[Paper Review] A Neuromorphic Paradigm for Online Unsupervised Clustering

James E. Smith|arXiv (Cornell University)|Apr 25, 2020

Advanced Memory and Neural Computing56 references4 citations

TL;DR

This paper proposes a neuromorphic computational paradigm for online unsupervised clustering inspired by neuroscientific principles, using a cognitive column with temporal coding, excitatory neurons, winner-take-all inhibition, and spike-timing-dependent plasticity (STDP) for localized, efficient learning. Simulations show performance comparable to k-means on a semi-synthetic benchmark, demonstrating dynamic adaptation to streaming data and robust clustering in real time.

ABSTRACT

A computational paradigm based on neuroscientific concepts is proposed and shown to be capable of online unsupervised clustering. Because it is an online method, it is readily amenable to streaming realtime applications and is capable of dynamically adjusting to macro-level input changes. All operations, both training and inference, are localized and efficient. The paradigm is implemented as a cognitive column that incorporates five key elements: 1) temporal coding, 2) an excitatory neuron model for inference, 3) winner-take-all inhibition, 4) a column architecture that combines excitation and inhibition, 5) localized training via spike timing de-pendent plasticity (STDP). These elements are described and discussed, and a prototype column is given. The prototype column is simulated with a semi-synthetic benchmark and is shown to have performance characteristics on par with classic k-means. Simulations reveal the inner operation and capabilities of the column with emphasis on excitatory neuron response functions and STDP implementations.

Motivation & Objective

• To develop a biologically inspired, online unsupervised clustering method suitable for real-time streaming data.
• To enable dynamic adaptation to macro-level input changes through localized, incremental learning.
• To implement a cognitive column architecture that integrates neuroscientific principles for efficient training and inference.
• To demonstrate competitive clustering performance against classic k-means using a semi-synthetic benchmark.
• To validate the role of STDP and temporal coding in enabling robust, online clustering.

Proposed method

• The paradigm is implemented as a cognitive column integrating five core elements: temporal coding, excitatory neuron inference, winner-take-all inhibition, column-level excitation-inhibition integration, and localized STDP-based training.
• Temporal coding encodes input features in spike timing, enabling efficient representation of continuous data streams.
• Excitatory neurons compute activation based on spike input, with responses tuned to temporal patterns and input frequency.
• Winner-take-all inhibition ensures sparse, non-overlapping activation across neurons, promoting distinct cluster representation.
• Spike-timing-dependent plasticity (STDP) enables online weight updates based on pre- and post-synaptic spike timing, allowing continual learning without retraining.
• The entire system operates in a localized, event-driven manner, minimizing computational overhead and supporting real-time deployment.

Experimental results

Research questions

• RQ1 Can a neuromorphic system based on neuroscientific principles achieve competitive unsupervised clustering performance in an online, streaming setting?
• RQ2 How do temporal coding and STDP contribute to the formation of stable, distinct clusters in dynamic input environments?
• RQ3 To what extent does the cognitive column architecture enable efficient, localized learning without centralized coordination?
• RQ4 How does the system adapt to macro-level changes in input distribution over time?
• RQ5 Does the proposed paradigm match or exceed the performance of classic k-means in online clustering tasks?

Key findings

• The neuromorphic clustering paradigm achieves performance on par with classic k-means on a semi-synthetic benchmark, validating its effectiveness.
• The system demonstrates dynamic adaptability to macro-level input changes, maintaining cluster stability during data distribution shifts.
• Simulations confirm that excitatory neuron response functions are sensitive to input timing and frequency, enabling robust feature encoding.
• STDP-based training enables continuous, incremental learning with localized weight updates, supporting real-time deployment.
• The cognitive column architecture successfully integrates excitation and inhibition, resulting in sparse, non-overlapping activation patterns ideal for clustering.
• The method is computationally efficient and well-suited for streaming, real-time applications due to its online, localized, and event-driven nature.