new generation of dendritic neural circuits modeled on the brain

A novel artificial neural circuit mimicking brain dendrites

Visual abstract of the neuromorphic dendritic network computation. Credit: Baek et al

Following the swift evolution of AI technologies, engineers across the globe are devising new architectures and hardware that imitate the structure and functions of the human brain.

Current brain-inspired technologies predominantly emulate the activity of neurons, focusing less on the comprehensive architecture of neural elements and their information processing contributions.

Tsinghua University researchers have introduced an innovative neuromorphic computational architecture designed to mimic the synaptic connections and dendritic tree structures of neurons.

Nature Electronics recently published a paper introducing a novel brain-like artificial system realized via a computational model of multi-gate silicon nanowire transistors and ion-doped sol-gel films.

Carlo Vittorio Cannistraci, a corresponding author, revealed to Tech Xplore that the concept of using neuron dendrite connectivity sparsity and morphology for efficient AI design came to him during his master's studies at the Polytechnic of Milano in Italy.

I was also captivated by the elegance of brain mechanisms like 'silent synapses,' which activate when they encounter increased electrical stimulation.

Cannistraci, inspired by his earlier studies and research interests, set out to computationally model complex brain mechanisms. In collaboration with researchers at Tsinghua University, he aimed to emulate the morphology of dendrites and synaptic functions through a neuromorphic computing model.

Carlo requested that I delve into 'Dendritic Computation' following our successful collaboration on 'neurotransistors,' noting their potential to emulate dendritic properties," shared Eunhye Baek, one of the corresponding authors, with Tech Xplore.

In collaboration with Prof. Luping Shi, I was engaged in the pursuit of enhancing a neuromorphic visual sensor system, recognizing the promising prospects offered by this strategy.

The Dendristor mimicking dendritic morphology. Credit: Adapted from Nature Electronics (2024). DOI: 10.1038/s41928-024-01171-7

I have consistently aimed to develop dynamic information processing systems that mimic brain and neuron functions. Dendritic computation particularly captivates me due to its broad spectrum of dynamic and complex properties, which remain relatively unexplored in neuromorphic engineering.

The predominant focus of current neuromorphic computing research has centered on replicating synaptic processes linked to learning and emulating the generation of neuronal spikes. Many of these studies have portrayed dendrites merely as basic transmission lines, thereby overlooking the functional aspects associated with their distinctive morphology.

Baek elaborated that dendrites leverage their arboreal structure to map spatially distributed signals, demonstrating plasticity specific to individual branches and integrating multiple synapses.

Each dendritic branch exhibits distinct sensitivity to signals of particular directionality, thereby specializing in processing spatiotemporal signals. Our research centers on exploring these intricate dendritic functions.

Cannistraci, Shi, Baek, and their team devised a pioneering device named the 'dendristor,' designed to emulate the morphology and operational characteristics of biological dendrites. This device utilizes multi-gate transistors coated with an ion-doped sol-gel film to replicate dendritic computations.

Baek explained that this film mimics dendritic branches by facilitating the movement of doped ions akin to those in neuronal dendrites, thereby modulating the transistor's current to mirror fluctuations in dendritic membrane potential. Our research showcases the dendristor's capability for nonlinear dendritic integration and directional selectivity.

Alongside the dendristor device, the recent paper from this research group introduces an artificial silent synapse. In this setup, the voltage across the dendristor's branches within the sol-gel film ensures that synaptic inputs activate only upon reaching a specific threshold, thereby enhancing the system's capability to discriminate the direction of moving visual stimuli.

Baek also described our creation of a neuromorphic dendritic neural circuit that calculates the direction of moving signals, drawing inspiration from neural circuits found in the retina and visual cortex. This circuit demonstrates the capability to detect signals moving in two dimensions and depth, integrating them to reconstruct the directional movement of objects in three-dimensional space.

By closely emulating the sparse connectivity observed in dendritic neurons, the novel neuromorphic computing approach introduced by Cannistraci, Baek, and their colleagues has demonstrated significant energy efficiency gains. This system has shown the ability to detect motion using fewer neurons compared to conventional artificial neural networks (ANNs).

3D visual motion perception of dendritic network neural circuits. Credit: Nature Electronics (2024). DOI: 10.1038/s41928-024-01171-7

The primary advantage of this new architecture lies in its ability to extend beyond mere replication of the functional aspects of biological neurons. Unlike other existing neuromorphic computing platforms, it faithfully reproduces the structure and sparse connectivity of neurons, encompassing the morphology of dendrites and the principles of silent synapses.

According to Baek, "Despite the diverse approaches in neuromorphic research targeting intelligence, our study uniquely highlights the critical role of neuronal and synaptic connection morphology in dynamic signal processing."

Our approach involved mimicking the biological process by which neurons establish functional neural circuits through sparsely mapped synaptic inputs, underscoring the critical role of this morphology in enhancing efficient neuromorphic information processing.

A notable achievement of this research team was being the first to illustrate how the spatial distribution of inhibitory and silent synapses can regulate signal processing in neuromorphic systems. This insight has the potential to influence the design of computational models and architectures that simulate silent synapses.

Cannistraci highlighted, "Sparsity and morphology have been overlooked in the construction of future AI systems." Our study is the first to illustrate how harnessing these features from authentic brain networks can facilitate the development of next-generation neuromorphic neural networks for efficient AI.

The recent initiatives undertaken by Cannistraci, Baek, and their team are expected to usher in innovative approaches to engineering neuromorphic systems based on semiconductor devices. Their proposed brain-inspired design stands to catalyze advancements in energy-efficient devices and AI tools, paving the way for sustainable computing solutions.

The researchers' upcoming studies will focus on expanding their artificial neural circuits with advanced inhibitory connections to improve the classification of dynamic visual signals. They plan to closely replicate neural connections observed in the early stages of brain development.

Cannistraci elaborated, "Our aim is to pioneer new neuromorphic dendritic network architectures capable of deep learning and addressing diverse AI tasks beyond visual perception, including time series analysis and auditory tasks."

Additionally, our goal is to develop multimodal circuits capable of processing and correlating sensory inputs from diverse modalities, such as visual and acoustic stimuli. Lastly, we aim to extend this sparse and morphological computing paradigm to traditional types of artificial neural networks implemented on digital hardware.

Source: