INSPIRE: Understanding Local-Global Representation Learning via Controlled Self-Supervised Training and Representation Interpretability

Self-supervised learning (SSL) has become a powerful paradigm for visual representation learning, yet it remains insufficiently understood how different training strategies shape the balance between global semantic structure and local visual detail. This study introduces the INSPIRE Framework, a unified framework for systematically investigating local-global representation learning in SSL through controlled training, downstream evaluation, and representation interpretability analysis. Within this framework, we develop the INSPIRE Method, a contrastive SSL approach that constructs merged training views by combining target images with distractor content and aligning global image representations with region-level features. By varying the number of merged regions, INSPIRE explicitly controls spatial granularity during pretraining. We evaluate INSPIRE through controlled ImageNet-100 experiments (130K images) with a ResNet-18 backbone, large-scale ImageNet-1K pretraining (1.28million images) with a ResNet-50 backbone, and transfer to fine-grained recognition benchmarks. We further introduce the INSPIRE Representation Interpretability Investigation Methodology, which combines representation similarity analysis with scale-controlled spatial perturbations to examine spatial sensitivity across SSL methods and network layers. Experiments show that representation quality varies non-monotonically with spatial granularity, with intermediate granularity often producing stronger transfer performance than coarser or more fragmented alternatives. The analysis further reveals distinct spatial sensitivity profiles across SSL objectives and progressively increasing perturbation sensitivity across network depth. Runtime experiments using single-GPU and multi-GPU distributed pretraining demonstrate the practical scalability of the framework. These findings suggest that SSL representations are better understood as occupying positions along a continuum of spatial sensitivity rather than as strictly local or global learners.

@article{Alharthi2026,
  title     = {INSPIRE: Understanding Local-Global Representation Learning via Controlled Self-Supervised Training and Representation Interpretability},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {7},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Reem Alharthi and Rashid Mehmood and Aiiad Albeshri and Abdulaziz A. Almuzaini},
  doi       = {10.14569/IJACSA.2026.0170731},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170731}
}