Learning and replaying spatiotemporal sequences: A replication study

Jette Oberländer, Younes Bouhadjar, Abigail MorrisonFrontiers in Integrative Neuroscience 2022 · DOI

TL;DR

This paper successfully replicates a biologically plausible spiking neural network model for learning and replaying spatiotemporal sequences. By re-implementing the model in the NEST simulator, the authors not only confirm the original findings but also identify and correct several implementation errors and inconsistencies, making the model more robust and accessible for future research.

What Problem Does This Paper Address?

The brain's ability to learn and replay sequences of events is fundamental to many cognitive functions, from motor control and language processing to planning and reasoning. While artificial neural networks have shown success in sequence learning, many models lack biological plausibility or have limited functionality. A previous study by Maes et al. (2020) proposed a promising spiking recurrent neural network (RNN) model that could learn and replay spatiotemporal sequences using biologically inspired components. However, for such models to be useful for further scientific inquiry, their replicability and accuracy are crucial. This paper addresses the challenge of verifying and refining the Maes et al. (2020) model by undertaking a rigorous replication study. The authors aim to identify any discrepancies between the original paper's description and its implementation, and to provide a corrected, independently implemented version.

Methods

The core of this study is a replication of the Maes et al. (2020) model. The authors re-implemented the entire model using the open-source neural simulator NEST and its modeling language NESTML. This approach allowed them to:

  • Describe the model in detail: The paper provides a thorough explanation of the model's architecture, including the recurrent neural network (RNN) for temporal processing and the read-out layer for spatial information.
  • Identify and correct errors: Through careful examination of the original paper and its reference implementation, the authors discovered and addressed several issues, including:
  • Implementation errors: Mistakes in the original code related to numerical integration methods (e.g., Forward Euler for STDP), indexing, and synaptic conductance calculations.
  • Parameter inconsistencies: Discrepancies between parameters stated in the text, tables, and code, as well as missing or unspecified parameters (e.g., synaptic delays).
  • Absent or inconsistent information: Missing details about learning protocols and synaptic normalization, and differing definitions of low-pass filtering in the text versus the code.
  • Physical unit mismatches: Errors in the units of various equations, requiring the introduction of correction constants.
  • Use a domain-specific simulator: Employing NEST and NESTML provided advantages such as integrated concepts for inhibition and excitation, built-in unit checking, and a more robust simulation environment compared to general-purpose programming platforms.

The replication process involved two main phases: first, training the RNN to develop sequential dynamics (a "neural clock"), and second, training the read-out layer to associate specific temporal patterns with spatial elements of a sequence.

Datasets

Not discussed in the paper. The model is trained on simulated data generated by the network itself based on specific learning protocols.

Key Insights and Ideas

  • Replication is crucial for scientific progress: This study highlights how essential independent replication is for identifying subtle but critical errors in model implementations, which can significantly impact results and hinder further research.
  • Spiking RNNs can learn complex sequences: The core finding of the original work, that a spiking RNN can learn and replay spatiotemporal sequences, is confirmed. This demonstrates the potential of biologically plausible neural models for complex cognitive tasks.
  • The "neural clock" is a key mechanism: The model's ability to discretize time through sequential activation of neuronal clusters (forming a "neural clock") is a central innovation, providing a temporal backbone for sequence learning.
  • Decomposition of learning: The model effectively separates temporal learning (in the RNN) from spatial learning (in the read-out layer), allowing for flexible mapping of sequence elements to time intervals.
  • Challenges in biological plausibility: While the model uses biologically inspired components, the "supervised" nature of the training input for the clock mechanism raises questions about how such a process might be implemented biologically without explicit supervision.

Key Results

  • The authors successfully replicated the main findings of Maes et al. (2020), confirming that the recurrent network develops precise sequential dynamics and that the read-out layer can learn to replay higher-order spatiotemporal sequences.
  • Minor quantitative deviations were observed, attributed to differences in simulation frameworks (NEST's event-based approach vs. time-based in the original implementation) and random number generators.
  • The paper identified and corrected several implementation errors, parameter inconsistencies, and physical unit mismatches in the original model, leading to a more accurate and maintainable implementation.
  • The re-implemented model is now better positioned to serve as a foundation for future research into spatiotemporal sequence learning.

Limitations and Future Work

  • Biological plausibility of training: The authors note that the training protocol for the RNN's "neural clock" relies on a sequential external input that mimics supervision, and the biological mechanisms supporting such a signal are not yet identified.
  • Limited sequence duration and temporal resolution: The current clock period (~470 ms) and the cluster activation duration (~15 ms) limit the length and fine-grained temporal details of sequences that can be learned. Extending these capabilities would likely require significant network redesign or parameter tuning.
  • Parallel sequence replay: While the model can learn sequences in parallel, replaying them separately requires an external inhibitory signal that is difficult to explain biologically.
  • Alphabet token redundancy: Learning parallel sequences with identical elements requires separate read-out neurons for each instance, even if they represent the same alphabetic token.

Future work could focus on developing more biologically plausible training mechanisms for the temporal backbone, extending the model's capacity for longer and more temporally precise sequences, and exploring more efficient mechanisms for parallel sequence replay.

Summary updated

Abstract

Learning and replaying spatiotemporal sequences are fundamental computations performed by the brain and specifically the neocortex. These features are critical for a wide variety of cognitive functions, including sensory perception and the execution of motor and language skills. Although several computational models demonstrate this capability, many are either hard to reconcile with biological findings or have limited functionality. To address this gap, a recent study proposed a biologically plausible model based on a spiking recurrent neural network supplemented with read-out neurons. After learning, the recurrent network develops precise switching dynamics by successively activating and deactivating small groups of neurons. The read-out neurons are trained to respond to particular groups and can thereby reproduce the learned sequence. For the model to serve as the basis for further research, it is important to determine its replicability. In this Brief Report, we give a detailed description of the model and identify missing details, inconsistencies or errors in or between the original paper and its reference implementation. We re-implement the full model in the neural simulator NEST in conjunction with the NESTML modeling language and confirm the main findings of the original work.
BibTeX
@inproceedings{oberlander2022learning,
  author = {Jette Oberländer and Younes Bouhadjar and Abigail Morrison},
  title = {{Learning and replaying spatiotemporal sequences: A replication study}},
  booktitle = {Frontiers in Integrative Neuroscience},
  year = {2022},
  doi = {10.3389/fnint.2022.974177},
}

Loading related content...

Figure