Access this paper (research preview draft).
Abstract:
Traditional genetic evolutionary techniques, while powerful for optimizing neural network topologies, suffer from prohibitive computational costs due to population-based evaluations. In contrast, gradient-based methods like those in transformers offer efficiency but introduce probabilistic errors, limiting their reliability in high-stakes domains such as healthcare. This paper introduces Evolutionary Adaptive Intelligence (EAI), a novel “scalar” adaptive evolutionary architecture that identifies and mutates error-prone nodes in a targeted manner using genetic algorithms (GAs), guided by a deterministic fitness function. By focusing on sparse subsets of nodes rather than entire networks, EAI achieves significant efficiency gains, enabling deployment on consumer-grade hardware.
In BibTeX format
@misc{oscilon2021-eai, title={Targeted Evolutionary Computing Approach for Efficient and Deterministic {AI}}, author={Abhiragh and Ashik Varma and Arun John Varghese}, year={2021}, month={January}, note={Research preview paper. Software available to approved researchers via osclion.ai}, url={https://osclion.ai/papers/eai-preview-2021.pdf}}
Or in textual form:
Abhiragh, Ashik Varma, and Arun John Varghese.Targeted evolutionary computing approach for efficient and deterministic AI.2021. Research preview paper. Software available to approved researchers via osclion.org.
If your work builds directly on Oscilon's sparse node identification, targeted GA mutations, or deterministic fitness thresholding, please acknowledge the Oscilon research preview and describe the specific preview version (e.g., v0.1) used.
As Oscilon progresses—with contributions from approved researchers on industrial benchmarks, edge deployment, and convergence guarantees—we will publish formal papers and establish a stable DOI for citation.
