Directed Structural Adaptation to Overcome Statistical Conflicts and Enable Continual Learning

Past decade has shown that complex networks should be at Structural adaptation Structural adaptation in NNs the core of any AI system that needs to be of robust use hasn't gained as much attention as other aspects of this technology, in any task of reasonable complexity. It has, however, been as many of these methods involve an additional step unfortunate that over the same period, the field of machine and often don't provide a significant benefit compared to the learning (ML) has been stuck in the twin limiting paradigms added complexity. One subfield in literature, called "neural of static topologies and statistical fine-tuning, attempting architecture search," focuses on optimizing the architecture to make up for the limitations of both of these by using itself explicitly (Liu, Simonyan, and Yang 2018; Shin, brute force, in form of overparameterization and computational Packer, and Song 2018; Baker et al. 2016; Stanley et al. requirements accompanying it. Limitations imposed 2019; Liu et al. 2017; Miikkulainen et al. 2019). Some other by these paradigms also prevent solving the crucial problem works view "structural adaptation" as starting from scratch of "catastrophic forgetting" in continual learning. In this or growth, sometimes referred to as Artificial Embryogenesis work, we first propose a novel method of structural adaptation, (Kowaliw et al. 2014), often using evolutionary algorithms.