Overview

A set of supervised and self-supervised style experiments executed as part of a deep learning curriculum and self-directed study. The goal was to compare architectures (EfficientNet, ViT, custom ResNet) under controlled augmentation and compute budgets, and to explore interpretability methods (Grad-CAM) and perceptual loss variants for downstream image tasks. :contentReference[oaicite:9]9

Representative Experiments

• EfficientNetV2 fine-tuning: Fine-tuned on Oxford Flowers-102 with custom augmentations and multi-GPU training; validated class-balance strategies and augmentation schedules.
• ViT & Representational Probes: Trained ViT-B32 on a domain-specific PV-fault dataset, compared head-only vs full-finetune methods.
• ResNet-36 & custom activations: Implemented a ResNet-36 from scratch, experimented with an analytic custom activation and benchmarked against ReLU baselines.
• Interpretability: Generated Grad-CAM maps to inspect class discriminative regions and evaluate whether features align with human-perceptible cues.

Outcomes

• Demonstrated practical knowledge of end-to-end vision pipelines (dataset curation → augmentation → multi-GPU training → evaluation).
• Gained fluency in model selection trade-offs, fine-tuning strategies, and basic interpretability techniques useful for downstream research and production prototypes. :contentReference[oaicite:10]10

Notes

These experiments are compact and reproducible; they serve as methodological building blocks for larger perception projects (e.g., Mesquite MoCap visual fusion, Happenstance).