Deficits in Prediction Ability Trigger Asymmetries in Behavior and Internal Representation

Anja Philippsen, Yukie NagaiFrontiers in Psychiatry 2020 · DOI

TL;DR

This paper uses recurrent neural networks (RNNs) to model how imbalances in predictive coding—the brain's ability to predict sensory input—can lead to different behavioral and internal representation outcomes. They found that both too much and too little reliance on predictions can impair a network's internal representations, even if its basic task performance remains intact, suggesting a potential computational basis for the diverse symptoms seen in conditions like autism spectrum disorder (ASD).

What Problem Does This Paper Address?

The brain constantly tries to predict incoming sensory information based on prior experiences. This "predictive coding" framework suggests that mismatches between predictions and actual sensory input (prediction errors) are key to learning and perception. Aberrations in this process, such as relying too heavily on predictions or being overly sensitive to sensory input, are hypothesized to underlie various developmental disorders, particularly autism spectrum disorder (ASD). However, understanding how these imbalances manifest computationally and how they relate to the observed heterogeneity in ASD symptoms remains a challenge. This paper aims to explore this by computationally modeling how parametric changes in predictive coding mechanisms affect both a model's observable behavior and its internal representations.

Methods

The researchers employed a stochastic continuous-time recurrent neural network (S-CTRNN), a type of RNN well-suited for modeling time-series data and internal representations. They trained these networks to perform a specific task: drawing simple 2D trajectories (ellipses and "eight" shapes).

The core of their approach involved manipulating two key parameters within the S-CTRNN model:

  1. External Contribution Parameter ($X_{train}$): This parameter controls how much the network relies on incoming sensory data versus its own internal predictions. A high $X_{train}$ means strong reliance on sensory input, while a low $X_{train}$ means strong reliance on predictions.
  2. Aberrant Precision Parameter ($K$): This parameter modifies how the network estimates the variance (or uncertainty) of sensory input. $K=0$ represents normal estimation, $K>0$ represents overestimation of variance (leading to less weight given to prediction errors), and $K<0$ represents underestimation (leading to more weight given to prediction errors, potentially causing "overlearning").

Crucially, these parameters were modified during the training process, allowing the researchers to observe their impact on the development of the network's internal representations, not just its final performance. They evaluated networks based on both their behavioral output (how well they reproduced the trained trajectories) and the quality of their internal representations (analyzed using principal component analysis of the context layer activations).

Datasets

The study used a synthetically generated dataset consisting of eight different 2D trajectories: four ellipses and four "eight" shapes, each repeated three times. Gaussian noise was added to these trajectories to simulate environmental uncertainty. This dataset was designed to be simple enough for analysis but complex enough to require the network to learn and represent different features (shape, position, noise level).

Key Insights and Ideas

  • Behavior vs. Internal Representation Asymmetry: The most striking finding is that a network's ability to perform a trained task (behavioral output) doesn't always reflect the quality of its internal representations. A network can perform well on the training data even if its internal "understanding" of the patterns is poor.
  • U-Shaped Impairment: Both too little ($X_{train} \approx 0$) and too much ($X_{train} \approx 1$) reliance on sensory input (i.e., extreme $X_{train}$ values) led to impaired internal representations and generalization capabilities. This suggests that a "sweet spot" of prediction reliance is crucial for healthy cognitive development.
  • Aberrant Precision's Nuance: While over- or underestimating sensory noise ($K \neq 0$) didn't significantly impact task performance, it did differently affect internal representations. Negative $K$ (underestimation) led to better separation of internal representations compared to positive $K$ (overestimation).
  • Potential for ASD Subtypes: The U-shaped impairment suggests that ASD, which is often linked to prediction deficits, might not be a single entity. Instead, different subtypes of ASD could arise from either too weak or too strong a reliance on predictions, leading to diverse symptom profiles.
  • Internal Representation as a Predictor: The quality of internal representations (measured by how well different trajectories were separated in the network's internal state space) was a better predictor of generalization ability to novel tasks than raw behavioral performance on trained tasks.

Key Results

  • Task Performance: Networks generally achieved good task performance (low prediction error) across a wide range of $X_{train}$ and $K$ values, with only extreme $X_{train}$ values (close to 0) showing slower learning.
  • Internal Representation Quality: Optimal internal representation quality (indicated by good separation of trajectories in the internal state space) was observed for moderate $X_{train}$ values (around 0.4) and negative $K$ values. Extreme $X_{train}$ values (both low and high) led to impaired internal representations.
  • Generalization: Networks with better internal representation quality tended to generalize better to an untrained trajectory. This was particularly evident for moderate $X_{train}$ and negative $K$ values.
  • Correlation Disconnect: There was a weak correlation between behavioral performance and internal representation quality, highlighting that good performance on trained tasks doesn't guarantee good internal understanding or generalization.

Limitations and Future Work

  • Simplified Model: The study uses a relatively simple RNN model and a simplified task. The authors acknowledge that future work should explore more complex network architectures and tasks to bridge the gap between computational models and real-world cognitive phenomena.
  • Fixed Parameters: The model assumes fixed parameter values during training. In reality, individuals can adapt their reliance on predictions based on context. Future research should investigate how this adaptability might differ in ASD and how it influences learning.
  • Abstract Representation: The study focuses on abstract representations. Further work is needed to connect these findings to specific neural mechanisms in the brain.
  • ASD Subtypes: The paper proposes that different ASD subtypes might arise from extreme ends of the prediction reliance spectrum. This is a hypothesis that requires further experimental validation.

The authors suggest that future research should focus on developing more sophisticated models, investigating adaptive parameter settings, and conducting experimental studies to test these hypotheses, particularly in the context of developmental disorders like ASD.

Summary updated

Abstract

Predictive coding is an emerging theoretical framework for explaining human perception and behavior. The proposed underlying mechanism is that signals encoding sensory information are integrated with signals representing the brain's prior prediction. Imbalance or aberrant precision of the two signals has been suggested as a potential cause for developmental disorders. Computational models may help to understand how such aberrant tendencies in prediction affect development and behavior. In this study, we used a computational approach to test the hypothesis that parametric modifications of prediction ability generate a spectrum of network representations that might reflect the spectrum from typical development to potential disorders. Specifically, we trained recurrent neural networks to draw simple figure trajectories, and found that altering reliance on sensory and prior signals during learning affected the networks' performance and the emergent internal representation. Specifically, both overly strong or weak reliance on predictions impaired network representations, but drawing performance did not always reflect this impairment. Thus, aberrant predictive coding causes asymmetries in behavioral output and internal representations. We discuss the findings in the context of autism spectrum disorder, where we hypothesize that too weak or too strong a reliance on predictions may be the cause of the large diversity of symptoms associated with this disorder.
BibTeX
@inproceedings{philippsen2020deficits,
  author = {Anja Philippsen and Yukie Nagai},
  title = {{Deficits in Prediction Ability Trigger Asymmetries in Behavior and Internal Representation}},
  booktitle = {Frontiers in Psychiatry},
  year = {2020},
  doi = {10.3389/fpsyt.2020.564415},
}
Figure