Neural networks, also known as artificial neural networks (ANNs), are a relatively new technology that has affected many fields including engineering, mathematics, and natural language processing . Their design is inspired by the brain, which, at the cellular level, is a network of neurons that communicate with one another through electrical signals. Likewise, a neural network is a network of processing units, called nodes, which can hold, modify, and communicate pieces of information. Neural networks have multiple applications in the field of artificial life, including the study of emergent behavior and self-replication, simulation of certain biological processes, and modeling of signaling pathways and phenotype expression .
Feedforward Neural Networks
Among the most conceptually basic varieties of neural network is the feedforward neural network . Unlike biological neurons, whose communication is complex and multi-directional, feedforward neural networks communicate strictly unidirectionally. Information flows from the input layer, through each successive middle, or “hidden”, layer, and then to the output layer .
Deep Learning
Deep learning is a variety of machine learning wherein the system learns to recognize intricate patterns in large datasets. Neural networks provide the necessary framework for deep learning to occur. Feedforward neural networks are often employed as the underlying machinery for deep learning models . Their capacity for this complexity arises from the utilization of more complex functions (i.e. non-linear rather than linear ones) to perform their calculations, which means an increased ability to process information at great depth. This is what yields recognition of patterns in data that would otherwise be nearly impossible to achieve. This ability, however, hinges on the size of the datasets upon which the neural network can be trained. A sufficiently large training dataset will allow for a neural network to become highly effective at parsing similar data .
Applications in Artificial Life
Below are further examples of the utilization of neural networks and deep learning in the field of artificial life. The articles provide examples of the potential of using modified neural networks or pairing neural networks with other mathematical models to understand complex biological systems.
- Self-Replication in Neural Networks
- Development of a hybrid model for a partially known intracellular signaling pathway through correction term estimation and neural network modeling
- A neural-mechanistic hybrid approach improving the predictive power of genome-scale metabolic models
Limitations
Although neural networks have proven their utility in artificial life and show promise to assist in some of its most intractable problems, there still exist many limitations to what neural networks can achieve. Specifically, deep learning models have vastly improved in verbal reasoning, but they still struggle with complex reasoning over sequences, including reasoning with sparse data, remembering information established early in a long sequence, and logical reasoning . In deep learning specifically, models have evolved to the point that we can no longer understand the computational steps taken by the models to reach their conclusions. Some of these conclusions have proven accurate and thus useful, but there remains caution and concern surrounding the prospect of utilizing the insights produced by machines that we cannot understand theoretically, especially in fields that affect the general public . This may pose an especially challenging limitation in artificial life, as the mechanics of the system (i.e. the part of deep learning models we cannot understand) are of significant importance when modeling biological structures.