Making up your mind: Enhanced perceptual decision-making induced by stochastic resonance during non-invasive brain stimulation: Stochastic resonance in perceptual decision-making

Perceptual decision-making relies on the gradual accumulation of noisy sensory evidence until a specified boundary is reached and an appropriate response is made. It might be assumed that adding noise to a stimulus, or to the neural systems involved in its processing, would interfere with the decision process. But it has been suggested that adding an optimal amount of noise can, under appropriate conditions, enhance the quality of subthreshold signals in nonlinear systems, a phenomenon known as stochastic resonance. Here we asked whether perceptual decisions obey these stochastic resonance principles by adding noise directly to the visual cortex using transcranial random noise stimulation (tRNS) while participants judged the direction of motion in foveally presented random-dot motion arrays. Consistent with the stochastic resonance account, we found that adding tRNS bilaterally to visual cortex enhanced decision-making when stimuli were just below, but not well below or above, perceptual threshold. We modelled the data under a drift diffusion framework to isolate the specific components of the multi-stage decision process that were influenced by the addition of neural noise. This modelling showed that tRNS increased drift rate, which indexes the rate of evidence accumulation, but had no effect on bound separation or non-decision time. These results were specific to bilateral stimulation of visual cortex; control experiments involving unilateral stimulation of left and right visual areas showed no influence of random noise stimulation. Our study is the first to provide causal evidence that perceptual decision-making is susceptible to a stochastic resonance effect induced by tRNS, and that this effect arises from selective enhancement of the rate of evidence accumulation for sub-threshold sensory events.


22
The authors declare no competing financial interests. above, perceptual threshold. We modelled the data under a drift diffusion framework to 48 isolate the specific components of the multi-stage decision process that were influenced by 49 the addition of neural noise. This modelling showed that tRNS increased drift rate, which 50 indexes the rate of evidence accumulation, but had no effect on bound separation or non-51 decision time. These results were specific to bilateral stimulation of visual cortex; control 52 experiments involving unilateral stimulation of left and right visual areas showed no 53 influence of random noise stimulation. Our study is the first to provide causal evidence that 54 perceptual decision-making is susceptible to a stochastic resonance effect induced by tRNS,55 and that this effect arises from selective enhancement of the rate of evidence accumulation 56 for sub-threshold sensory events.

Results and Discussion 70
Noise is an intrinsic property of all biological systems [2]. Typically, noise is viewed as being 71 detrimental for neuronal computations and the behaviors they regulate [2,3], including 72 decision-making [4]. A key limiting factor in decision-making arises from noisy 73 representations of sensory evidence in the brain [5,6]. On this view, noisy sensory 74 information representations are not optimal, and this leads to errors in decisions. However, 75 small amounts of noise added to a nonlinear system can increase the stimulus quality by 76 increasing the signal-to-noise ratio (SNR) [7]. This phenomenon is known as stochastic 77 resonance, and its expression has been demonstrated in different sensory modalities [8][9][10]. 78 Stochastic resonance occurs when an optimal amount of noise is added to a sub-threshold 79 signal, which makes the signal cross a decision threshold, and therefore enhances detection 80 performance ( Figure 1). 81 82 83 Figure 1: Stochastic resonance occurs when an optimal level of noise is added to a subthreshold signal. In this example the signal alone (red sinusoid) remains below the perceptual threshold (dotted line). Adding an optimal amount of noise (grey line) periodically raises the stimulus above the system threshold. If the added noise is too weak, the threshold is not crossed. Conversely, if the noise is too strong the signal remains buried and cannot be discriminated from the noise[1].

116
We used a random-dot-motion task (RDM-task) as the perceptual decision-making task of 117 choice (see Supplemental Information). The RDM-task is widely used in studies of 118 perceptual decision-making, and has well characterized neural correlates [24,25]. 119 Participants fixated on a centrally presented array of randomly moving dots within which a 120 proportion of the dots moved coherently in a common direction (leftward or rightward; see 121 Figure 2B). Participants judged the common direction of movement (two-alternative forced-122 choice/2-AFC) as quickly and accurately as possible. The difficulty of the task was 123 parametrically manipulated by altering the proportion of signal dots that moved coherently in 124 a given trial (3%, 6%, 12%, 25% or 50% coherence). The benefit of this task is that it allows 125 for the continuous accumulation of sensory evidence over a period of several hundred 126 milliseconds, which facilitates investigation of the underlying processes involved in decision-127 making [21,26,27]. 128

129
If the stochastic resonance model applies to perceptual decision-making, then the addition of 130 relatively small amounts of noise should enhance motion discrimination performance. The 131 added noise will likely increase the quality of the sensory evidence for coherent motion trials 132 in which the signal is just below threshold, but not for trials in which the signal is well below 133 framework used to model perceptual decision-making in the dot motion task. In the model, evidence is accumulated over time until a response boundary is crossed. t is the non-decision time, which includes the time taken to execute a motor response. v is the drift rate, which reflects the rate at which sensory evidence is accumulated. This parameter is taken as an index of the quality of sensory information. a represents the boundary separation (correct at the top, incorrect at the bottom), indicating how much information is needed to make a decision. B: Schematic of the random dot-motion task in which participants judged whether signal dots moved on average to the left or right. Task difficulty was titrated by altering the proportion of coherently moving dots (shown with arrows attached, for purposes of illustration) amongst randomly moving dots. In this example the coherent motion is rightward, but in the experiment the dots were equally likely to move toward the left or right. For display purposes, we depicted here the outline of the imaginary circle in which the dots were presented. or above threshold. Conversely, the addition of large amounts of noise should either have no 134 effect on perceptual thresholds, or should impair performance slightly for displays at or 135 above threshold [8]. We therefore applied four different tRNS intensities (0.25, 0.375, 0.5 136 and 0.75 mA; 100-640 Hz zero-mean Gaussian white noise) while participants engaged in the 137 RDM-task. These tRNS intensities result in current densities that we have shown previously 138 are able to induce a stochastic resonance effect in a visual contrast detection task [18]. In Experiment 1, we stimulated visual cortex bilaterally with tRNS in 15 participants (see 149 Figure 3 and Figure S2). The coherence levels of 3% and 6% were subthreshold (average 150 detection performance < 0.63%), i.e., performance was below the detection threshold, which 151 corresponded to 75% correct in our task. For the analysis, we calculated the group %correct- where i denotes each of the 4 tested noise intensities. As shown in the left panel in Figure 3, 159 for the 6% coherence condition, which was just below threshold in the no-tRNS condition, 160 motion discrimination performance improved when tRNS was applied at a relatively low 161 intensity, whereas performance remained unaffected for the other coherence levels and noise 162 intensities. To quantify these effects, we performed a 4 (tRNS intensity) x 5 (coherence level) 163 within-subjects ANOVA on the %CCI data. There was a significant interaction between 164 coherence level and tRNS-intensity (F(12,156) = 2.47 p < 0.01, Cohen's f= 0.43). To isolate 165 the source of this interaction, one-way ANOVAs were conducted for each coherence level 166 separately. For the 6% coherence condition only (red symbols in Figure 3), performance was 167 significantly affected by the different tRNS intensities (F(3,39) = 3.56 p = 0.02 Cohen's 168 f=0.52 ). There were no other significant main effects or interactions for the coherence 169 conditions of 3%, 12%, 25% or 50%. Post-hoc tests were conducted to compare performance 170 in the 6% coherence condition at each noise level against the baseline. All p-values were 171 corrected for multiple comparisons. These comparisons revealed that a tRNS intensity of 172 0.25mA significantly enhanced motion discrimination performance relative to baseline (t(13) 173 = 3.39 p corrected < 0.02). A similar enhancement was evident for the 6% coherence level at an 174 intensity of .375mA, but this effect did not survive our stringent correction for multiple 175 comparisons, (t(13) = 2.53, p corrected > 0.1). These results suggest that perceptual decision-176 making for sensory stimuli that are just below threshold can be improved by adding a small 177 amount of neural noise over bilateral visual cortex, consistent with predictions arising from 178 the stochastic resonance principle [8]. 179   180   181 Next we employed the drift diffusion framework to accurately model the processes involved 182 in decision-making based on the accuracy and response time data obtained from the decision-183 making task. Specifically, we used the hierarchical drift diffusion model (HDDM,[30]) to 184 determine which aspect of decision-making was affected by tRNS. We normalized the 185 obtained DDM-parameters relative to the zero noise condition in the same way as the 186 behavioral data, as described above. As shown in the right panel of Figure 3, the drift rate 187 was markedly affected by tRNS for the 6% coherence condition, whereas it appears to be 188 unaffected for the remaining coherence levels. We submitted the drift-rate parameter to a 5 x 189

237
Analysis of the baseline data in all three experiments revealed no interaction between 238 coherence level and tRNS intensity (repeated measures ANOVA with within-subjects factor 239 of coherence level and between-subjects factor of experiment, F(2,39) = 1.15, p > .32), 240 suggesting that the stochastic resonance-effect observed in Experiment 1 was not driven by 241 differences in baseline performance between the three experiments. Across all three 242 experiments, there was a highly significant main effect of coherence level on performance, as 243 expected. For completeness, we also report here a small number of significant main effects 244 which are not related to the central stochastic resonance hypothesis under examination in this 245 study (see also Figure S2). First, there was a small but consistent main effect of tRNS 246 intensity on accuracy during right visual cortex stimulation, F(3,39) = 3.13 p = .036, Cohen's 247 f =0.49. Post-hoc contrasts revealed that this effect was driven by overall poorer performance 248 for the .25mA tRNS intensity, regardless of motion coherence level, t(69) = -2.78 p corrected < 249 0.03. This decrease in performance was mirrored by a significant main effect of tRNS-250 intensity on drift rate (see Figure 5B and Table S1), (F(3,39) = 4.54 p < .01, Cohen's f = 0.59, 251 which was again specific to the .25mA tRNS intensity, (t(69) = 2.67 p corrected = .04), 252 regardless of motion coherence level. Second, there was a significant main effect of 253 coherence level on bound-separation during stimulation of the right visual cortex, F(4,52) = 254 3.09 p = .024, Cohen's f = 0.4 (see Supplemental Information). Post-hoc tests showed that the 255 bounds were significantly closer together for the highest (50%) coherence condition, t(55) = 256 3.16 p corrected < .04, relative to baseline), but there were no significant effects on bound 257 separation for the other coherence levels. 258 259

Conclusions 260
We have shown that adding a small amount of noise bilaterally to the visual cortex can 261 enhance perceptual decision-making in a motion discrimination task, particularly for 262 subthreshold stimuli (6% coherence). When modeled as a drift-diffusion process, this tRNS-263 induced performance improvement coincided with an increase in the rate of evidence 264 accumulation, reflected as a change in the model's drift-rate parameter. The same model 265 revealed no change in either bound-separation or non-decision time, suggesting that an 266 optimal level of neural noise exclusively improves perceptual decision-making by enhancing 267 sensory information quality, consistent with a stochastic resonance mechanism ([8-10], see 268 Supplemental Information for the model fits). In line with previous work [18], we showed 269 that the stochastic resonance effect was strongest when appropriate tRNS intensities were 270 added to the 6% coherence condition, i.e. to a subthreshold stimulus, as indicated by the 271 average baseline detection accurarcy of 60%. Note that all tRNS intensities and coherence 272 levels were randomized over participants to account for any aftereffects, fatigue or learning 273 effects across conditions. 274 275 There was no evidence for a stochastic resonance effect when noise was applied unilaterally 276 to the visual cortex. This absence of a performance-enhancing effect for unilateral tRNS was 277 not due to differences in baseline performance between the groups: detection performance in 278 the 6% coherence condition was similar across experiments (Experiment 1 -60%; 279 Experiment 2 -58%; Experiment 3 -57%). Modelling of the electrical field for each 280 electrode montage (Figure 4) indicated a higher peak current when the tRNS was applied 281 bilaterally than in the unilateral stimulation conditions. It is unlikely that this apparent 282 difference in current densities prevented a stochastic resonance effect for the unilateral 283 stimulation conditions, however, because the same absolute current densities during bilateral 284 stimulation were also reached during unilateral stimulation but at higher tRNS intensities. 285

286
The visual stimuli employed in our motion discrimination task were always presented in the 287 centre of the display, and thus would have been processed initially by visual areas in both the 288 left and right hemsipheres [35]. Given that unilateral visual cortex stimulation did not 289 influence motion-discrimination performance, it is most parsimonious to conclude that visual 290 areas in both hemispheres must be stimulated concurrently with tRNS for the stochastic 291 resonance effect to occur. Because of the relatively diffuse nature of transcranial electrical 292 stimulation in general [36], it is not possible to determine the specific anatomical regions that 293 mediate the stochastic resonance effect we observed. The primary visual cortex (V1) [37] and 294 motion area V5/MT are both crucial for the processing of dynamically moving visual stimuli 295 [38][39][40]. These two areas are highly interconnected, so our bilateral stimulation protocol 296 might have impacted motion processing in area V5/MT, enhanced signal quality in area V1, 297 or both. Further work using more focal stimulation techniques (e.g., transcranial magnetic 298 stimulation) will be needed to pinpoint the visual areas responsible for the stochastic 299 resonance effects we report here. 300

301
Our results are in line with recent work that employed a similar task to show that decision-302 making is sensitive to the addition of noise to visual motion stimuli [41]. Critically, our 303 findings extend these results by demonstrating that a stochastic resonance effect can be 304 induced in a decision-making task when noise is directly applied to the visual cortex with 305 tRNS [42,43]. Moreover, we are the first to show that this stochastic resonance effect 306 enhances the quality of information processing as indicated by an accelerated rate of evidence 307 accumulation. The underlying mechanism for the observed tRNS effect is not completely 308 understood. However, single unit recordings have shown that sensory neurons in the visual 309 cortex are sensitive to a stochastic resonance mechanism, e.g., there is an increase in the SNR 310 of the firing rate of neurons when an optimal level of noise is applied to a visual stimulus 311 [44]. This is likely due to the recruitment of voltage-gated sodium channels by the noise [45-312 47].