Mixed-Precision Quantization Population Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural network (DNN) models are inefficient for execution in limited environments like mobile devices due to their large size and high resource requirements, and existing optimization methods, such as mixed-precision quantization, face challenges in simultaneously optimizing contradictory objectives like accuracy, size, latency, and power consumption.
Innovation Solution
A processor-implemented method that determines convergence criteria for a population of instances with different mixed-precision quantizations, generates new instances based on converging features, and updates the population to optimize mixed-precision quantization, using techniques like random, genetic algorithm, Bayesian optimization, and evolutionary optimization to efficiently generate a multi-objective optimization population.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If mixed-precision quantization is applied to optimize DNN models for limited environments, then model size and execution efficiency are improved, but achieving optimal balance across multiple contradictory objectives (accuracy, size, latency, power consumption) becomes complex and computationally intensive
Solution Approach 1:
The patent segments the optimization process by dividing the population into multiple subsets, where each subset is optimized for specific objectives. This allows independent optimization of different aspects (accuracy, size, latency, power consumption) without requiring simultaneous optimization of all objectives, thereby reducing overall computational complexity while maintaining execution efficiency.
Solution Approach 2:
The patent applies preliminary action by pre-defining multiple objectives and their weightings before the optimization process begins. The system pre-establishes the trade-off relationships and evaluation criteria, which guides the subsequent optimization iterations and reduces the complexity of real-time decision-making during the optimization process.
2Manufacturing precision
If multiple initial instances are generated using optimization algorithms, then the quality of optimized population is improved, but generation time and computational resources increase
Solution Approach 1:
The patent applies partial action by generating only a limited number of initial instances (e.g., 5-10 instances) rather than exhaustively searching the entire solution space. This partial generation approach, combined with the subsequent iterative optimization process, achieves sufficient optimization quality while significantly reducing the time and computational resources required compared to complete exhaustive search.
Solution Approach 2:
The patent utilizes parameter changes by dynamically adjusting the number of initial instances, population size, and optimization iterations based on the specific problem requirements and available computational resources. This allows the system to adapt the generation process to achieve optimal quality-time balance for different scenarios without fixed overhead.
3Reliability
If population updates are performed iteratively, then convergence to optimal solution is improved, but processing time and energy consumption increase
Solution Approach 1:
The patent implements feedback mechanisms where the optimization process continuously monitors convergence progress and adjusts the number of iterations accordingly. When convergence criteria are met or the improvement plateau is detected, the process terminates early, preventing unnecessary additional iterations. This feedback-driven approach ensures reliable convergence while minimizing energy consumption by avoiding excessive processing.
Solution Approach 2:
The patent applies dynamics by making the optimization process adaptive and flexible in its execution. The system dynamically adjusts key parameters such as population size, number of instances per subset, and iteration count based on real-time performance metrics and convergence progress. This dynamic adjustment allows the system to achieve reliable convergence when needed while reducing energy consumption during early stages or when optimal solutions are readily found.
Data Source
AI summary
A processor-implemented method with neural network processing includes: determining whether a portion of a population comprising a plurality of instances to which different mixed-precision quantizations are applied for a neural network satisfies convergence criteria; generating, in response to the determination that the portion satisfies the convergence criteria, a new instance using the portion; and updating the population by adding the new instance to the population.