Teaching Recurrent Neural Networks to Modify Chaotic Memories by Example

Jason Z. Kim, Zhixin Lu, Erfan Nozari, George J. Pappas, Danielle S. BassettarXiv 2020 · OpenReview

TL;DR

This paper shows that recurrent neural networks (RNNs) can learn to manipulate complex, chaotic data representations by simply observing examples. By training an RNN on variations of the Lorenz chaotic system, it learns to perform translations, transformations, and even predict bifurcations of this data, demonstrating a powerful new mechanism for how neural systems might learn to process and modify information.

What Problem Does This Paper Address?

While computers excel at structured data processing, biological systems handle complex, dynamic, and often unstructured sensory input with remarkable flexibility. A key question in neuroscience and AI is how neural systems learn to represent and manipulate this complex information. Specifically, how do they learn to modify their internal representations of dynamic processes, like those found in sensory streams or motor control? Existing models often require pre-programmed network structures or separate components for memory and computation, failing to capture the integrated, learned nature of biological information processing. This paper aims to bridge this gap by demonstrating that a single, unified RNN can learn to modify its internal representations of chaotic systems through example-based learning, mirroring aspects of biological learning.

Methods

The core of the methodology lies in using a reservoir computing framework. This involves a fixed, randomly connected recurrent neural network (the "reservoir") where only a simple linear readout layer is trained.

  1. Lorenz System as Data: The chaotic Lorenz system is used as the complex data source. Its time series are manipulated (translated, linearly transformed, or altered to induce bifurcations) to create training examples.
  2. Training with Examples: The RNN is driven by these manipulated Lorenz time series. The training objective is to learn a readout matrix (W) that maps the reservoir's internal states to the input Lorenz data.
  3. Feedback Loop: Once trained, the output of the readout layer is fed back as the input to the reservoir, creating an autonomous system. This allows the reservoir to evolve its internal states without external input, ideally mimicking the learned transformations.
  4. Control Signal: A crucial element is the introduction of a control signal (c). By varying this signal, the trained RNN can interpolate and extrapolate the learned transformations (translations, stretches, shears) and even predict bifurcations of the Lorenz attractor, far beyond the specific examples it was trained on.
  5. Quadratic Reservoir Dynamics: For theoretical analysis, the authors derive the reservoir's dynamics in a "weakly nonlinear" or quadratic regime. This allows for a more tractable mathematical understanding of how the control signal influences the reservoir's internal states and learned operations.

Datasets

The primary "dataset" used is generated from the Lorenz chaotic system. The authors create variations of the Lorenz attractor's time series by:

  • Translations: Shifting the attractor along one or more of its spatial dimensions (x1, x2, x3).
  • Linear Transformations: Applying stretch and shear operations to the attractor.
  • Bifurcations: Modifying parameters of the Lorenz system to induce a qualitative change in its behavior (e.g., from stable fixed points to a chaotic manifold).

These generated time series serve as both training data and the basis for evaluating the RNN's ability to generalize and extrapolate learned operations.

Key Insights and Ideas

  • Learning Complex Dynamics from Examples: The most striking insight is that RNNs can learn to manipulate complex, chaotic representations (like the Lorenz attractor) by simply observing examples of transformations. This suggests a powerful, generalizable learning mechanism that doesn't require explicit programming of the transformations.
  • Unified Representation and Computation: The reservoir computing framework allows a single network to both represent the complex dynamics and learn to compute transformations on that representation, mirroring how biological systems might integrate these functions.
  • Extrapolation Beyond Training Data: The RNN's ability to interpolate and extrapolate translations and transformations over a much wider range of the control signal than it was trained on is remarkable. This suggests the learned operations are robust and generalizable.
  • Inferring Bifurcations: The RNN's capacity to predict the behavior after a bifurcation, having only been trained on data before the bifurcation, is a significant finding. This implies the network is not just memorizing patterns but inferring underlying structural changes.
  • Control Signal as a "Dial": The control signal (c) acts like a dial, allowing continuous manipulation of the learned representation. This provides a mechanism for fine-grained control over the RNN's internal state.

Key Results

  • Successful Translation: The RNN learned to translate its representation of the Lorenz attractor along arbitrary directions (x1, x2, x3) and could extrapolate these translations far beyond the trained range of the control signal.
  • Successful Transformation: The RNN learned to apply stretch and shear transformations to the Lorenz representation, again demonstrating interpolation and extrapolation capabilities.
  • Bifurcation Prediction: Trained on stable Lorenz trajectories, the RNN could accurately predict the chaotic manifold that emerges after a bifurcation, even without direct training examples of the post-bifurcation state.
  • Simultaneous Learning: The framework demonstrated the ability to learn multiple operations (e.g., translations in different directions) simultaneously by using multiple control signals.
  • Mechanism of Learning: The paper provides a theoretical framework (quadratic reservoir dynamics) to explain how the control signal influences the reservoir's internal dynamics to achieve these learned operations, relating it to scaling and shifting.

Limitations and Future Work

  • Mechanism of Connectivity Stabilization: The authors acknowledge a lack of clear understanding regarding how the network's internal connectivity ultimately stabilizes the chaotic manifold. While progress has been made in controlling chaos, a guarantee for successful training across all parameters is missing.
  • Training Example Spacing: The exact spacing of training examples required for learning linear relationships between state differentials and control parameters is not fully specified.
  • Quadratic Approximation: The theoretical analysis relies on a quadratic approximation of reservoir dynamics, which is accurate within a certain range. Future work could explore higher-order approximations or different regimes.
  • Generalizability to Other Systems: While the Lorenz system is a rich example, future work could explore the applicability of this framework to other complex dynamical systems, including real-world time-series data like speech or music.
  • Inferring Higher-Order Structure: The ability to infer bifurcations from pre-bifurcation data suggests the network is learning higher-order nonlinear structure. Understanding this inferential capability more deeply is a key future direction.
  • Neural Mass Models: The results obtained in the quadratic regime are expected to hold for common neural mass models, suggesting a unifying framework for learning and computation in dynamical neural models.

Summary updated

Abstract

The ability to store and manipulate information is a hallmark of computational systems. Whereas computers are carefully engineered to represent and perform mathematical operations on structured data, neurobiological systems perform analogous functions despite flexible organization and unstructured sensory input. Recent efforts have made progress in modeling the representation and recall of information in neural systems. However, precisely how neural systems learn to modify these representations remains far from understood. Here we demonstrate that a recurrent neural network (RNN) can learn to modify its representation of complex information using only examples, and we explain the associated learning mechanism with new theory. Specifically, we drive an RNN with examples of translated, linearly transformed, or pre-bifurcated time series from a chaotic Lorenz system, alongside an additional control signal that changes value for each example. By training the network to replicate the Lorenz inputs, it learns to autonomously evolve about a Lorenz-shaped manifold. Additionally, it learns to continuously interpolate and extrapolate the translation, transformation, and bifurcation of this representation far beyond the training data by changing the control signal. Finally, we provide a mechanism for how these computations are learned, and demonstrate that a single network can simultaneously learn multiple computations. Together, our results provide a simple but powerful mechanism by which an RNN can learn to manipulate internal representations of complex information, allowing for the principled study and precise design of RNNs.
BibTeX
@misc{kim2020teaching,
  author = {Jason Z. Kim and Zhixin Lu and Erfan Nozari and George J. Pappas and Danielle S. Bassett},
  title = {{Teaching Recurrent Neural Networks to Modify Chaotic Memories by Example}},
  year = {2020},
  eprint = {2005.01186},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2005.01186},
}
Figure