Improved probabilistic inference as a general learning mechanism with action video games

Abstract

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 because 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. By 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 [16] 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.

Figure 1. Visual Motion Direction Discrimination

A. Task. Subjects viewed a dynamic random dot motion display and were asked to indicate the net direction of motion (left or right, here the correct answer would be right). On every trial, some proportion of the dots moved coherently (top panel = 50% coherence, middle panel = 25% coherence, bottom panel = 0% coherence) either to the left or to the right, while the remaining dots were replotted randomly. By parametrically varying the number of coherently moving dots from very few to many, full psychometric and chronometric curves could be obtained. B. Behavior. While VGPs and NVGPs demonstrated equivalent accuracy - p = .65, p-eta² = .01 (top panel), VGPs responded substantially faster than NVGPs - F(1,21) = 18.9, p < .001, p-eta² = .47 (bottom panel). Importantly, this factor interacted with motion coherence due to a greater decrease in RTs at low than high coherence – F(6, 126) = 3.5, p < .001, p-eta² = .15. In this and all other psychometric and chronometric curve figures, error bars correspond to between-subject standard error. C. Drift Diffusion Model (DDM). The accumulation of the noisy sensory evidence is simulated by the diffusion of a particle upward or downward until a decision bound is reached. DDM models generate psychometric and chronometric curves that are constrained by three main variables [14]: (1) the rate at which information is accumulated over time, (2) the height of the decision bound at which the system stops accumulating evidence and a decision is made and, (3) the non-decision time, an additive amount of time that is common to all tasks and reflects non-decision processes such as motor planning and execution. To quantitatively assess the individual contribution of integration rate, decision bound and non-decision time, RT and accuracy data were simultaneously fit to each subject’s data with the proportional-rate diffusion model as in Palmer and colleagues[14]. The fits were good and equivalent in the two groups (r²_VGP = .93, r²_NVGP = .90, p = .36). The rate of integration was greater in the VGP than the NVGP group (t(21) = 2.6, p = .02, Cohen’s d = 1.13, top panel), while the opposite result was observed for the decision bound (t(21) = 3.6, p = .002, Cohen’s d = 1.57, middle panel). No difference was observed between the groups in non-decision time (p > .7, Cohen’s d = 0.14 bottom panel), eliminating an additive, post-decision process as a possible source of group differences. Data from individual subjects were fit separately and error bars correspond to between-subject standard error.

Figure 2. Neural Model – Motion Direction Discrimination

A. Neural Network Architecture. The network consists of two interconnected layers of neurons with Gaussian tuning curves. In MT, the sensory layer, the tuning curves are for direction of motion, while in LIP, the integration layer, the tuning curves are for saccade direction, as a proxy in our case for a left/right decision. Note that we do not mean to say that LIP is the only area involved in this process---the label is used mostly for convenience (the same is true for the MT label in the input layer). The layers differ by their connectivity and dynamics. The MT neurons send feed-forward connection to the LIP neurons. Each LIP neuron receives a Gaussian pattern of weight centered on the MT neuron with the same preferred direction. The LIP neurons also have lateral connections to implement short range excitation and long range inhibition as well as a long time constant (1s) allowing them to integrate their input. Each panel indicates a representative pattern of activity in terms of spike count 200 ms into a trial for the sensory layer (MT, bottom panel) and the integration layer (LIP, top panel). B. Neural Model Fit. As with the DDM, the fits were good and equivalent for the two groups. The neural model captures the change in performance from NVGP to VGP with a 55% increase in the conductance of the feed-forward connections between the sensory (MT) and the integration (LIP) layers and, in contrast to the DDM, nearly no change in the bound height (a 1% decrease, which is within the resolution of our numerical maximization procedure; see Supplemental Section 2A). The conductance controls the amount of information transmitted from the sensory layer to the integration layer per unit time, or the strength of the feed-forward connections. It is important to note that this effect is not analogous to a simple change in sensitivity in DDMs. While increasing the conductance does increase the amount of information transmitted from the sensory to the integration layer (which intuitively should increase accuracy), it also induces large fluctuations in the membrane potential of the neurons. These fluctuations lead the network to reach the bound faster, thus lowering the percentage of correct responses. These two effects cancel one another over a wide range of parameters, allowing a single change in feed-forward strength to alter reaction time while leaving the psychometric curve nearly unchanged. Model fit corresponds to best fit to the mean data rather than the mean of individual fits.

Figure 3. Auditory Tone Location Discrimination

A. Task. A pure tone embedded in a white noise mask was presented in one ear, while white noise alone was presented in the other (both ears being normalized to the same mean amplitude). The subjects’ task was to indicate with a button press the ear in which the tone was present as quickly and accurately as possible. In a manner consistent with adjusting the coherence level of the motion stimulus, the ratio of the amplitude of the target tone to the white noise mask was manipulated in order to test performance across the range of possible accuracy levels and reaction times (high amplitude - top panel, low amplitude- bottom panel). B. Behavior. As in Experiment 1, VGPs and NVGPs demonstrated equivalent accuracy - p = .32, p-eta² = .05 (top panel), VGPs responded substantially faster than the NVGPs - F(1,21) = 20.6, p < .001, p-eta² = .50 (bottom panel) and the RT difference between groups was greater at lower signal-to-noise ratios (SNR) - F(7,147) = 5.2, p < .001, p-eta² = .2 C. Drift Diffusion Model. The rate of integration was greater in the VGP than the NVGP group - t(21) = 3.8, p = .001, Cohen’s d = 1.66 (top panel), while the opposite result was observed for the decision bound - t(21) = 2.6, p = .02, Cohen’s d = 1.13 (middle panel). No difference was observed between the groups in non-decision time - p > .05, Cohen’s d = .82 (bottom panel). Data from individual subject were fit separately and error bars correspond to between-subject standard error.

Figure 4. Critical Duration Experiment Results

Accuracy of the VGPs and NVGPs for two levels of visual coherence (Panel A) and two levels of auditory SNRs ( Panel B) as a function of presentation duration (see Supplement Information). Individual subject data were modeled as a simple exponential rise to an asymptote (VGP = thinner lines; NVGP = thicker lines) using %Correct(t) = λ(1- e^−β(t-δ))+50%, where lambda (λ) is the level of asymptotic performance, beta (β) is the rate at which accuracy grows as a function of time, and delta (δ) is the intercept or the time at which accuracy rises above chance levels [35] . In both the DDM and the neural model, faster accumulation of information predicts greater rate (β) value in the VGPs. This prediction was confirmed (visual motion: VGP: 8.1 +/− 1.1, NVGP: 4.8 +/− 1.1 - F(1,21) = 4.6, p = .045, p-eta² = .19; auditory tone: VGP: 24.5 +/− 3.9, NVGP: 12.9 +/− 3.8 - F(1,21) = 4.5, p = .044, p-eta² = .18) demonstrating again a greater rate of sensory integration in VGP.