Simulating developmental diversity: Impact of neural stochasticity on atypical flexibility and hierarchy

Abstract

Introduction: Investigating the pathological mechanisms of developmental disorders is a challenge because the symptoms are a result of complex and dynamic factors such as neural networks, cognitive behavior, environment, and developmental learning. Recently, computational methods have started to provide a unified framework for understanding developmental disorders, enabling us to describe the interactions among those multiple factors underlying symptoms. However, this approach is still limited because most studies to date have focused on cross-sectional task performance and lacked the perspectives of developmental learning. Here, we proposed a new research method for understanding the mechanisms of the acquisition and its failures in hierarchical Bayesian representations using a state-of-the-art computational model, referred to as in silico neurodevelopment framework for atypical representation learning.

Methods: Simple simulation experiments were conducted using the proposed framework to examine whether manipulating the neural stochasticity and noise levels in external environments during the learning process can lead to the altered acquisition of hierarchical Bayesian representation and reduced flexibility.

Results: Networks with normal neural stochasticity acquired hierarchical representations that reflected the underlying probabilistic structures in the environment, including higher-order representation, and exhibited good behavioral and cognitive flexibility. When the neural stochasticity was high during learning, top-down generation using higher-order representation became atypical, although the flexibility did not differ from that of the normal stochasticity settings. However, when the neural stochasticity was low in the learning process, the networks demonstrated reduced flexibility and altered hierarchical representation. Notably, this altered acquisition of higher-order representation and flexibility was ameliorated by increasing the level of noises in external stimuli.

Discussion: These results demonstrated that the proposed method assists in modeling developmental disorders by bridging between multiple factors, such as the inherent characteristics of neural dynamics, acquisitions of hierarchical representation, flexible behavior, and external environment.

Keywords: Bayesian brain; autism spectrum disorder (ASD); computational psychiatry; flexibility; neural network; neural noise; predictive coding; representation learning.

Figure 1

The scheme of “‘in silico neurodevelopment framework for atypical representation learning” proposed in this study. (A) The agent modeled by the hierarchical Bayesian neural network model (PV-RNN) must learn the hierarchical and probabilistic structure hidden in the observations in the developmental learning process. (B) The inherent characteristics of neural dynamics and environmental factors are simulated as experimental manipulation to understand divergence in the developmental process. z_t and x_t represent latent and observed variables, respectively.

Figure 2

(A) The simulation experiments based on the proposed framework. In the experiments, as behavioral and cognitive task, flexibility task was used. To understand atypical developmental process, (a) the stochasticity in neural dynamics of lower layer, (b) noise level of observation signal was manipulated. dist. represents distribution. (B) An example of training sequences in the simulation experiments. These sequences repeated state transitions to LEFT or RIGHT (“target state”). The probability that the transition from HOME to LEFT is likely to occur is determined by “transition bias.” Transition bias was set to 0.76 (LEFT-biased sequences), and “signal noise” was set to low (stable environment condition) in the presented sequence. In the test phase, the transition bias switched at the middle point in the sequence to quantify flexibility.

Figure 3

The graphical representation of PV-RNN architecture (left). PV-RNN constructs hierarchical generation process in which the higher layer has larger time constant (slow neural dynamics) while the lower layer has smaller time constant (fast neural dynamics), as shown in (bottom-right). In (top-right), the inference process of z_t is illustrated. The right superscripts of symbols (i.e., p and q) are used to distinguish prior and approximate posterior distributions. The inference of posterior latent units $z_t^q$ is performed by propagating the errors in the reverse direction of arrows and updating the adaptive variables a_t. This figure is simplified to improve readability, and detailed and accurate information of PV-RNN is shown in Supplementary Methods 1.3, 1.4.

Figure 4

An example of flexibility tasks under normal meta-prior condition. In the top of figure, RNN generations and test sequence are plotted on two-dimensional plane (left) and along time axis (right). The unit0 and unit1, unit2 and unit3, and unit4 and unit5 reflect the lower layer, middle layer, and higher layer, respectively. The latent units coding mean parameters of Gaussian distributions are plotted in the figure rather than z_t itself. In the figure, only the 128 steps before and after switching of the transition bias are plotted.

Figure 5

(A) Example of flexibility tasks under strong meta-prior condition. The arrows and arrowheads represent perseveration errors and timing mismatches, respectively. The latent units coding mean parameters of Gaussian distributions were plotted in figure rather than z_t itself. (B) Example of flexibility tasks under weak meta-prior condition. The range of color plot adjusted to activities of higher latent units although the max and min values in lower- and middle-units surpassed the ranges of those plotted. In the figures (A, B), only the 128 steps before and after switching of the transition bias are plotted.

Figure 6

The quantitative evaluation about behavioral flexibility (A) and cognitive flexibility (B). MP represents meta-prior.

Figure 7

The results of latent space traversal under normal meta-prior condition. The properties of generated sequences (y-axis) changed depending on fixed activation values (x-axis) of one particular unit. Changes in the number of steps staying with HOME states (A) and the numbers of transition to LEFT states (B) were plotted. (C) Changes in the variances of generations were plotted when activities of units inferring variances of latent units were fixed. Irrelevant lines are plotted in a pale color to improve readability.

Figure 8

The results of latent space traversal under strong (A, B) and weak meta-prior condition (C, D). Irrelevant lines are plotted in a pale color to improve readability.

Figure 9

(A–D) The generative hierarchy under each condition. MP represents meta-prior.

Figure 10

(A) The training sequence in noisy environment condition in which transition bias was set to 0.76 (LEFT-biased sequences). (B–E) The interaction effect between the environment and meta-prior. The results of statistical test were showed under only strong meta-prior condition.

Figure 11

The results of simulation experiments were graphically summarized. The networks with normal neural stochasticity were able to acquire hierarchical representations, including higher-order representation, and exhibited good behavioral and cognitive flexibility. When the neural stochasticity was high in the learning process, top-down generation using higher-order representation (i.e., generative hierarchy) was impaired, although the flexibility did not differ from that of the normal settings. On the other hand, when the neural stochasticity was low in the learning process, the networks demonstrated reduced flexibility and abnormal hierarchical representation. However, this altered acquisition of higher-order representation and flexibility was ameliorated by increasing the level of noises in external stimuli.