Samar Agnihotri

Anika Shrivastava, Renu Rameshan, Samar Agnihotri

Masked Autoencoders (MAEs) achieve impressive performance in image classification tasks, yet the internal representations they learn remain less understood. This work started as an attempt to understand the strong downstream classification performance of MAE. In this process we discover that representations learned with the pretraining and fine-tuning, are quite robust - demonstrating a good classification performance in the presence of degradations, such as blur and occlusions. Through layer-wise analysis of token embeddings, we show that pretrained MAE progressively constructs its latent space in a class-aware manner across network depth: embeddings from different classes lie in subspaces that become increasingly separable. We further observe that MAE exhibits early and persistent global attention across encoder layers, in contrast to standard Vision Transformers (ViTs). To quantify feature robustness, we introduce two sensitivity indicators: directional alignment between clean and perturbed embeddings, and head-wise retention of active features under degradations. These studies help establish the robust classification performance of MAEs.

Mohd Saif Ali Khan, Karthik RM, Samar Agnihotri

Efficient uplink processing in distributed massive multiple-input multiple-output (D-mMIMO) systems requires both effective local combining and scalable decoding to significantly mitigate inter-user interference. Recent zero-forcing (ZF)-based combining schemes, such as partial full-pilot ZF (PFZF) and protected weak PFZF (PWPFZF), rely on heuristic threshold-based user grouping that may lead to inefficient utilization of spatial degrees of freedom across access points (APs). To address this limitation, we propose adaptive pilot-aware local combining strategies, generalized PFZF (G-PFZF) and generalized PWPFZF (G-PWPFZF), that dynamically allocate spatial degrees of freedom based on local channel conditions and replace heuristic grouping with a decentralized pilot-level optimization framework. Thus providing substantial performance gains over conventional PFZF and PWPFZF. Further, centralized decoding has recently emerged as a promising technique for interference suppression in D-mMIMO systems. However, it incurs substantial fronthaul overhead and computational costs. We develop a decentralized large-scale fading decoding (d-LSFD) scheme in which each AP computes LSFD weights using only locally available channel statistics. We derive a lower bound on the signal-to-interference-plus-noise ratio that explicitly quantifies the performance gap between the proposed d-LSFD scheme and centralized LSFD (c-LSFD), and identifies conditions under which the proposed decentralized solution approaches the centralized optimum. Numerical results demonstrate that the proposed generalized combining and the d-LSFD scheme together achieve significantly higher sum spectral efficiency in comparison to any combination of existing local combining and decoding schemes, while also substantially reducing the computational cost and fronthaul overhead.

Anika Shrivastava, Renu Rameshan, Samar Agnihotri

Understanding the latent spaces learned by deep learning models is crucial in exploring how they represent and generate complex data. Autoencoders (AEs) have played a key role in the area of representation learning, with numerous regularization techniques and training principles developed not only to enhance their ability to learn compact and robust representations, but also to reveal how different architectures influence the structure and smoothness of the lower-dimensional non-linear manifold. We strive to characterize the structure of the latent spaces learned by different autoencoders including convolutional autoencoders (CAEs), denoising autoencoders (DAEs), and variational autoencoders (VAEs) and how they change with the perturbations in the input. By characterizing the matrix manifolds corresponding to the latent spaces, we provide an explanation for the well-known observation that the latent spaces of CAE and DAE form non-smooth manifolds, while that of VAE forms a smooth manifold. We also map the points of the matrix manifold to a Hilbert space using distance preserving transforms and provide an alternate view in terms of the subspaces generated in the Hilbert space as a function of the distortion in the input. The results show that the latent manifolds of CAE and DAE are stratified with each stratum being a smooth product manifold, while the manifold of VAE is a smooth product manifold of two symmetric positive definite matrices and a symmetric positive semi-definite matrix.

Samar Agnihotri, Renu Rameshan, Ritwik Ghosal

Lossless image compression is required in various applications to reduce storage or transmission costs of images, while requiring the reconstructed images to have zero information loss compared to the original. Existing lossless image compression methods either have simple design but poor compression performance, or complex design, better performance, but with no performance guarantees. In our endeavor to develop a lossless image compression method with low complexity and guaranteed performance, we argue that compressibility of a color image is essentially derived from the patterns in its spatial structure, intensity variations, and color variations. Thus, we divide the overall design of a lossless image compression scheme into three parts that exploit corresponding redundancies. We further argue that the binarized version of an image captures its fundamental spatial structure. In this first part of our work, we propose a scheme for lossless compression of binary images. The proposed scheme first learns dictionaries of $16\times16$, $8\times8$, $4\times4$, and $2\times 2$ square pixel patterns from various datasets of binary images. It then uses these dictionaries to encode binary images. These dictionaries have various interesting properties that are further exploited to construct an efficient and scalable scheme. Our preliminary results show that the proposed scheme consistently outperforms existing conventional and learning based lossless compression approaches, and provides, on average, as much as $1.5\times$ better performance than a common general purpose lossless compression scheme (WebP), more than $3\times$ better performance than a state of the art learning based scheme, and better performance than a specialized scheme for binary image compression (JBIG2).

Siddhartha Sarma, Samar Agnihotri, Joy Kuri

The problem of secure communication in Amplify-and-Forward (AF) relay networks with multiple eavesdroppers is considered. Assuming that a receiver (destination or eavesdropper) can decode a message only if the received SNR is above a predefined threshold, we introduce SNR based optimization formulations to calculate optimal scaling factors for relay nodes in two scenarios. In the first scenario, we maximize the achievable rate at the legitimate destination, subject to the condition that the received SNR at each eavesdropper is below the target threshold. Due to the non-convex nature of the objective function and eavesdroppers' constraints, we transform variables and obtain a Quadratically Constrained Quadratic Program (QCQP) with convex constraints, which can be solved efficiently. When the constraints are not convex, we consider a Semi-definite relaxation (SDR). In the second scenario, we minimize the total power consumed by all relay nodes, subject to the condition that the received SNR at the legitimate destination is above the threshold and at every eavesdropper, it is below the corresponding threshold. We propose a semi-definite relaxation of the problem in this scenario and also provide an analytical lower bound.

Siddhartha Sarma, Samar Agnihotri, Joy Kuri

We have evaluated the optimal secrecy rate for Amplify-and-Forward (AF) relay networks with multiple eavesdroppers. Assuming i.i.d. Gaussian noise at the destination and the eavesdroppers, we have devised technique to calculate optimal scaling factor for relay nodes to obtain optimal secrecy rate under both sum power constraint and individual power constraint. Initially, we have considered special channel conditions for both destination and eavesdroppers, which led us to analytical solution of the problem. Contrarily, the general scenario being a non-convex optimization problem, not only lacks an analytical solution, but also is hard to solve. Therefore, we have proposed an efficiently solvable quadratic program (QP) which provides a sub-optimal solution to the original problem. Then, we have devised an iterative scheme for calculating optimal scaling factor efficiently for both the sum power and individual power constraint scenario. Necessary figures are provided in result section to affirm the validity of our proposed solution.