Asymmetric Distillation and Information Retention in Capacity-Constrained Cross-Modal Transfer

arXiv:2603.06698v1 Announce Type: new
Abstract: Knowledge distillation between asymmetric architectures often induces severe geometric constraints on the learned representation space. In this work, we investigate the Dimensional Collapse phenomenon when distilling a 500M parameter global Vision Transformer (CLIP ViT-B/32) into strictly capacity-constrained, local-receptive-field CNNs (0.5M to 8.0M parameters) on the CIFAR-10 dataset. By employing strictly centered Singular Value Decomposition (SVD) and Variance-based Shannon Entropy Effective Rank, we isolate true structural variance from mean-vector artifacts. Our empirical results demonstrate a capacity-agnostic phase transition: while the Teacher exhibits an Effective Rank of 88.68, all Student models experience severe dimensional collapse to an intrinsic Effective Rank of ~16. By probing robustness, we uncover that this 81% reduction in effective dimensionality strips away the Teacher’s inherent noise immunity (which retains 89.35% accuracy under sigma=0.1 Gaussian noise). Furthermore, information-theoretic analysis using InfoNCE reveals a critical trade-off within this bottleneck: excess Student capacity densely packs the collapsed subspace for clean data, but induces severe brittleness (43.76% at sigma=0.1). Conversely, extreme capacity constraints (0.5M parameters) act as a robust low-pass filter, preserving higher noise immunity (54.84%). Explicit input augmentation fails to restore the larger model’s robustness, proving this fragility is a fundamental geometric limitation of asymmetric cosine distillation.