Towards a more general understanding of the algorithmic utility of recurrent connections

Brett W. Larsen, Shaul DruckmannPLOS Computational Biology 2022 · OpenReview

TL;DR

This paper investigates the algorithmic advantages of recurrent neural networks (RNNs) over feedforward networks for tasks requiring global information processing. By developing and analyzing a "tag propagation" framework, the authors demonstrate that RNNs excel at tasks involving spatial and temporal connectivity, even outperforming state-of-the-art feedforward architectures in certain regimes. This work suggests that recurrent connections are crucial for solving complex problems that mirror biological decision-making and offers a path towards more powerful AI.

What Problem Does This Paper Address?

While modern artificial neural networks (ANNs) have achieved remarkable success, they predominantly rely on feedforward architectures. This stands in stark contrast to biological brains, which extensively utilize lateral and recurrent connections. This discrepancy raises a fundamental question: does the prevalence of recurrent connections in biological systems contribute to their superior computational power, and if so, what specific computational advantages do they offer? Understanding these advantages is crucial for bridging the gap between artificial and natural intelligence and for developing more sophisticated AI systems. The authors aim to provide a more general understanding of the algorithmic utility of recurrent connections, particularly for tasks that inherently require processing global information.

Methods

The core of the paper's methodology revolves around a "tag propagation" algorithm, inspired by early AI research and biological plausibility. This algorithm efficiently solves connectivity problems by sequentially propagating a "tag" (e.g., a "connected" label) through a network.

Key methodological aspects include:

  • Edge-Connected Pixel Task: A novel benchmark designed to precisely quantify network performance on a spatial connectivity task. This task involves identifying pixels connected to the image edge via paths of "on" pixels.
  • Large-Scale Architecture Search: The authors systematically trained and evaluated a wide range of feedforward, recurrent, and hybrid network architectures on the edge-connected pixel task using stochastic gradient descent (SGD). This involved extensive hyperparameter optimization and training from multiple random initializations to ensure robust findings.
  • Hybrid Architectures: The paper explores combining feedforward and recurrent components. A "time-dependent switching network" was designed to leverage the speed of feedforward networks for early predictions and the accuracy of recurrent networks for longer computations.
  • Generalized Tag Propagation: The tag propagation concept was extended to handle multiple interacting tags, enabling the solution of more complex problems.
  • Competitive Foraging Task: A new, abstract decision-making task was introduced to demonstrate the utility of generalized tag propagation for temporal prediction and decision-making in dynamic environments. This task involves an agent deciding whether to pursue food based on its own and competitors' potential reachability over time.
  • Analytical Solutions: For both the edge-connected pixel task and the competitive foraging task, analytical solutions based on tag propagation were derived to serve as benchmarks and provide theoretical grounding.

Datasets

  • Edge-Connected Pixel Task Dataset: The authors generated custom datasets for the edge-connected pixel task. These datasets consisted of binary images with controlled distributions of path lengths and distractor pixels to ensure challenging examples for network training and evaluation.
  • Competitive Foraging Task Dataset: Custom datasets were generated for the competitive foraging task, featuring random environments with barriers, and specified locations for the animal, competitors, and food.

Key Insights and Ideas

  • Recurrence for Global Computations: The study strongly supports the idea that recurrent connections are fundamentally advantageous for tasks requiring global information processing, such as determining connectivity across an entire image or predicting future states.
  • Tag Propagation as a General Framework: The "tag propagation" algorithm, and its generalization to multiple tags, proves to be a powerful and versatile framework for solving a range of connectivity and prediction problems. It elegantly transforms global computations into local ones.
  • Hybrid Architectures Offer Robustness: Combining feedforward and recurrent networks into hybrid architectures allows for a flexible trade-off between computation time and accuracy. This is particularly beneficial when computational resources are limited or vary dynamically.
  • Time vs. Depth Trade-off: Feedforward networks excel when computation time is limited (few layers/steps), while recurrent networks become superior when ample time is available (more layers/steps). Hybrid models bridge this gap.
  • Inductive Bias Matters: Even with significant architectural constraints (e.g., masked weights enforcing locality), networks can still learn effective solutions, suggesting that appropriate inductive biases can guide learning towards efficient recurrent-like computations.
  • Biological Plausibility: The findings align with the observation of recurrent connections in biological brains and suggest their role in complex cognitive functions like decision-making and prediction.

Key Results

  • Edge-Connected Pixel Task:
    • Recurrent tag propagation networks achieved near-perfect accuracy when allowed sufficient time steps.
  • Deep feedforward networks performed better than RNNs when computational time was severely limited (few layers/steps).
  • Hybrid architectures significantly outperformed pure feedforward or recurrent networks across a range of computational time constraints. For instance, a hybrid model achieved 7.28% error on path pixels after 5 timesteps, compared to 63.52% for pure tag propagation.
  • Masked recurrent networks, incorporating locality constraints, also achieved performance comparable to tag propagation at higher layer numbers.
  • Competitive Foraging Task:
    • Trained recurrent networks accurately learned the multi-tag propagation for animal and competitor reachability.
  • The best trained recurrent decision network achieved 90.2% accuracy on the task.
  • Convolutional networks (CNNs and ResNet-20) were less efficient and, in some cases, less accurate than the recurrent tag propagation solution, despite having more parameters. For example, a ResNet-20 achieved 93.86% accuracy, but the tag propagation solution achieved 0% error.

Limitations and Future Work

  • Abstract Tasks: The tasks studied are highly abstracted. While they provide conceptual understanding, extending these findings to more complex, real-world scenarios with richer sensory inputs and motor outputs remains a future direction.
  • Unknown Locality Structures: The paper notes that in many real-world tasks, the precise locality structure is unknown. Future work could explore how to learn or infer these structures.
  • Interpretability of Complex Dynamics: While tag propagation offers some interpretability, understanding the dynamics of complex recurrent networks with many interacting tags remains challenging.
  • Biological Constraints: Future work could explore how biological constraints (e.g., genetic encoding, developmental shaping) might influence the implementation of recurrent architectures in the brain.
  • Experimental Paradigms: The authors suggest that experimental paradigms involving animals making multiple, distinct predictions about future states would be invaluable for testing these hypotheses and deriving further constraints.
  • Hierarchical Recurrence: Exploring explicit hierarchical organization of processing levels within generalized tag propagation frameworks is suggested as a promising avenue.

Summary updated

Abstract

Author summary In striking contrast to the majority of current-day artificial neural network research which primarily focuses on feedforward architectures, biological brains make extensive use of lateral and recurrent connections. This raises the possibility that this difference makes a fundamental contribution to the gap in computational power between real neural circuits and artificial neural networks. Thus, despite the difficulty of making effective comparisons between different network architectures, developing a more detailed understanding of the computational role played by such connections is a pressing challenge. Here, we leverage the computational capabilities of large-scale machine learning to robustly explore how differences in architectures affect a network’s ability to learn tasks that require propagation of global information. We first focus on the task of determining whether two pixels are connected in an image which has an elegant and efficient recurrent solution: propagate a connected label or tag along paths. Inspired by this solution, we show that it can be generalized in many ways, including propagating multiple tags at once and changing the computation performed on the result of the propagation. Strikingly, this simple expansion of the tag propagation network is sufficient to solve a crucial abstraction to temporal connectedness at the core of many decision-making problems, which we illustrate for an abstracted competitive foraging task. Our results shed light on the set of computational tasks that can be solved efficiently by recurrent computation and how these solutions may relate to the structure of neural activity.
BibTeX
@article{larsen2022towards,
  author = {Brett W. Larsen and Shaul Druckmann},
  title = {{Towards a more general understanding of the algorithmic utility of recurrent connections}},
  journal = {PLOS Computational Biology},
  year = {2022},
  url = {https://openreview.net/forum?id=L5R-6UsZJV},
}
Figure