Brain-like Visual Neural Network Forward Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning vision algorithms face challenges such as lack of explicit position encoding, reliance on error back-propagation and gradient descent, limited ability to combine and abstract multiple dimensions of information, absence of reverse neural pathways, and inefficient training processes, which hinder accurate shape and position recognition and generalization.
Innovation Solution
A brain-like visual neural network with forward-learning and meta-learning functions, comprising primary and composite feature encoding modules with excitatory and inhibitory connections, supporting bidirectional information processing and employing synaptic plasticity to encode and abstract visual features efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If error back-propagation and gradient descent are used for training, then the neural network can learn from data, but the training cost is high and it requires a large number of partial differential operations
Solution Approach 1:
The patent replaces the traditional back-propagation training mechanism with a forward-learning mechanism inspired by biological neural systems. Instead of using gradient descent with partial differential operations, the system uses forward propagation with local learning rules at each neuron, eliminating the need for complex mathematical operations and significantly reducing training time while maintaining learning effectiveness.
Solution Approach 2:
The training process is segmented into local learning events at individual neurons rather than global gradient computations. Each neuron independently updates its weights based on local activity patterns, dividing the complex training task into many simple parallel operations that can be executed efficiently without requiring expensive back-propagation through the entire network.
2Device complexity
If only forward neural pathway is implemented, then the network structure is simple, but it cannot support top-down information processing and lacks reverse neural pathway functionality
Solution Approach 1:
The patent implements dynamic bidirectional pathways where neural connections can actively transmit information in both forward and reverse directions. The reverse pathway is not a static structural addition but a dynamic capability that activates when needed for top-down processing, allowing the network to flexibly switch between bottom-up feature extraction and top-down contextual processing modes.
Solution Approach 2:
The same neural connections serve multiple functions: they participate in both forward propagation for feature extraction and reverse propagation for contextual modulation. This multi-functionality allows a single network architecture to handle both bottom-up and top-down information processing without requiring separate dedicated pathways, maintaining structural simplicity while enhancing versatility.
3Device complexity
If implicit position encoding is used, then the network is simpler without special position encoding circuits, but it cannot flexibly combine visual features at any position and has weak generalization ability
Solution Approach 1:
The patent introduces an explicit position encoding mechanism as an intermediary layer that transforms spatial position information into a format suitable for neural processing. This position encoding acts as a mediator between the raw visual input and the feature combination operations, enabling the network to flexibly combine features from any positions while maintaining a relatively simple overall architecture.
Solution Approach 2:
The patent transforms position information from spatial coordinates into an encoded dimensional representation that can be easily manipulated by neural operations. By converting position data into a different dimensional form through explicit encoding, the network gains the ability to flexibly combine visual features across different positions without requiring complex spatial reasoning capabilities.
4Reliability
If traditional deep learning algorithms are used, then the training process is well-established, but it requires a large amount of training data and has long training cycles
Solution Approach 1:
The patent implements forward-learning mechanisms where neurons automatically adjust their weights based on local activity patterns without requiring external gradient signals. This self-service learning capability allows the network to adapt efficiently to new data with fewer training examples and shorter training cycles, while maintaining stable learning through biologically-inspired plasticity rules.
Data Source
AI summary
A brain-like visual neural network with forward-learning and meta-learning functions is provided, comprising the primary feature encoding module, the composite feature encoding module, comprising the active attention mechanism and automatic attention mechanism, having neural loops that explicitly encode the location information of visual features, having forward neural pathways and reverse neural pathways, supporting upper and lower bi-directional information processing, adopting a variety of plasticity process with biological rationality, with the ability to conduct forward-learning to fast encode the visual representation of the input image or video information to memory information, and to conduct information abstraction process and information component adjustment process to obtain the common feature information and different feature information between objects, forming information channels with multiple information dimensions and information abstraction degrees, improving generalization ability while retaining detailed information.