|Home | About | Journals | Submit | Contact Us | Français|
Action video game play benefits performance in an array of sensory, perceptual and attentional tasks that go well beyond the specifics of game play [1-9]. That a training regimen may induce improvements in so many different skills is notable as the majority of studies on training-induced learning report improvements on the trained task but limited transfer to other, even closely related tasks [10, but see also 11-13]. Here we ask whether improved probabilistic inference may explain such broad transfer. Using a visual perceptual decision making task [14, 15], the present study shows for the first time that action video-game experience does indeed improve probabilistic inference. A neural model of this task  establishes how changing a single parameter, namely, the strength of the connections between the neural layer providing the momentary evidence and the layer integrating the evidence over time captures improvements in action-gamers behavior. These results were established in a visual, but also in a novel auditory task, indicating generalization across modalities. Thus, improved probabilistic inference provides a general mechanism for why action video game playing enhances performance in a wide variety of tasks. In addition, this mechanism may serve as a signature of training regimens that are likely to produce transfer of learning.
To test the hypothesis that video-game experience leads to improved probabilistic inference, VGPs and NVGPs were first compared on a standard coherent dot motion direction discrimination task . In this task, subjects viewed a dynamic random dot motion display and were asked to indicate the direction of coherent motion (left or right) as quickly and accurately as possible by means of a key press (Figure 1A).
One of the benefits of this task is that psychometric and chronometric curves are well captured by diffusion-to-bound models (DDM) or variations thereof [17-27] as well as a recent neural model of decision making , which, unlike the standard DDMs, has a clear probabilistic interpretation. This interpretation is based on recent theories in which neural patterns of activity represent probability distributions over the encoded variables [28-31]. Critically, this allows us to assess whether the changes that arise as a result of video-game experience meet our standards for improved probabilistic inference. These can be defined rigorously in the task we chose by considering decision making from a probabilistic perspective. Before committing to a choice the best a subject can do is to compute the probability that each choice is correct given all the evidence up to the present time, a distribution also known as the posterior distribution over choices (which we denote p(c|e) where c are the choices, and e is the evidence). This computation requires knowledge of p(e|c), the statistics of the evidence. Indeed, through Bayes rule, p(c|e) is proportional to p(e|c). In general, subjects do not initially know the statistics of the evidence, which in turn means that the calculated posterior distribution over choices is not the optimal one. However, through repeated exposure to a task, subjects have the opportunity to learn these statistics and, as a result, can make decisions on the basis of a more accurate posterior distribution. This type of improved inference is what we mean by making better use of the evidence.
The DDM fits to the data provide support for this hypothesis. The results show an overall decrease in decision bound, no change in non-decision time and crucially a greater integration rate in the VGP population (Figure 1B and 1C). This latter result indicates increased sensitivity in VGPs suggesting that VGPs may indeed make better use of the evidence than NVGPs. Yet to definitively make this point, it is necessary to demonstrate that VGPs use better posterior distributions over choices than those used by NVGPs. As the DDM used here does not provide a natural way to compute a posterior distribution (particularly when the coherence varies over trials), we took advantage of a recent neural model of probabilistic decision making  in which the state of the network encodes the posterior distribution over choices given the evidence (Figure 2A).
We first used this model to fit the psychometric and chronometric curves of VGPs and NVGPs using a numerical optimization procedure (Figure 2B; for details see Supplemental Section 2A). Data from VGPs were well modeled by assuming only an increase in the strength of the feedforward connectivity between the two layers of the network representing areas MT and LIP as compared to NVGPs. A change in this single parameter naturally accounts for the finding that only RTs change whereas accuracy stays constant. Consistent with this proposal, Law and Gold [32, 33] recently showed that improved performance on a motion discrimination task following extensive training on the motion task itself is also captured by a strengthening the connections between the sensory and integration layers.
We then asked whether action video-game experience is indeed associated with a better posterior distribution for a fixed amount of evidence. In the case of a binary decision task (i.e. left versus right), the quality of this posterior distribution can be assessed by calculating the log odds (i.e. the log of the ratio of the probability that the dots move rightward over the probability that the dots move leftward). Because the log odds in the model reflects the quality of the evidence available to the decision maker, if action video-game players do make better use of the evidence, we should find an increase in the absolute value of the log odds in their model. This is indeed what was observed, with the log odds being 19% higher on average across all coherences, confirming that, according to the probabilistic population network, VGPs make better use of the evidence than NVGPs (for details see Supplemental Section 2A).
A key question concerns the generalization of this finding to other tasks and domains. To ask whether the improved probabilistic inference noted in VGPs generalizes to other tasks and domains, an auditory analog of the motion direction task was developed (Figure 3A, for details see Supplemental Section 2B). As in the motion task, performance on the task requires the accumulation of information over time, and thus the intuitions regarding sensitivity and decision bound are similar and can be modeled in the same manner. The results in Experiment 2 mirror those of Experiment 1 quite closely (Figure 3B). In the DDM, the integration rate parameter was significantly greater in VGPs than in NVGPs confirming greater sensitivity to the stimulus in VGPs. The decision bound parameter was significantly smaller in VGPs than in NVGPs and no significant difference was observed in the non-decision time parameter. The neural model confirmed that the difference between NVGP and VGP can be captured by changing solely the strength of the feed-forward connections (by 51%, Supplementary Information, Figure S1A). We further confirmed that VGPs do indeed make a better use of the evidence in this paradigm, with log odds being 48% higher on average across all levels of signal to noise ratio in VGPs as compared to NVGPs.
While our hypothesis is that extensive video-game experience is at the root of these changes, it could also be the case that VGPs are individuals who have been born with improved abilities at performing probabilistic inferences. To establish that video-game experience is sufficient to drive a better use of the evidence, NVGPs underwent 50 hours of video-game training on either action video-games or a control video-game. If action video-game experience does improve probabilistic inferences in neural circuits, larger improvements should be noted in the action-trained group than in the control group. This prediction was confirmed as the results of this training study are nearly identical to the results of Experiments 1 and 2 (Figure S2, for details see Supplemental Section 2C). Critically, better use of the evidence was noted only in the action group at post-test. The log odds for the motion task were 16% (respectively 31% for auditory) higher on average across all levels of coherence (respectively SNR) at post-test as compared to pre-test for the action group (see Supplementary Information, Figure S3). The results of this experiment therefore confirm the causal relationship between action video-game experience and improvements in probabilistic inference.
A remarkable feature of our results is that action video game experience leads to a decrease in RTs while leaving accuracy unchanged. At first sight, this pattern could be consistent with a reduction in non-decision time, for instance, through a decrease in motor latencies. However, a change in non-decision time alone cannot capture the present data because it predicts a constant difference between VGP and NVGP reaction times as coherence varies, whereas we report larger reaction times differences at lower than higher signal strengths. This was confirmed by our fitting procedure, which revealed that the sensitivity and the bound height must be changed to capture our data, and that non-decision time is in fact not significantly different across populations. Nonetheless, to confirm our key result, namely, that the sensitivity to stimuli has increased between NVGPs and VGPs, we ran our motion and auditory tasks in a condition in which stimuli were presented for fixed durations and the subjects were instructed to watch/listen to the entire stimulus before reporting their decision. We then plotted the accuracy of the response as a function of duration. Since this task does not involve speeded reaction time, non-decision time is irrelevant in accounting for these data (for details see Supplemental Section 2D).
As predicted by higher sensitivity in VGPs than in NVGPs, we found the rate with which accuracy increases as a function of duration (the parameter β in Figure 4) to be higher in VGPs [34, 35]. Simulations of the drift diffusion model with fixed duration confirmed that a change in sensitivity between VGPs and NVGPs is required to account for the data in Figure 4. We first ran a simulation in which the sensitivity for VGPs and NVGPs were set to be equal, while the bounds heights were independently adjusted for VGPs and NVGPs to maximize the fit to the experimental data. This simulation revealed that it is impossible to capture the faster rise of accuracy for VGPs, particularly for short durations (Supplementary Information, Figures S4A and S4B). Next, to determine whether the change in sensitivity estimated from Experiments 1 and 2 can account for the critical duration data, we ran simulations in which the sensitivity ratios between VGP and NVGP were the same as estimated from the reaction time experiments (E1 and E2). Under these conditions, the model captures the results in Figure 4 well thus confirming that sensitivity must be higher in VGPs compared to NVGPs (Supplementary Information, Figures S4C and S4D).
Action video-game experience results in more efficient use of sensory evidence. Importantly, these improvements are not restricted to the visual modality, but appear in the auditory modality as well. Moreover, 50 hours of action video-game training led to qualitatively similar results in a group of NVGPs establishing a causative relationship between action video-game experience and improved probabilistic inference.
To establish the role of improved statistical inference in VGP performance, full chronometric and psychometric curves were measured. Using DDMs, increased sensitivity (along with a compensatory decrease in bound) provided the best fit to the data, thus confirming the hypothesis that video game play results in a more efficient use of the evidence. That statistical inference is improved in gamers was further confirmed by establishing enhanced performance in gamers in fixed duration paradigms. Note that these results were far from being a foregone conclusion. While faster RTs in gamers were expected, this could have resulted from a speed-accuracy tradeoff through a change in the decision bound and/or faster motor execution through a change in the non-decision time, without concomitant changes in sensitivity. In contrast, the six experiments presented establish the role of increased sensitivity in gamers’ performance.
The neural model framework provides converging evidence for this claim as it captures VGP behavior by enhancing the connection strength between the layer providing the momentary sensory evidence and the layer integrating the evidence, leading to a more efficient use of the evidence. Increasing the connection strength not only increases the amount of information per unit of time transmitted to the output layer but also induces larger fluctuations in the membrane potential of neurons, which in turn lead the network to reach the bound faster. The net effect on accuracy is negligible since the output layer receives more information per unit of time, but integrates information over shorter durations. As a result, the percentage of correct responses at the bound remains stable. A novel contribution from the neural model therefore is that it naturally accounts for one striking regularity of the data with just one parameter change, and crucially this parameter controls the quality of the statistical inference. One might wonder why the nervous system does not keep the feedforward weights set to a maximum value since our results show that increasing the feedforward weights lead to better performance. This might be in part because of the metabolic cost of maintaining synapses with high efficacies. But, more importantly, it is not the case that increasing the feedforward weights always increases performance. We have found analytically that there is an optimal value of the feedforward weights, beyond which a further increase worsens performance (Beck, J., Bejjanki, V., & Pouget, A., Fisher information in stimulus correlated spiking networks, submitted). The optimal value of the weights depends on the statistics of the spike trains in the sensory layer.
The fact that action gamers perform better probabilistic inference is appealing as it provides a unified mechanism to explain why action video-game training improves participants’ performance in seemingly widely different tasks such as contrast detection, visual search, multiple object tracking, letter recognition with flankers and decision making [1, 6, 7, 9, 36]. However, such an improvement may also have much wider implications. Decision making can be formalized as a probabilistic inference in which subjects must compute the posterior distribution over choices given the evidence at each time step [37-39]. This iterative process requires that the posterior distribution at the previous time step be multiplied by the likelihood of the evidence at the present time step (see equation A1 in Supplementary Information). To be done optimally (i.e., without loss of information) this process requires both that the multiplication step be faithfully executed and also that the subjects have perfect knowledge of the likelihood function, which is to say, perfect knowledge of the statistics of their sensory input. It is highly unlikely that our subjects came to the task equipped to perform optimally from the get-go. At the beginning of the experiment, our subjects were naive to the visual and auditory stimuli. Their lack of knowledge of the statistics of our particular stimuli would initially result in a poor estimate of the likelihood of the visual or auditory evidence. Similarly, subjects were initially naïve to the inference task used in our studies. Consequently, they had to learn the statistics of our particular stimulus set (more accurate likelihood of the evidence) as well as the appropriate inference for this task (performing the multiplication that permits the update over time of the posterior distribution). One possibility is that VGPs perform better in those tasks because they learn a better model of the stimuli used, and/or perform the inference more accurately , which in turn leads to a better posterior distribution.
This proposal is closely related to a dominant idea in the field of perceptual learning by which learning occurs through template matching by reweighing the connectivity between sensory and decision stages [40-42]. Law and Gold [32, 33] have recently documented how the effective connectivity between MT and LIP changes as monkeys are trained and then tested with motion stimuli similar to ours. Our results build on this work, but establish that such a reweighing is not necessarily the product of extensive training with a specific class of stimuli as is standard in the field of perceptual learning. Indeed, improvements after action game training are not limited to playing the game itself, but generalize to new tasks. Gamers, we propose, acquire the ability to dynamically retune the connectivity between the momentary evidence layer and the layer integrating the evidence based on the statistics of the very task they are performing.
This type of learning may be a consequence of the nature of action video game training. Unlike standard learning paradigms, which have a highly specific solution, there is no such specific solution in action video games as situations are rarely, if ever, repeated. Thus, the only characteristics that can be learned are how to rapidly and accurately learn the statistics on the fly and how to accumulate this evidence more efficiently. We can only speculate as to what might be the neural mechanisms responsible from this remarkable transfer of learning, as there are several possibilities. First, the neural circuits involved in evidence accumulation might be shared across modalities, a plausible explanation given that the cortical areas that have been implicated in evidence accumulation are often multimodal . Second, these areas are likely to be under the influence of shared fronto-parietal networks, providing a common source for performance improvements, a view supported by the impact of working memory training in cognitive improvements [44-46]. Finally, learning might involve the global release of neuromodulators that improve probabilistic inferences across all circuits, as has been proposed for instance for noradrenalin . An important issue for future work will be to determine the relative role of these mechanisms in fostering such a wide transfer of learning.
We thank Ashley Anderson, Renjie Li and David Feder for help with subjects, Mike Sugarman for help with figure preparation, Jan Drugowitsch for help with DDM simulations, and Josh Tenenbaum for helpful discussions. This work was supported by National Institute of Health grant EY016880 and Office of Naval Research grant N00014-07-1-0937.3 to DB, a NIDA grant #BCS0346785 and a research grant from the James S. McDonnell Foundation to AP, as well as award P30 EY001319 from the National Eye Institute. The content of the manuscript is solely the responsibility of the authors and does not necessarily represent the official views of the National Eye Institute or the National Institute of Health.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.