible to produce synaptic noise by this method, so it would be impossible to test for the presence of SR. All failed slices are documented here as a conservative measure because it was not possible to determine in real time whether each failure was due to lack of SR or of connectivity. Therefore noise improved detection in at least 31% of the slices tested, a number that could be much higher if failures due to poor connectivity could have been ignored. The effect of adding the DC current in a cell is shown in Fig.3 A. Before the addition of noise, the signal (which causes negative artifacts in the raw data) is not detected. During the DC pulse interval, 50% (6/12) of the pulses are “detected,” causing action potentials. Detection vanishes when the pulse ends. The data are from the same cell and DC amplitude as Fig. 2 A. Two possible EPSPs, marked by stars, are visible during the pulse. The calculated SNR from 13 of the successful experiments are shown in Fig. 3 B. (Four of the 17 cells tested were not included in these measurements because the input currents could not be measured.) Of the 13 cells that were tested for multiple known current levels, there were only 2 instances in which the SNR did not increase when the current was raised. A nonparametric sign test was performed to test the hypothesis that increasing the current increases SNR. For a null hypothesis that current and SNR are not related, the probability of failure is at least 50% for each cell, depending on the number of observations. Using this conservative probability, the sign test for 13 cells and 2 failures shows that increasing the DC current significantly increases the SNR (P < 0.011). Since increased current also significantly increased the noise (see previous section), it follows that increasing the noise within the range of this experiment improves detection of subthreshold signals in CA1 cells. It is interesting to note that the SNR for 20 μA was clearly improved compared with 0 μA despite the fact that noise (Fig. 2 B) was not significantly increased at that current level.

Fig. 3.

Signal detection improvement. A: application of the current pulse improves signal detection. Signal pulses are referenced below the data. Scale bars are 10 mV and 1 s. Action potentials are top-truncated. B: increasing current improves signal detection. Plot of data from 13 cells with varying current amplitudes, with mean and 1 SD for each current shown.

Figure 3 B displays two important hallmarks of SR: noise elicits detection of a subthreshold signal and detection improves as the noise is increased. Another characteristic of SR is a tapering of SNR at higher noise levels; however, this region of detection could not be tested because the noise produced by the DC pulse could not be increased sufficiently without damaging the slice. These data show that significant changes in detection do occur as the noise increases, even when the increments in noise are very small.


The position of the DC electrode was changed in three cells to test the hypothesis that the improvement in signal detection is due to increased noise and not to an artifact of the DC pulse. As shown in Fig.4 A, the DC electrode was moved away from CA3 cells to a position equidistant from the CA1 region. A DC pulse administered from the secondary position had no effect on detection (n = 3). After replacing the DC electrode over CA3, the improved detection returned (Fig. 4 B), showing that the improved detection was dependent on CA3 depolarization and not due to any field effect on CA1.

Fig. 4.

Experimental controls. A: the noise electrode was moved to a new position equidistant from CA1 to test whether the improved detection was due to a field effect. B: data fromA. Current applied near CA3 cells improves detection, but does not when applied to the other region (n = 4). Detection returns after replacing the electrode near CA3 (n = 2). Current was 80 μA. C: slices were tested before and after an incision was made between the CA1 and CA3 regions to test dependence on synaptic connections.D: after the cut, slices that before had shown increased noise and improved detection during a pulse lost both effects (see Table 2). Current was 80 μA in all cases.


To test the hypothesis that increased noise/detection was dependent on synaptic activity in CA3 cells, three slices with increased noise during the DC pulse and improved detection were incised as shown in Fig. 4 C, severing the Schaffer collaterals. Following the incision, intracellular readings were obtained for several neurons. In every case there was no change in noise or in signal detection for all of the postincision cells during the DC pulse (see Fig. 4 D,Table 2). From the results shown in the two experiments in Fig. 4, we conclude that the increased noise variance in CA1 is due to axonal signals in the Schaffer collaterals and the improved signal detection is dependent on this endogenous noise source.

View this table:
Table 2.

Results of Schaffer collateral incisions

Simulation results

A simulation was performed to examine the effect of adding noise to the cell beyond the range possible in the brain slice experiment. The simulation allowed performing multiple trials on the same cell as well as modulation of noise to levels much higher than those possible in the slice. As in the brain slice, there was no detection for the baseline noise, and the noise variance was augmented by increasing the mean frequency of Poisson distributed synaptic events.

Noise was added to the cell with a range from 12,000 to 700,000 μV2. The signal amplitude was set below threshold for noise variance of 40,000 μV2. The signal threshold and the low end of the noise range are similar to values measured experimentally. Increasing the noise amplitude to 120,000 μV2 progressively improved signal detection in the model (Fig.5 A). Noise variance of 70,000 μV2 improved detection to about 40% of the signal pulses (SNR = 21, compared with zero detection at 40,000 μV2). The maximum SNR occurred with noise at 112,000 μV2, which detected 70% of the pulses (SNR = 118). The steep increase in SNR for small noise intensity is characteristic of SR theory and is very similar to the results obtained experimentally.

Fig. 5.

Simulation results. A: signal detection with increasing noise. The signal was subthreshold for baseline noise, detectable for 70,000 μV2 noise, and overwhelmed by 350,000 μV2 noise. Inset: power spectrum for the 2 noise levels. The 2.5-Hz spike is present in both, but background noise is much higher for the larger noise input. B: intracellular results from 4 cells compared with simulated data and the SR equation fitted to the simulations. The variance in each cell was normalized to its resting level. For noise <10 times baseline, increasing the noise causes an increase in SNR.

Higher levels of noise begin to corrupt the signal and lower SNR. The data for this cell are slightly below those predicted by the SR equation. This is in contrast to the response of neurons to low-frequency, amplitude-modulated noise, which can exceed the predicted values at high variance (Stacey and Durand 2000; Wiesenfeld and Moss 1995). The decreased values in this simulation were due in large part to an effect not addressed by SR theory: oscillatory neuronal behavior. As the noise variance increased to higher values, the neuron began to fire nearly periodically, as shown in Fig. 5 A (bottom). The firing frequency increased with increasing noise, independently of the signal frequency. The power spectrum in Fig. 5 A shows that the signal frequency is present in the output but is overwhelmed by the large hump centered at 6 Hz. The oscillations due to high noise greatly diminished the ability of the cell to detect the signal, lowering the SNR below the value predicted by SR.

The results of the simulated data were fit to the SR equation and compared with the intracellular data obtained from four CA1 cells (Fig.5 B). The noise variance in each cell was normalized to its baseline level to facilitate comparison of the different cells. All five cells show similar qualitative effects within the experimental range of noise up to 10 times baseline. The figure shows that the SNR in this range increases with increasing noise. This positive slope lies on the left side of the SR curve. One cell (cell 3) actually detected 100% of the inputs during the 50- and 80-μA pulses. As noise increases beyond the experimental range, detection in the simulated results is corrupted as the neurons begin to oscillate but is still similar to SR theory.


Choice of noise sources

The key element in testing the effects of SR on signal detection is to ensure that the noise levels used are within the physiological range. Previous work dealing with SR in the hippocampus (Gluckman et al. 1996; Stacey and Durand 2000) did not report measurements of intracellular noise. In both cases, noise was evoked using random stimuli that created global dendritic signals, which are not good representations of the noise in CA1 (Lindner et al. 1995). We have sought to overcome these shortcomings by modulating the endogenous noise sources in the slice. These sources include random release of neurotransmitter as well as uncorrelated true synaptic events (Stacey and Durand 2000), which have roughly the same quantal size (Larkman et al. 1997). Other methods for modulating synaptic noise such as application of phorbol esters (Hestrin et al. 1990;Malenka et al. 1987), hyperosmolar solution (Malgaroli and Tsien 1992; McBain and Dingledine 1992; Stevens and Tsujimoto 1995; Wyllie et al. 1994), or protocol to evoke long-term potentiation (LTP) (Malgaroli and Tsien 1992) are not suitable for this experiment because they also change the synaptic transmission properties of the input signal itself. It was therefore necessary to develop a novel method of noise modulation that does not affect synaptic transmission. By applying a cathodic extracellular current to the CA3 region, the CA3 cells were depolarized without affecting the CA1 cells. Depolarizing the CA3 cells increased their probability of firing spontaneously. Since each CA3 cell could fire independently at random intervals, this method created an increase in independent, random synaptic events spatially and temporally distributed on the CA1 apical dendrites. The noise produced was therefore locally uncorrelated, a good representation of the physiological CA1 noise. The range of noise presented to the cells should lie within physiological limits for the improvement in signal detection to be considered relevant. The noise variance measured at rest varied from 4,000 to 90,000 μV2. This variance is comparable to published values of 10,000–40,000 μV2 for CA1 cells in slices (Sayer et al. 1989; Wahl et al. 1997). With the application of the current pulse, the variance reached maximum levels of 300,000 μV2, which was six times the baseline noise. That level of noise was too small to generate action potentials in the CA1 cells. It was not possible to raise the noise any higher. Thus the noise evoked in vitro was below the level of a detected synaptic event, clearly within the range of inputs normally presented to a CA1 cell.

The simulation also produced noise by increasing the frequency of distributed, independent noise sources. To produce the full range of SR activity, the noise was raised to higher levels than those generated in the slice. The plot in Fig. 5 B shows the variance reaching values 58 times the resting variance (700,000 μV2). This high level of noise could exist in the cell if the hippocampus were in a very active state. Several conditions in the slice increase noise variance, such as bath calcium (Raastad et al. 1992), temperature (Finch et al. 1990), paired pulse modulation (Mennerick and Zorumski 1995), LTP (Malgaroli and Tsien 1992), and random axonal firing (Turner 1988; Wahl et al. 1997). Some of these sources alone are capable of increasing the variance significantly, up to 30 times. Uncorrelated action potentials from CA3 produce even higher noise variance. The variance in a slice can potentially increase by at least 100 times with combinations of these effects.

Even larger noise sources are present in vivo. They are behavior dependent and contain theta rhythms (White et al. 1998), random pulse trains (Leung 1982), and sharp waves (Kamondi et al. 1998), and they may express active and passive states like other areas of the brain (Wilson and Kawaguchi 1996) that increase baseline variance 100-fold (Destexhe and Paré 1999). Thalamic signals are also quite significant (Bertram and Zhang 1999). These high levels, however, probably would be detrimental to normal brain function after extended periods because they evoke constant, random firing in CA1. All these noise sources provide a wide range of variance that may be present in a CA1 cell, potentially reaching very high levels at times. The simulations reported here generated noise within only a small portion of this range, evoking noise that is easily within the physiological range of the slice and only a fraction of the potential range in vivo.

SR in simulated and in vitro CA1 cells

SR predicts that SNR will increase steeply for low noise levels and then fall gradually after peak detection. The experimental data (Fig. 3 B) show significant improvement in signal detection as current is increased. Somatic noise variance also increases as the current is raised, but the changes are less significant (Fig.2 B). Comparing these two figures indicates that average detection improved even for current that was too small to increase the noise significantly (20 μA). The observation that small changes in noise produce significant SNR improvement is a hallmark of SR. Due to the limits on input current to avoid tissue damage and the limited connectivity that exists between CA3-CA1 in a slice preparation (Bernard and Wheal 1994), it was not possible to produce noise large enough to evoke an action potential due to noise alone. Because of this, the analysis of the in vitro experiment was limited to the left side of the SR curve where noise is subthreshold.

The simulation was able to generate the entire range of noise amplitudes to allow curve fitting with the SR equation. Although the simulated cell does deviate slightly from the equation at high noise levels, it shows the characteristic properties of SR. Comparison with the experimental data suggests that some neurons actually have higher detection at low noise amplitude than the simulated cells (cells 3 and 4). We predict that, if presented with higher noise levels, the in vitro CA1 cells would begin to fire randomly, inevitably decreasing the SNR. We hypothesize based on the simulations that the cells' response would approximate the SR curve in Fig.5 B.

Physiological relevance of SR in CA1 neurons

CA1 neurons perform the complicated task of integrating tens of thousands of synaptic inputs. Since CA1 neurons are the final integrators in the hippocampal circuit (Traub and Miles 1991), evaluating signal detection in these neurons should have broad implications in understanding basic hippocampal function: the formation of memory. For a synaptic event to produce an action potential, up to 300 EPSPs must occur simultaneously (Andersen 1990; Sayer et al. 1989). Many of these synapses are electrotonically distant and greatly attenuated at the soma (Spruston et al. 1993). This results in nearly countless combinations that can evoke an action potential (Bernard and Wheal 1994). What determines whether a particular signal will be detected? A straightforward method of signal detection is for 300 CA3 cells to fire a deterministic and simultaneous barrage at a particular CA1 cell. But the broad spectrum of activity and the propensity for random events in the CNS suggest that this deterministic model is not sufficient to explain all signal detection in CA1. SR provides a means to explain the stochastic nature of the system.

Recently, SR was suggested as a possible means of both using noise to improve signal detection and overcoming the attenuation effects in passive CA1 dendrites (Stacey and Durand 2000). Special noise modalities have also been shown to be able to further improve the response beyond traditional SR (Feng and Tirozzi 2000;Locher et al. 2000). To show that SR is actually involved in signal detection in CA1 cells, it was necessary to show that the effect occurred under physiological conditions with endogenous noise. Improvement in signal detection occurred with very small amounts of noise, clearly in the physiological range. Detection is not solely dependent on 300 presynaptic neurons simultaneously firing but, for a signal close to threshold, also can be improved by any synaptic activity uncorrelated to the signal. The varied noise sources within the hippocampus thus potentially provide the CNS with a method to control detection.


We have developed a method to modulate synaptic noise in CA1 cells without changing the synaptic properties of the network. This method allowed us to test the role of noise in improving signal detection in rat hippocampal CA1 cells, as predicted by SR theory (Stacey and Durand 2000). A computer model verified the experiment and extended the results to higher noise levels. These results show that noise can influence signal detection significantly. The effect is especially prominent for signals close to threshold even with small changes in endogenous noise. The large number of synaptic connections and noise sources in the hippocampus can produce a broad range of variance, providing the CNS with a powerful method to influence detection of subthreshold signals merely by changing the characteristics of the noise in the system.


The authors thank J. Hammel for helpful suggestions.

This work was supported by The Whitaker Foundation and the Medical Scientist Training Program at Case Western Reserve University.


  • Address for reprint requests: D. M. Durand, Dept. of Biomedical Engineering, Case Western Reserve University, CB Bolton Rm. 3510, 10900 Euclid Ave., Cleveland, OH 44106 (E-mail:dxd6{at}


View Abstract