Biologically Plausible Neural Network with Local Connectivity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional artificial neural networks often ignore or contradict principles of biological neural networks, leading to limitations in achieving biological intelligence and efficient learning mechanisms, with issues such as non-neuron units, weight inconsistency, non-local connectivity, and random initialization being biologically impossible or implausible.
Innovation Solution
A neural network configuration that mimics biological neural networks by using structured local patterns of connections, alternating layers of excitatory and inhibitory neurons, and learning rules based on output neurons, with artificial neurons following a simple step function and positive/negative synaptic weights, and topographically arranged neurons with dendrites extending to connect based on position and orientation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional artificial neural networks use random initialization and non-biological connection patterns, then training flexibility is improved, but biological plausibility deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-configuring neurons with biologically plausible properties (topographic arrangement, selective connectivity based on orientation and position) before training begins. This preliminary configuration establishes biologically realistic constraints while still allowing learning to occur within those constraints, resolving the contradiction between biological plausibility and training flexibility.
Solution Approach 2:
The patent implements local quality by making each neuron's connectivity specific to its local properties in the topography. Neurons connect based on their position and orientation characteristics rather than random connections, creating locally adapted connection patterns that are biologically plausible while maintaining overall network adaptability through the distributed nature of these local connections.
2Productivity
If neurons are arranged in topography with selective connectivity, then learning efficiency is improved, but network complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the neural network into distinct layers with specific functional roles (input layer, hidden layers with orientation-selective neurons, output layer). Each segment has specialized connectivity patterns based on topographic position, which organizes complexity into manageable functional units while improving learning efficiency for specific tasks like orientation detection.
3Measurement precision
If extensive training is used to achieve high accuracy, then classification performance is improved, but training time increases
Solution Approach 1:
The patent applies preliminary action by pre-configuring the network with biologically plausible connection patterns based on topographic position and orientation selectivity before training. This preliminary structure allows the network to achieve high classification accuracy with significantly less training time, as the biologically realistic constraints guide learning more efficiently than random initialization.
Data Source
AI summary
Certain aspects of the present disclosure provide systems and methods for configuring and training neural networks. The method includes models of individual neurons in a network that avoid certain biologically impossible or implausible features of conventional artificial neural networks. Exemplary networks may use patterns of local connections between excitatory and inhibitory neurons to provide desirable computational properties. A network configured in this manner is shown to solve a digit classification problem.