Description
This study examines whether altered pruning in a generative neural network leads to more restricted or repetitive outputs, using Autism Spectrum Disorder (ASD) as a cautious computational analogy rather than as a biological equivalence. A compact denoising diffusion probabilistic model (DDPM) with a U-Net backbone was trained on CIFAR-10, and pruning percentage was treated as the primary structural variable. Instead of relying on a single assumed baseline, the final implemented study evaluated 0%, 10%, 15%, 20%, 35%, and 50% pruning under controlled conditions: the same dataset, the same epoch budget, the same random seed, and 1,000 generated samples per condition. Output diversity and repetition were measured with CLIP-embedding-based pairwise cosine distance, nearest-neighbor similarity, and a near-duplicate rate. The completed seed 0, epoch 10 sweep showed that repetition was lowest at 20% pruning (mean nearest-neighbor similarity = 0.9576), while heavier pruning degraded diversity and increased repetition. The results do not support a simple linear claim that less pruning always causes more repetitive behavior or that more pruning always improves the model. Instead, the descriptive pattern is best interpreted as a shallow U-shaped trend within the tested range: both dense and heavily pruned regimes can increase repetition, while an intermediate pruning region produces the most flexible generative behavior. Because the final figure set is based on a completed single-seed sweep rather than a fully averaged multi-seed study, the findings should be read as careful empirical evidence for this configuration rather than as a universal claim. Together, these results show that connectivity structure can affect generative expressivity in diffusion models.
Publication Date
4-2026
Thesis Advisor
Jake Qualls
Thesis Committee
Jason Causey, Karen Yanowitz, Carmen Williams
Degree Name
BS, Computer Science
Expected Graduation
5-2026
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
Recommended Citation
Momen, Sara, "Altered Pruning in a Generative Neural Network" (2026). Honors Theses. 11.
https://arch.astate.edu/honors-thesis/11
