Rama Chellappa

Guoyizhe Wei, Rama Chellappa

Vision Transformers (ViTs) have delivered remarkable progress through global self-attention, yet their quadratic complexity can become prohibitive for high-resolution inputs. In this work, we present ViT-Linearizer, a cross-architecture distillation framework that transfers rich ViT representations into a linear-time, recurrent-style model. Our approach leverages 1) activation matching, an intermediate constraint that encourages student to align its token-wise dependencies with those produced by the teacher, and 2) masked prediction, a contextual reconstruction objective that requires the student to predict the teacher's representations for unseen (masked) tokens, to effectively distill the quadratic self-attention knowledge into the student while maintaining efficient complexity. Empirically, our method provides notable speedups particularly for high-resolution tasks, significantly addressing the hardware challenges in inference. Additionally, it also elevates Mamba-based architectures' performance on standard vision benchmarks, achieving a competitive 84.3% top-1 accuracy on ImageNet with a base-sized model. Our results underscore the good potential of RNN-based solutions for large-scale visual tasks, bridging the gap between theoretical efficiency and real-world practice.

Azadeh Alavi, Vishal M Patel, Rama Chellappa

It is proven that encoding images and videos through Symmetric Positive Definite (SPD) matrices, and considering the Riemannian geometry of the resulting space, can lead to increased classification performance. Taking into account manifold geometry is typically done via embedding the manifolds in tangent spaces, or Reproducing Kernel Hilbert Spaces (RKHS). Recently, it was shown that embedding such manifolds into a Random Projection Spaces (RPS), rather than RKHS or tangent space, leads to higher classification and clustering performance. However, based on structure and dimensionality of the randomly generated hyperplanes, the classification performance over RPS may vary significantly. In addition, fine-tuning RPS is data expensive (as it requires validation-data), time consuming, and resource demanding. In this paper, we introduce an approach to learn an optimized kernel-based projection (with fixed dimensionality), by employing the concept of subspace clustering. As such, we encode the association of data points to the underlying subspace of each point, to generate meaningful hyperplanes. Further, we adopt the concept of dictionary learning and sparse coding, and discriminative analysis, for the optimized kernel-based projection space (OPS) on SPD manifolds. We validate our algorithm on several classification tasks. The experiment results also demonstrate that the proposed method outperforms state-of-the-art methods on such manifolds.