I am Siva, a Staff Machine Learning engineer at Calico, where I develop deep learning models aimed at decoding the complexities of aging. My primary research interests lie in the application of Bayesian principles to deep learning, particularly within the medical domain. I am particularly drawn to topics such as generative models, representation learning, domain adaptation, and uncertainty estimation. I have a keen interest in research that integrates diverse concepts in machine learning to create novel approaches. I completed my Master's degree at Boston University under the guidance of Professor Brian Kulis.
Embedded Files
Selected Publications
Selected Publications
Sivaramakrishnan Sankarapandian, Shivesh Chaudhary, Zhenghao Chen, Jun Xu, ICML 2026 AI4Science
Standard Masked Autoencoders (MAEs) for medical imaging rely on fixed patch sizes, which act as spatial filters that struggle to capture features smaller than the patch. To address this efficiently, we introduce Mosaic MAE, an architecture trained dynamically across varying patch sizes and conditioned on scale via adaptive LayerNorm. Experiments on multi-scale datasets and DXA bone scans show Mosaic MAE outperforms FlexiMAE, successfully extracting both fine-grained details and coarse semantic structures using a single pretrained model
Matt Sooknah*, Ramprakash Srinivasan*,Sivaramakrishnan Sankarapandian*, Zhenghao Chen, Jun Xu, medArxiv 2026
Traditional Genome-wide association studies (GWAS) are limited by the need for predefined phenotypes. To overcome this, we introduce Vector2Variant (V2V), a computationally efficient framework that uses high-dimensional data (like imaging) to scan for genetic associations without rigid phenotype specifications. V2V identifies the multivariate axis where genetic carriers and noncarriers differ most, generating biologically meaningful "projection phenotypes." When applied to UK Biobank data, V2V successfully identified both established biology (like CASP9 in renal failure) and novel associations (like LRRIQ1 in cardiovascular disease)
Sivaramakrishnan Sankarapandian*, Ramprakash Srinivasan*, Matt Sooknah, Elena Sorokin, Jun Xu, NeurIPS 2024 AIDrugX workshop
Traditional Genome-Wide Association Studies (GWAS) of imaging-derived phenotypes rely on manual feature annotation, which is time-consuming, subjective, and potentially overlooks valuable information in medical scans. We developed an embedding-derived phenotypes (EDPs) approach that combines self-supervised representation learning with ICD-10 codes to automatically transform medical scans into disease-relevant phenotypes, eliminating the need for manual annotation.Our method demonstrated superior performance in identifying biologically meaningful genetic associations compared to conventional approaches using either ICD-10 codes or manual phenotypes alone, offering a more efficient and comprehensive approach to phenotype derivation from medical imaging data.
Matt Sooknah, Sivaramakrishnan Sankarapandian, Ramprakash Srinivasan, Johannes Riegler, Jun Xu, NeurIPS 2024 AIDrugX workshop
In recent years, many studies have utilized cardiac magnetic resonance imaging (cMRI) to define image-derived phenotypes (IDPs) relating to heart structure and function for genome-wide association studies (GWAS). These IDPs are traditionally defined manually from volume, strain, and geometric parameters. Here we introduce an unsupervised learning approach that extracts spatiotemporal representations from cMRI videos in a large human cohort of ~68,000 subjects from the UK BioBank. The resulting representations can be used to predict age and manually crafted IDPs accurately. We further use these representations to define IDPs to capture both known and potential novel genetic associations. Our work suggests that unsupervised learning can be used to extract rich, unbiased information from medical videos with applications to genetic discovery.
Shivesh Chaudhary∗,Sivaramakrishnan Sankarapandian∗, Matt Sooknah, Joy Pai, Caroline McCue, Zhenghao
Chen1, Jun Xu, MICCAI 2024
Fluorescence microscopy is crucial for biological discovery, but image quality is limited by resolution, sample sensitivity, and other factors. Computational denoising methods can enhance signal-to-noise ratio by bypassing these constraints. Current methods are often supervised, requiring paired noisy and clean images, limiting their applicability. Baikal introduces a novel approach using Denoising Diffusion Probabilistic Models (DDPM), trained unconditionally to capture microscopy image complexities. It outperforms existing self-supervised methods across diverse datasets, offering promise for broader application without paired data.
Sivaramakrishnan Sankarapandian, Jun Xu, Zhenghao Chen, NeurIPS 2022 LMRL workshop
Implementation of Spatial LDA uses variational inference for learning model parameters and unfortunately does not scale well with dataset size and does not lend itself to speed-up via GPUs / TPUs. As researchers begin to collect larger in-situ multiplexed imaging datasets, there is a growing need for more scalable approaches for analysis of microenvironments. Here we propose a VAE-style network which we call Neural Spatial LDA extending the auto-encoding Variational Bayes formulation of classical LDA. We show Neural Spatial LDA achives significant speed-up over Spatial LDA while at the same time recovering similar topic distributions thus enabling its use in large data domains.
Sivaramakrishnan Sankarapandian, et. al, ICCV 2021 CDPath workshop
We present a pathology deep learning system (PDLS) that performs hierarchical classification of digitized whole slide image (WSI) specimens into six classes defined by their morphological characteristics, including classification of “Melanocytic Suspect” specimens likely representing melanoma or severe dysplastic nevi . We trained the system on 7,685 images from a single lab (the reference lab), including the the largest set of triple-concordant melanocytic specimens compiled to date, and tested the system on 5,099 images from two distinct validation labs. We achieved Area Underneath the ROC Curve (AUC) values of 0.93 classifying Melanocytic Suspect specimens on the reference lab, 0.95 on the first validation lab, and 0.82 on the second validation lab.
Sivaramakrishnan Sankarapandian, Brain Kulis, NeurIPS 2020 Machine Learning and Physical Sciences workshop
Higgins et al. introduce β-VAEs, which penalizes the KL divergence between the variational posterior and the prior using the hyperparameter β. Bottleneck-VAEs increase the capacity of the information bottleneck as the training progresses thus offering better reconstructions than β-VAEs. In this work, we show that Bottleneck-VAEs and β-VAEs are closely connected and propose a decreasing schedule for the hyperparameter β in β-VAEs that controls the information capacity similar to the hyperparameter C in the objective function of Bottleneck-VAEs
Julianna D. Ianni, Rajath E. Soans, Sivaramakrishnan Sankarapandian, et.al Nature Scientific Reports
We present and validate a deep learning system which classifes digitized dermatopathology slides into 4 categories. The system is developed using 5,070 images from a single lab, and tested on an uncurated set of 13,537 images from 3 test labs, using whole slide scanners manufactured by 3 diferent vendors. The system’s use of deep-learning-based confdence scoring as a criterion to consider the result as accurate yields an accuracy of up to 98%, and makes it adoptable in a real-world setting
Sivaramakrishnan Sankarapandian, et al, NeurIPS 2018 CDNNRIA workshop
In this paper, we employ a simple regularizer on the number of hidden units in the networks, which we refer to as adaptive network regularization (ANR). This method places a penalty on the number of hidden units per layer, designed to encourage compactness and flexibility of the network architecture. This penalty acts as the sole tuning parameter over the network size, increasing simplicity during training. We motivate this model using small-variance asymptotics—a Bayesian neural network with a Poisson number of units per layer becomes our model in the small-variance limit
Page updated
Google Sites
Report abuse