# Sequence learning, prediction, and replay in networks of spiking neurons

**Authors:** Younes Bouhadjar, Dirk J. Wouters, Markus Diesmann, Tom Tetzlaff  
**arXiv:** [2111.03456](https://arxiv.org/abs/2111.03456)  
**Date:** 2021-11-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.03456v2/

## Key points

- Neocortex performs universal computations through sequence learning, prediction, and replay
- Hierarchical Temporal Memory (HTM) algorithm embodies these processes
- HTM algorithm relies on abstract discrete-time neuron and synapse dynamics
- Authors propose a continuous-time implementation of the temporal-memory (TM) component of HTM using spiking neurons with biophysically interpretable variables and parameters
- Proposed model learns high-order sequences using structural Hebbian synaptic plasticity mechanism and rate-based homeostatic control
- Model enables dynamic self-organization of narrow sequence-specific subnetworks
- Subnetworks facilitate propagation of sparse, synchronous activity for robust predictions and autonomous replay of learned sequences
- Continuous-time implementation strengthens link to biology and allows for better evaluation through electrophysiological and behavioral data analysis
- Advancement in bridging computational models with biological mechanisms for sequence learning and prediction in networks of spiking neurons.

## Summary

The neocortex is capable of performing universal computations through sequence learning, prediction, and replay. The Hierarchical Temporal Memory (HTM) algorithm is a computational model that embodies these processes. However, the HTM algorithm relies on abstract discrete-time neuron and synapse dynamics, which limits its biological plausibility. In this study, the authors propose a continuous-time implementation of the temporal-memory (TM) component of the HTM algorithm using a network of spiking neurons with biophysically interpretable variables and parameters. The proposed model learns high-order sequences by employing a structural Hebbian synaptic plasticity mechanism supplemented with rate-based homeostatic control. This combination, along with nonlinear dendritic input integration and local inhibitory feedback, enables dynamic self-organization of narrow sequence-specific subnetworks. These subnetworks facilitate the propagation of sparse, synchronous activity, leading to robust and context-specific predictions of future sequence elements as well as autonomous replay of previously learned sequences. By incorporating more biologically realistic features into the TM component of the HTM algorithm, this continuous-time implementation strengthens the link to biology. It allows for better evaluation of the TM hypothesis through electrophysiological and behavioral data analysis by providing a comprehensive exploration of its capabilities in learning sequences and generating predictions. Furthermore, it highlights its potential relevance to understanding neural processing in biological systems. Overall, this study presents an important advancement in bridging computational models with biological mechanisms by developing a more biologically plausible implementation for sequence learning and prediction in networks of spiking neurons.