LISA conditions the latent geometry of deterministic autoencoders for ex-post generation by projecting codes onto a hypersphere and encouraging directional coverage with a repulsive kernel regularizer. Coupled with projected multivariate Gaussian or Gaussian-mixture samplers, it improves generation across image benchmarks without stochastic encoders, adversarial training, or learned priors. Grammar-LISA extends the approach to structured arithmetic expressions.
@article{mazumder2026lisa,title={{LISA}: Latent Inference on the Sphere for A-posteriori sampling},author={Mazumder, Alokendu and Garg, Chirag and Baruah, Tirthajit and Rathore, Punit},year={2026},month=sep,journal={Transactions on Machine Learning Research},url={https://openreview.net/forum?id=f0agX08y69},}
TMLR
Global Linear Convergence of Inexact TD under Generalized Smoothness
@article{mazumder2026td,title={Global Linear Convergence of Inexact TD under Generalized Smoothness},author={Mazumder, Alokendu and Ananta, Ila and Rathore, Punit},year={2026},journal={Transactions on Machine Learning Research},url={https://openreview.net/forum?id=Xvyehi4izc},}
TMLR
On Hamming-Lipschitz Type Stability of Subdominant (Min-Max) Ultrametric: Theory and Simple Proofs
@article{mazumder2026ultrametric,title={On Hamming-Lipschitz Type Stability of Subdominant (Min-Max) Ultrametric: Theory and Simple Proofs},author={Mazumder, Alokendu and Roy, Arnab and Rathore, Punit},year={2026},journal={Transactions on Machine Learning Research},url={https://openreview.net/forum?id=R4ASOCp3uM},}
IEEE TAI
On Convergence of ADAM with Data Dependent Stepsize
Alokendu Mazumder, Rishabh Sabharwal, Bhartendu Kumar*, and 4 more authors
IEEE Transactions on Artificial Intelligence, 2026
We propose a data and dynamics driven constant stepsize for Adam that removes the need for costly tuning and avoids the pitfalls of decaying schedulers. Our estimate matches the best stepsizes found by exhaustive search, performs on par with popular schedulers, and scales to ImageNet and LLM fine-tuning. We further provide theory showing that Adam with our stepsize converges to critical points for smooth non-convex objectives.
@article{mazumder2026adam,title={On Convergence of ADAM with Data Dependent Stepsize},author={Mazumder, Alokendu and Sabharwal, Rishabh and Kumar, Bhartendu and Tayal, Manan and Garg, Chirag and Roy, Arnab and Rathore, Punit},year={2026},journal={IEEE Transactions on Artificial Intelligence},url={https://ieeexplore.ieee.org/document/11361385},}
ICASSP
PosSpec-Net: Enhancing Radar-Based Speech Signals via U-Nets and Positional Vectors
Yash Soni, Alokendu Mazumder, Prashant Mishra, and 1 more author
In ICASSP 2026 Radar Acoustic Speech Enhancement Grand Challenge · 4th Place, 2026
@inproceedings{soni2026posspec,title={PosSpec-Net: Enhancing Radar-Based Speech Signals via U-Nets and Positional Vectors},author={Soni, Yash and Mazumder, Alokendu and Mishra, Prashant and Rathore, Punit},year={2026},booktitle={ICASSP 2026 Radar Acoustic Speech Enhancement Grand Challenge · 4th Place}}
2025
IEEE TNNLS
Fractional Gradient Descent with Matrix Stepsizes for Non-Convex Optimization
We establish convergence guarantees for fractional gradient descent on matrix-smooth non-convex functions and show how matrix smoothness accelerates convergence. We introduce Compressed Fractional Gradient Descent (CFGD) with matrix-valued stepsizes, proving faster convergence than scalar stepsizes in both single-node and distributed settings. This work studies fractional gradient descent in federated/distributed optimisation.
@article{mazumder2025fractional,title={Fractional Gradient Descent with Matrix Stepsizes for Non-Convex Optimization},author={Mazumder, Alokendu and Vyas, Keshav and Rathore, Punit},year={2025},journal={IEEE Transactions on Neural Networks and Learning Systems},url={https://ieeexplore.ieee.org/document/11298544},}
ICLR (W)
DIME: Deterministic Information Maximizing Autoencoders
Alokendu Mazumder, Chirag Garg, Tirthajit Baruah, and 1 more author
In Workshop on Deep Generative Models @ ICLR 2025, 2025
@inproceedings{mazumder2025dime,title={DIME: Deterministic Information Maximizing Autoencoders},author={Mazumder, Alokendu and Garg, Chirag and Baruah, Tirthajit and Rathore, Punit},year={2025},booktitle={Workshop on Deep Generative Models @ ICLR 2025},}
Low-Rank Autoencoder (LoRAE) extends autoencoders with a low-rank regularizer to learn compact and implicit low-dimensional latent spaces. It achieves tighter theoretical error bounds and strong empirical performance in image generation and downstream classification.
@inproceedings{mazumder2024lorae,title={Learning Low-Rank Latent Spaces Using Simple Deterministic Autoencoders: Theoretical \& Empirical Insights},author={Mazumder, Alokendu and Baruah, Tirthajit and Kumar, Bhartendu and Sharma, Rishab and Pattanaik, Vishwajeet and Rathore, Punit},year={2024},booktitle={IEEE/CVF Winter Conference on Applications of Computer Vision},url={https://openaccess.thecvf.com/content/WACV2024/html/Mazumder_Learning_Low-Rank_Latent_Spaces_With_Simple_Deterministic_Autoencoder_Theoretical_and_WACV_2024_paper.html},}
2023
ICCV (W)
DeepVAT: A Self-Supervised Technique for Cluster Assessment in Image Datasets
Alokendu Mazumder, Tirthajit Baruah, Akash Kumar, and 3 more authors
@inproceedings{mazumder2023deepvat,title={DeepVAT: A Self-Supervised Technique for Cluster Assessment in Image Datasets},author={Mazumder, Alokendu and Baruah, Tirthajit and Kumar, Akash and Murthy, Pagadla Krishna and Pattanaik, Vishwajeet and Rathore, Punit},year={2023},booktitle={IEEE/CVF ICCV Workshops 2023}}
2021
ICIP
Perceptual Quality Assessment of DIBR Synthesized Views Using Saliency Based Deep Features
Shubham Chaudhary, Alokendu Mazumder, Deebha Mumtaz, and 2 more authors
In IEEE International Conference on Image Processing, 2021
We propose an efficient no-training perceptual quality metric for DIBR synthesized views that exploits saliency-guided deep features extracted from a pretrained VGG-16 network. By focusing on perceptually important regions and fusing feature maps using cosine similarity, the method captures geometric distortions on standard benchmarks.
@inproceedings{chaudhary2021dibr,title={Perceptual Quality Assessment of DIBR Synthesized Views Using Saliency Based Deep Features},author={Chaudhary, Shubham and Mazumder, Alokendu and Mumtaz, Deebha and Jakhetiya, Vinit and Subudhi, Badri N.},year={2021},booktitle={IEEE International Conference on Image Processing},url={https://ieeexplore.ieee.org/document/9506607},}
OCEANS
Fusion-UWnet: Multi-channel Fusion-based Deep CNN for Underwater Image Enhancement
Pious Pradhan, Alokendu Mazumder, Srimanta Mandal, and 1 more author
@inproceedings{pradhan2021fusion,title={Fusion-UWnet: Multi-channel Fusion-based Deep CNN for Underwater Image Enhancement},author={Pradhan, Pious and Mazumder, Alokendu and Mandal, Srimanta and Subudhi, Badri N.},year={2021},booktitle={IEEE OCEANS}}
under review & work in progress
Draft
On Convergence of Adagrad Under Generalized Smoothness
