Neural Network Dummy Layer for Reverse Decision Variable Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models cannot deduce decision variables in reverse for cell culture processes, leading to labor-intensive and resource-consuming design and improvement processes, and this challenge is not limited to cell culture but also applies to other neural network training processes.
Innovation Solution
A method involving a dummy layer with artificial neurons connected to a pre-trained predictive model, where the bias value of activation functions is set to 0, and an optimizer adjusts the weight values of these neurons to match the target result, allowing the deduction of decision variables.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning predictive models are used for cell culture process design, then model accuracy and convergence speed are improved, but the ability to deduce decision variables in reverse is lost, resulting in labor-intensive and resource-consuming design processes
Solution Approach 1:
The patent applies inversion by reversing the traditional forward prediction process. Instead of inputting decision variables to predict outcomes, the system inputs target outcomes and uses the trained predictive model to deduce the decision variables that would produce those outcomes. This is achieved by treating the predictive model as a differentiable function and using gradient-based optimization to solve for input parameters given desired outputs, thereby enabling reverse deduction of decision variables while maintaining model accuracy.
2Adaptability or versatility
If numerous decision variables are manually tried during process design, then comprehensive parameter exploration is achieved, but the design and improvement process becomes labor-intensive and resource-consuming
Solution Approach 1:
The system applies self-service by automatically performing the parameter exploration and optimization that would otherwise require manual intervention. The trained predictive model, combined with gradient-based optimization, autonomously deduces optimal decision variables for given target outcomes without requiring researchers to manually test numerous parameter combinations. This automation maintains comprehensive parameter exploration while dramatically improving design process efficiency by eliminating repetitive manual trials.
3Ease of manufacture
If fixed decision variables are used for cell preparation manufacturing, then production simplicity is improved, but the ability to handle high variability between cells from different individuals is lost
Solution Approach 1:
The patent applies dynamics by transitioning from static fixed decision variables to dynamic personalized decision variables. The system uses the trained predictive model to determine optimal decision variables specific to each individual's cell characteristics and desired outcomes. This allows the manufacturing process to adapt parameters based on individual variability while maintaining systematic control, thereby achieving both manufacturing simplicity through automation and adaptability through personalized parameter optimization.
Data Source
Figure 1A~1B
Figure 2
Figure 3
AI summary
The present invention provides a method for calculating decision variables. A dummy layer is added at an input layer of a trained neural network predictive model. The dummy layer includes a plurality of artificial neurons respectively connected to a corresponding input terminal of the input layer for the trained predictive model by a newly established link. The input value of each artificial neuron is set to 1, the bias value of the activation function is set to 0, and the output of the activation function is set to 1 when the input of the activation function is 1. The initial weight value of the newly established link is selected and set, and the weight values can be considered as decision variables, wherein the weight values can have ranges or other inter-conditional restrictions. The optimal solution is obtained using the optimizer built in a general machine-learning platform when the parameters of the trained predictive model are frozen, and only the weight values of the newly established links are adjusted. The training objective is set so that the output of the parameter predictive model matches the desired target result. At the end of the training, the weight values of the newly established link new are the feasible input decision variables. This invention allows the effective use of the general machine learning platforms and the built-in methods to find the optimal input parameters that achieve the expected results.