About This Project

Modern AI strategies relying on the Artificial Neural Network (ANN) have achieved massive gains in accuracy and performance over prior methodologies, but they still face flaws in terms of error rates and over-sized, inefficient models. However, natural intelligence has evolved to be accurate and efficient. I hypothesize that if neural network topology design is generated by simulating evolution, then the resulting networks will have higher accuracy and efficiency than existing designs.

Ask the Scientists

What is the context of this research?

Modern AI systems perform at high levels on the tasks they were specifically designed to perform, as with the mere 0.21% error on handwritten digit classification achieved at New York University. However, they often fail spectacularly when conditions change even slightly due to the catastrophic interference problem. Part of the problem is that Artificial Neural Networks, the systems behind modern AI, have only a shallow relation to how our brains have actually evolved to function. I hypothesize that AI solutions can broaden their capabilities and make gains in accuracy and efficiency by taking cues from the biological process of group evolution.

What is the significance of this project?

Little data has been collected on how biological mimicry can benefit AI since specific research on the topic only recently began, such as openAI beginning research in march of 2017. As a result, in-depth study is necessary to discern exactly how much performance can be gained, what features can be gained, and by what methods. Specifically, I study the mimicry of group evolution. The proposed experiment will test if group evolution can be used to achieve higher accuracy, lower computational cost, and scalability to any problem size. This data will shed additional light on the validity of biological mimicry, either opening the door to future research or indicating a need to focus on alternative methods.

What are the goals of the project?

My hypothesis is that modeling evolution can increase the efficiency and accuracy of neural networks. I will test the top 20 state of the art designs listed on http://rodrigob.github.io on the standard MNIST data set as well as the naive approaches of the multilayer perceptron and convolutional neural network to establish baseline measurements. From there, I will generate networks with evolution simulations that measure fitness in terms of prediction accuracy and inference time. The algorithm will be tweaked after each trial to improve realism and effectiveness. After 20 trials and iterations of the algorithm, I will compile the data to measure if the evolution simulation improved accuracy and efficiency and note what simulation features were beneficial.