Training neural networks with a hybrid optimization method of genetic and gradientdescent algorithms

Web Published:
1/7/2025
Description:

Training Neural Networks with a Hybrid Optimization Method of Genetic and Gradient Descent Algorithms

Princeton Docket # 25-4214

Neural networks are a powerful class of machine learning model that learn to perform useful operations such as classification, inference, and generation, and can be applied to a range of tasks, such as analysis and generating text, code, or images. Training these models, however, requires exponentially increasing amounts of data and computational power; they are also notoriously difficult to successfully optimize, potentially requiring several training attempts to attain satisfactory performance. Increasing the training efficacy of neural network models will therefore save time, reduce costs, and is an important step in unlocking further performance improvements.

Drawing inspiration from how nature combines evolution with biological adaptation to produce natural intelligence, a Princeton researcher has developed a new algorithm for training neural network models that is composed of two optimization loops: the outer loop employs a proprietary genetic algorithm to optimize the initial state of the model, and the inner loop deploys gradient-based parameter updates to fine-tune the model. The algorithm is agnostic to the model architecture and task: it can be applied to small bespoke models learning well-defined associations or on large generative models.

Novel methodological advancements were made to improve training speed and training efficacy of this algorithm. Models trained with this novel algorithm are over 20% faster than a standard genetic algorithm and are successfully trained at a rate almost 3x as high (increasing from a ~35% success rate to ~95%), as shown by testing with deep neural networks on a standard reinforcement learning benchmark. Efficacy was similarly improved (from ~70% to ~95%) over the standard and ubiquitous gradient-based method for training neural networks (stochastic gradient descent via backpropagation), thereby reducing the number of failed training runs and decreasing the amount of time and compute resources necessary to produce a successfully trained model. Taken together, this algorithm can be easily integrated into existing engineering workflows, is applicable to a diverse range of tasks and model architectures, and reduces the computational load and time required to successfully train neural network models.

 

Applications:

• Optimize the parameters of a deep neural network

• Improve training speed and increase the success rate of model training

• Potential to improve artificial intelligence learning and performance

 

 

Advantages:

• More efficient and robust than classic genetic algorithms

• Improve training speed and increase the success rate of model training

• Reduces computational cost by increasing training success rate

 

 

Stage of Development

A working version of the algorithm has been developed for an existing research project and has been validated as part of that project. Specifically, the project focused on simulating the evolution of astrobiological intelligence on non-Earth worlds. The model was used to train neural network models on a benchmark reinforcement learning task, and the training progression and efficacy was evaluated as part of this project. These results form the basis of the conclusion that this hybrid training algorithm boosts efficacy and robustness and is applicable to deep learning models.

 

 

Inventor

Benjamin Midler is the 2024 William G. Bowen fellow and a PhD candidate in the Department of Neuroscience at Princeton University. His current research uses empirical and computational methods to study both the neural dynamics of stress and astrobiology. Previously, he studied neuroscience and history at Stanford University where he was advised by James McClelland. His work has been published in leading journals and conferences.

 

Intellectual Property & Development status

Patent protection is pending.

Princeton is currently seeking commercial partners for the further development and commercialization of this opportunity.

 

Contact

Cortney Cavanaugh

Princeton University Tech Licensing & New Ventures • (609) 258 7256 • ccavanaugh@princeton.edu

Patent Information:
For Information, Contact:
Cortney Cavanaugh
New Ventures and Licensing associate
Princeton University
609-258-7256
ccavanaugh@princeton.edu
Inventors:
Benjamin Midler
Keywords: