Evolving reservoir computers reveals bidirectional coupling between predictive power and emergent dynamics

Hanna M. Tolle, Andrea I. Luppi, Anil K. Seth, Pedro A. M. MedianoarXiv 2024 · DOI · arXiv

Abstract

Biological neural networks can perform complex computations to predict their environment, far above the limited predictive capabilities of individual neurons. While conventional approaches to understanding these computations often focus on isolating the contributions of single neurons, here we argue that a deeper understanding requires considering emergent dynamics—dynamics that make the whole system "more than the sum of its parts". Specifically, we examine the relationship between prediction performance and emergence by leveraging recent quantitative metrics of emergence, derived from Partial Information Decomposition, and by modelling the prediction of environmental dynamics in a bio-inspired computational framework known as reservoir computing. Notably, we reveal a bidirectional coupling between prediction performance and emergence, which generalises across task environments and reservoir network topologies, and is recapitulated by three key results: 1) Optimising hyperparameters for performance enhances emergent dynamics, and vice versa; 2) Emergent dynamics represent a near sufficient criterion for prediction success in all task environments, and an almost necessary criterion in most environments; 3) Training reservoir computers on larger datasets results in stronger emergent dynamics, which contain task-relevant information crucial for performance. Overall, our study points to a pivotal role of emergence in facilitating environmental predictions in a bio-inspired computational architecture.

Introduction: Beyond Reductionism in Neural Prediction

In neuroscience and artificial intelligence, reductionist approaches often attempt to decode complex system behavior by analyzing individual components, such as single neurons. However, biological brains execute "non-trivial" computations that are not apparent when neurons are viewed in isolation. This "binding problem" highlights the challenge of understanding how distributed information is integrated into a unified whole.

This research addresses the concept of causal emergence. The authors investigate whether emergence—defined as predictive information present in the collective system that is absent in its individual parts—is a functional requirement for computational tasks or merely a mathematical byproduct. By utilizing Reservoir Computing (RC) as a model for biological networks, the study explores whether a system's capacity to forecast its environment is fundamentally linked to these synergistic, emergent dynamics.

Methodology: Reservoir Computing and Information Decomposition

The researchers utilized a bio-inspired framework known as Reservoir Computing. In this architecture, a fixed, non-linear recurrent neural network (the "reservoir") processes temporal inputs. Only a simple linear "readout" layer is trained to map the reservoir's internal states to the desired output.

The Reservoir Computing architecture (A) and the conceptual framework of emergence (B), where the whole system possesses properties not found in its individual components.

To quantify emergence, the study employed a metric derived from Partial Information Decomposition (PID) called ψ (psi). This metric measures "self-predictive" information within the system's forecast that cannot be attributed to individual neurons. Using Evolutionary Algorithms, populations of reservoirs were evolved by optimizing hyperparameters (such as spectral radius and connection density) for two distinct objectives: maximizing predictive performance (minimizing loss) and maximizing causal emergence (ψ).

The models were tested on six 3-dimensional chaotic dynamical systems, including the Lorenz Attractor and various Sprott systems. Additionally, the researchers incorporated structural data from the Human Connectome Project (HCP) to create bio-inspired reservoir topologies, comparing them against standard random networks.

A primary finding of the study is the bidirectional coupling between predictive accuracy and emergence. When networks were evolved to minimize prediction loss, they naturally developed higher levels of emergence. Conversely, optimizing specifically for emergence (ψ) resulted in a spontaneous improvement in predictive performance.

Evolutionary optimization for maximum emergence (ψ) results in a corresponding decrease in prediction loss over generations. The scatter plot confirms a strong negative correlation between emergence and loss across the population.

Key results include:

  • Sufficiency and Necessity: Emergence was found to be a near-sufficient criterion for success; networks exhibiting high emergent dynamics were almost universally effective predictors ($P(S|E) > 0.8$). In several complex environments, emergence was also a necessary condition for achieving high-fidelity forecasts.
  • Data-Driven Synergy: Increasing the size of the training dataset correlated with a higher probability of emergence. This suggests that mastering complex environmental patterns necessitates a shift from redundant, individual-neuron representations toward synergistic, collective dynamics.
  • General Principles of Dynamics: Interestingly, human connectome-based reservoirs did not significantly outperform random networks in these tasks. This indicates that the link between emergence and prediction is a general property of recurrent dynamics rather than being strictly dependent on specific biological wiring at the scales tested.

Discussion and Conclusion

The study demonstrates that emergence is not just a philosophical concept but a quantifiable engine of neural prediction. Each neuron in the reservoir acts as a "weak learner," and emergence provides the mechanism to unite these signals into a powerful predictive system.

While the results are robust, the authors note that the ψ metric serves as a lower bound for emergence and may not capture all higher-order features. Future research is needed to establish direct causality through interventionist studies and to explore whether optimizing for emergence enhances transfer learning in unfamiliar environments. Scaling these models to larger, more specialized tasks will further clarify the advantages of bio-inspired topologies in fostering emergent computation.

Summary updated

Abstract

Biological neural networks can perform complex computations to predict their environment, far above the limited predictive capabilities of individual neurons. While conventional approaches to understanding these computations often focus on isolating the contributions of single neurons, here we argue that a deeper understanding requires considering emergent dynamics - dynamics that make the whole system "more than the sum of its parts". Specifically, we examine the relationship between prediction performance and emergence by leveraging recent quantitative metrics of emergence, derived from Partial Information Decomposition, and by modelling the prediction of environmental dynamics in a bio-inspired computational framework known as reservoir computing. Notably, we reveal a bidirectional coupling between prediction performance and emergence, which generalises across task environments and reservoir network topologies, and is recapitulated by three key results: 1) Optimising hyperparameters for performance enhances emergent dynamics, and vice versa; 2) Emergent dynamics represent a near sufficient criterion for prediction success in all task environments, and an almost necessary criterion in most environments; 3) Training reservoir computers on larger datasets results in stronger emergent dynamics, which contain task-relevant information crucial for performance. Overall, our study points to a pivotal role of emergence in facilitating environmental predictions in a bio-inspired computational architecture.
BibTeX
@misc{tolle2024evolving,
  author = {Hanna M. Tolle and Andrea I. Luppi and Anil K. Seth and Pedro A. M. Mediano},
  title = {{Evolving reservoir computers reveals bidirectional coupling between predictive power and emergent dynamics}},
  year = {2024},
  eprint = {2406.19201},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2406.19201},
}
Figure