Dual-Neural Driving Control for Self-Learned Trajectory Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current artificial intelligence driving devices rely on probability-based models, which are less effective compared to the neuron-based calculation of the human brain, necessitating a self-learning intelligent driving device that can mimic human cognition and improve training processes.
Innovation Solution
A self-learning intelligent driving device is proposed, comprising a first neural network module for action evaluation, a switching unit, a second neural network module for image evaluation, and a driving unit with a robotic arm, where the auxiliary AI module accelerates the training of the main AI module through deep learning processes, enabling self-learning capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If probability-based models are used in driving devices, then the device can operate with existing computational frameworks, but the device fails to achieve neuron-based calculation efficiency and self-learning capability
Solution Approach 1:
The patent divides the computational model into two distinct neural network modules: a hypothesis generation network and a hypothesis verification network. This segmentation allows each module to specialize in specific functions (generation vs. verification), enabling self-learning capability while managing complexity through modular architecture.
Solution Approach 2:
The patent introduces hypothesis coordinates as an intermediary representation between image input and motion output. This intermediary layer enables the system to learn and generate hypotheses about object states, facilitating self-learning while maintaining a structured computational framework that manages complexity.
2Productivity
If traditional deep learning training processes are used, then the training process can follow conventional methods, but the training process is time-consuming and lacks efficiency
Solution Approach 1:
The patent performs preliminary hypothesis generation before full training begins. The hypothesis generation network creates initial predictions that guide the training process, allowing the system to start from informed hypotheses rather than random initialization, thereby reducing training time while maintaining efficient learning.
Solution Approach 2:
The patent implements a feedback mechanism where the hypothesis verification network evaluates generated hypotheses and provides error signals back to the hypothesis generation network. This feedback loop accelerates convergence by directing learning toward accurate predictions, improving training efficiency while reducing the time needed to achieve competent performance.
3Adaptability or versatility
If the system generates and verifies hypotheses through dual neural network modules, then the device achieves self-learning capability, but the device complexity increases
Solution Approach 1:
The patent merges the hypothesis generation and verification functions into a unified training framework where both networks are trained jointly but with distinct objectives. This merging allows the system to achieve self-learning capability through their interaction while managing architecture complexity through shared computational infrastructure and coordinated training procedures.
Solution Approach 2:
The patent designs both neural network modules to use universal computational building blocks (convolutional layers, activation functions, optimization algorithms) that can be applied to different tasks. This multi-functionality allows the system to achieve self-learning capability across various driving scenarios while controlling complexity through reusable, standardized components.
Data Source
AI summary
A self-learning intelligent driving device including: a first neural network module for performing a corresponding action evaluation operation on an input image to generate at least one set of trajectory coordinates: a switching unit controlled by a switching signal, where when the switching signal is active, data received at a first port is sent to a second port, and when the switching signal is inactive, data received at the first port is sent to a third port; a second neural network module for performing a corresponding image evaluation operation on the at least one set of trajectory coordinates when the switching signal is active to generate at least one simulated trajectory image; and a driving unit having a robotic arm for generating at least one corresponding motion trajectory according to the at least one set of trajectory coordinates when the switching signal is inactive.