Abstract:Storage and retrieval of sequences require memory that is sensitive to the temporal order of features. For example, in human language, words that are stored in long-term memory are retrieved based on the order of phonemes. It is currently unknown whether Hebbian learning supports the formation of memories that are structured in time. We investigated whether word-like memories can emerge in a network of neurons with dendritic structures. Dendrites provide neuronal processing memory on the order of 100 ms and have been implicated in structured memory formation. We compared a network of neurons with dendrites and two networks of point neurons that have previously been shown to acquire stable long-term memories and process sequential information. The networks were equipped with voltage-based, spike-timing dependent plasticity (STDP) and were homeostatically balanced with inhibitory STDP. In the learning phase, networks were exposed to phoneme sequences and word labels, which led to the formation of overlapping cell assemblies. In the retrieval phase, networks only received phoneme sequences as input, and we measured the firing activity of the corresponding word populations. The dendritic network correctly reactivated the word populations with a success rate of 80%, including words composed of the same phonemes in a different order. The networks of point neurons reactivated only words that contained phonemes that were unique to these words and confused words with shared phonemes (success rate below 20%). These results suggest that the slow timescale and non-linearity of dendritic depolarization allowed neurons to establish connections between neural groups that were sensitive to serial order. Inhibitory STDP prevented the potentiation of connections between unrelated neural populations during learning. During retrieval, it maintained the dendrites hyperpolarized and limited the reactivation of incorrect cell assemblies. Thus, the addition of dendrites enables the encoding of temporal relations into associative memories.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is: how to achieve time - structure - sensitive memory under biological constraints, especially for lexical memory in language processing. Specifically, the author explored whether dendritic structures can support the ability to form and re - activate sequential memories through Hebbian plasticity, which are crucial for distinguishing words with the same phonemes but different orders. ### Background of the Paper and Problem Definition 1. **Speech Perception and Long - Term Memory** - Speech perception involves integrating continuous auditory information into ordered memories. In human languages, the storage and retrieval of words are based on the order of phonemes. - Currently, it is still unclear whether Hebbian learning supports the formation of time - structured memories. 2. **Hebbian Learning and Cell Assemblies** - The Hebbian learning theory holds that lexical memories are stored in long - term memory in the form of cell assemblies, which are formed by repeated co - activation. - However, due to the large number of phoneme overlaps in the human lexicon (such as "cat" and "tack"), how to distinguish different lexical memories based on the order of phonemes is a challenge. 3. **Sequence Detection Problem** - Biological neural networks need to have the ability to distinguish among similar input sequences. This involves how to encode temporal relationships in neural networks so that the network can activate corresponding memories according to the order of input. ### Research Methods and Models 1. **Model Design** - The researchers constructed a neuronal network with dendritic structures (Tripod network) and compared it with two point - neuron networks. - The network uses voltage - dependent spike - timing - dependent plasticity (vSTDP) and inhibitory STDP (iSTDP) to balance network activity. 2. **Experimental Setup** - Learning phase: The network was simultaneously exposed to phoneme sequences and word labels, forming overlapping cell assemblies. - Retrieval phase: The network only received phoneme sequences as input and measured the firing activity of the corresponding word populations. ### Main Findings 1. **Performance of Dendritic Networks** - The dendritic network successfully re - activated word populations with a success rate of 80%, including words composed of the same phonemes but different orders. - This indicates that the slow time scale and nonlinear depolarization of dendrites allow neurons to establish order - sensitive connections. 2. **Performance of Point - Neuron Networks** - Point - neuron networks can only re - activate words containing unique phonemes and perform poorly when dealing with words sharing phonemes (success rate less than 20%). 3. **Mechanism Analysis** - Dendritic memory combines the input phoneme sequence through a long - lasting depolarized state (about 100 milliseconds), supporting the formation of sequential memories. - Inhibitory STDP prevents the connection enhancement between irrelevant neural populations and maintains the hyperpolarized state of dendrites during the retrieval phase, limiting the re - activation of incorrect cell assemblies. ### Conclusion This study demonstrates the crucial role of dendritic structures in supporting the formation and retrieval of time - structured memories. Through the combination of Hebbian plasticity and dendritic memory, neural networks can effectively encode and distinguish lexical memories with similar inputs but different orders. This finding provides a new perspective for understanding how the brain processes and stores time - sensitive information.

Dendrites support formation and reactivation of sequential memories through Hebbian plasticity

A Hierarchical Organized Memory Model Using Spiking Neurons

Differences in the consolidation by spontaneous and evoked ripples in the presence of active dendrites

Striatum Expresses Region-Specific Plasticity Consistent with Distinct Memory Abilities

Sequential Memory: a Putative Neural and Synaptic Dynamical Mechanism.

Emergence and maintenance of modularity in neural networks with Hebbian and anti-Hebbian inhibitory STDP

Different Propagation Speeds of Recalled Sequences in Plastic Spiking Neural Networks

Compartmentalized dendritic plasticity in the retrosplenial cortex integrates memories across time

Postsynaptic burst reactivation of hippocampal neurons enables associative plasticity of temporally discontiguous inputs

A hippocampus CA3 spiking neural network model for storage and retrieval of sequential memory

A non-Hebbian code for episodic memory

Stably maintained dendritic spines are associated with lifelong memories

Long-term memory stabilized by noise-induced rehearsal

Divergent recruitment of developmentally defined neuronal ensembles supports memory dynamics

A history of spike-timing-dependent plasticity

Coexistence of Multiple Types of Synaptic Plasticity in Individual Hippocampal CA1 Pyramidal Neurons

STDP-based Associative Memory Formation and Retrieval

Interplay of Long- and Short-term Synaptic Plasticity in a Spiking Network Model of Rat's Episodic Memory

Co-dependent excitatory and inhibitory plasticity accounts for quick, stable and long-lasting memories in biological networks

Inhibitory Plasticity Enhances Sequence Storage Capacity and Retrieval Robustness

Spatial representation of temporal information through spike timing dependent plasticity