The Fractured Entangled Representation (FER) hypothesis proposes that gradient descent produces disordered internal structure: redundant, entangled representations that degrade under distribution shift or new information. In contrast, neuroevolution produces sparse, modular representations hypothesized to generalize better and adapt more readily to changing environments. This thesis asks whether that qualitative gap can be made quantitative and, if so, whether the resulting measurements can be turned into actionable training signals that improve network behavior. Using CPPNs as a visualizable testbed, metrics are developed that consistently separate evolved and gradient-trained networks. A differentiable penalty is then created that shapes SGD toward less fractured representations. Evaluated on digit classification, the penalty improves generalization in data-constrained settings and, more strikingly, keeps networks adaptable under continual learning even in regimes where baselines effectively stop learning: demonstrating that representational quality, not just task performance, is a tractable objective for gradient-based learning. More broadly, these results reframe representational geometry as a directly optimizable property: one that targets failure modes orthogonal to scale and composes with existing training methods rather than replacing them.

PDF

Technical Report, Department of Computer Sciences, The University of Texas at Austin, Austin, Texas, May 2026.

@techreport{xu:hthesis26, title={Measuring and Mitigating Fractured Entangled Representations in Neural Networks}, author={Kevin Xu}, month={May}, address={Austin, Texas}, institution={Department of Computer Sciences, The University of Texas at Austin}, type={Undergraduate Honors Thesis}, url="http://nn.cs.utexas.edu/?xu:honorsthesis26", year={2026} }