Arnab Kumar Mondal

Sékou-Oumar Kaba, Arnab Kumar Mondal, Yan Zhang et al.

Symmetry-based neural networks often constrain the architecture in order to achieve invariance or equivariance to a group of transformations. In this paper, we propose an alternative that avoids this architectural constraint by learning to produce canonical representations of the data. These canonicalization functions can readily be plugged into non-equivariant backbone architectures. We offer explicit ways to implement them for some groups of interest. We show that this approach enjoys universality while providing interpretable insights. Our main hypothesis, supported by our empirical results, is that learning a small neural network to perform canonicalization is better than using predefined heuristics. Our experiments show that learning the canonicalization function is competitive with existing techniques for learning equivariant functions across many tasks, including image classification, $N$-body dynamics prediction, point cloud classification and part segmentation, while being faster across the board.

Ayush Kaushal, Tejas Vaidhya, Arnab Kumar Mondal et al.

Rapid advancements in GPU computational power has outpaced memory capacity and bandwidth growth, creating bottlenecks in Large Language Model (LLM) inference. Post-training quantization is the leading method for addressing memory-related bottlenecks in LLM inference, but it suffers from significant performance degradation below 4-bit precision. This paper addresses these challenges by investigating the pretraining of low-bitwidth models specifically Ternary Language Models (TriLMs) as an alternative to traditional floating-point models (FloatLMs) and their post-training quantized versions (QuantLMs). We present Spectra LLM suite, the first open suite of LLMs spanning multiple bit-widths, including FloatLMs, QuantLMs, and TriLMs, ranging from 99M to 3.9B parameters trained on 300B tokens. Our comprehensive evaluation demonstrates that TriLMs offer superior scaling behavior in terms of model size (in bits). Surprisingly, at scales exceeding one billion parameters, TriLMs consistently outperform their QuantLM and FloatLM counterparts for a given bit size across various benchmarks. Notably, the 3.9B parameter TriLM matches the performance of the FloatLM 3.9B across all benchmarks, despite having fewer bits than FloatLM 830M. Overall, this research provides valuable insights into the feasibility and scalability of low-bitwidth language models, paving the way for the development of more efficient LLMs. To enhance understanding of low-bitwidth models, we are releasing 500+ intermediate checkpoints of the Spectra suite at https://github.com/NolanoOrg/SpectraSuite.

Arnab Kumar Mondal, Siba Smarak Panigrahi, Sékou-Oumar Kaba et al.

Equivariant networks are specifically designed to ensure consistent behavior with respect to a set of input transformations, leading to higher sample efficiency and more accurate and robust predictions. However, redesigning each component of prevalent deep neural network architectures to achieve chosen equivariance is a difficult problem and can result in a computationally expensive network during both training and inference. A recently proposed alternative towards equivariance that removes the architectural constraints is to use a simple canonicalization network that transforms the input to a canonical form before feeding it to an unconstrained prediction network. We show here that this approach can effectively be used to make a large pretrained network equivariant. However, we observe that the produced canonical orientations can be misaligned with those of the training distribution, hindering performance. Using dataset-dependent priors to inform the canonicalization function, we are able to make large pretrained models equivariant while maintaining their performance. This significantly improves the robustness of these models to deterministic transformations of the data, such as rotations. We believe this equivariant adaptation of large pretrained models can help their domain-specific applications with known symmetry priors.

Arnab Kumar Mondal, Siba Smarak Panigrahi, Sai Rajeswar et al.

The accurate modeling of dynamics in interactive environments is critical for successful long-range prediction. Such a capability could advance Reinforcement Learning (RL) and Planning algorithms, but achieving it is challenging. Inaccuracies in model estimates can compound, resulting in increased errors over long horizons. We approach this problem from the lens of Koopman theory, where the nonlinear dynamics of the environment can be linearized in a high-dimensional latent space. This allows us to efficiently parallelize the sequential problem of long-range prediction using convolution while accounting for the agent's action at every time step. Our approach also enables stability analysis and better control over gradients through time. Taken together, these advantages result in significant improvement over the existing approaches, both in the efficiency and the accuracy of modeling dynamics over extended horizons. We also show that this model can be easily incorporated into dynamics modeling for model-based planning and model-free RL and report promising experimental results.

Sanket Gandhi, Atul, Samanyu Mahajan et al.

Recent work has shown that object-centric representations can greatly help improve the accuracy of learning dynamics while also bringing interpretability. In this work, we take this idea one step further, ask the following question: "can learning disentangled representation further improve the accuracy of visual dynamics prediction in object-centric models?" While there has been some attempt to learn such disentangled representations for the case of static images \citep{nsb}, to the best of our knowledge, ours is the first work which tries to do this in a general setting for video, without making any specific assumptions about the kind of attributes that an object might have. The key building block of our architecture is the notion of a {\em block}, where several blocks together constitute an object. Each block is represented as a linear combination of a given number of learnable concept vectors, which is iteratively refined during the learning process. The blocks in our model are discovered in an unsupervised manner, by attending over object masks, in a style similar to discovery of slots \citep{slot_attention}, for learning a dense object-centric representation. We employ self-attention via transformers over the discovered blocks to predict the next state resulting in discovery of visual dynamics. We perform a series of experiments on several benchmark 2-D, and 3-D datasets demonstrating that our architecture (1) can discover semantically meaningful blocks (2) help improve accuracy of dynamics prediction compared to SOTA object-centric models (3) perform significantly better in OOD setting where the specific attribute combinations are not seen earlier during training. Our experiments highlight the importance discovery of disentangled representation for visual dynamics prediction.

Ayush Tripathi, Arnab Kumar Mondal, Lalan Kumar et al.

Airwriting Recognition is the problem of identifying letters written in free space with finger movement. It is essentially a specialized case of gesture recognition, wherein the vocabulary of gestures corresponds to letters as in a particular language. With the wide adoption of smart wearables in the general population, airwriting recognition using motion sensors from a smart-band can be used as a medium of user input for applications in Human-Computer Interaction. There has been limited work in the recognition of in-air trajectories using motion sensors, and the performance of the techniques in the case when the device used to record signals is changed has not been explored hitherto. Motivated by these, a new paradigm for device and user-independent airwriting recognition based on supervised contrastive learning is proposed. A two stage classification strategy is employed, the first of which involves training an encoder network with supervised contrastive loss. In the subsequent stage, a classification head is trained with the encoder weights kept frozen. The efficacy of the proposed method is demonstrated through experiments on a publicly available dataset and also with a dataset recorded in our lab using a different device. Experiments have been performed in both supervised and unsupervised settings and compared against several state-of-the-art domain adaptation techniques. Data and the code for our implementation will be made available at https://github.com/ayushayt/SCLAiR.

Arnab Kumar Mondal, Jose Dolz, Christian Desrosiers

We address the problem of segmenting 3D multi-modal medical images in scenarios where very few labeled examples are available for training. Leveraging the recent success of adversarial learning for semi-supervised segmentation, we propose a novel method based on Generative Adversarial Networks (GANs) to train a segmentation model with both labeled and unlabeled images. The proposed method prevents over-fitting by learning to discriminate between true and fake patches obtained by a generator network. Our work extends current adversarial learning approaches, which focus on 2D single-modality images, to the more challenging context of 3D volumes of multiple modalities. The proposed method is evaluated on the problem of segmenting brain MRI from the iSEG-2017 and MRBrainS 2013 datasets. Significant performance improvement is reported, compared to state-of-art segmentation networks trained in a fully-supervised manner. In addition, our work presents a comprehensive analysis of different GAN architectures for semi-supervised segmentation, showing recent techniques like feature matching to yield a higher performance than conventional adversarial training approaches. Our code is publicly available at https://github.com/arnab39/FewShot_GAN-Unet3D

Siba Smarak Panigrahi, Arnab Kumar Mondal

This work introduces a novel approach to achieving architecture-agnostic equivariance in deep learning, particularly addressing the limitations of traditional layerwise equivariant architectures and the inefficiencies of the existing architecture-agnostic methods. Building equivariant models using traditional methods requires designing equivariant versions of existing models and training them from scratch, a process that is both impractical and resource-intensive. Canonicalization has emerged as a promising alternative for inducing equivariance without altering model architecture, but it suffers from the need for highly expressive and expensive equivariant networks to learn canonical orientations accurately. We propose a new optimization-based method that employs any non-equivariant network for canonicalization. Our method uses contrastive learning to efficiently learn a canonical orientation and offers more flexibility for the choice of canonicalization network. We empirically demonstrate that this approach outperforms existing methods in achieving equivariance for large pretrained models and significantly speeds up the canonicalization process, making it up to 2 times faster.

Jay Nandy, Arnab Kumar Mondal, Anuj Rathore et al.

Large graph datasets make training graph neural networks (GNNs) computationally costly. Graph condensation methods address this by generating small synthetic graphs that approximate the original data. However, existing approaches rely on clean, supervised labels, which limits their reliability when labels are scarce, noisy, or inconsistent. We propose Pseudo-Labeled Graph Condensation (PLGC), a self-supervised framework that constructs latent pseudo-labels from node embeddings and optimizes condensed graphs to match the original graph's structural and feature statistics -- without requiring ground-truth labels. PLGC offers three key contributions: (1) A diagnosis of why supervised condensation fails under label noise and distribution shift. (2) A label-free condensation method that jointly learns latent prototypes and node assignments. (3) Theoretical guarantees showing that pseudo-labels preserve latent structural statistics of the original graph and ensure accurate embedding alignment. Empirically, across node classification and link prediction tasks, PLGC achieves competitive performance with state-of-the-art supervised condensation methods on clean datasets and exhibits substantial robustness under label noise, often outperforming all baselines by a significant margin. Our findings highlight the practical and theoretical advantages of self-supervised graph condensation in noisy or weakly-labeled environments.

Arnab Kumar Mondal, Stefano Alletto, Denis Tome

Understanding human motion from video is essential for a range of applications, including pose estimation, mesh recovery and action recognition. While state-of-the-art methods predominantly rely on transformer-based architectures, these approaches have limitations in practical scenarios. Transformers are slower when sequentially predicting on a continuous stream of frames in real-time, and do not generalize to new frame rates. In light of these constraints, we propose a novel attention-free spatiotemporal model for human motion understanding building upon recent advancements in state space models. Our model not only matches the performance of transformer-based models in various motion understanding tasks but also brings added benefits like adaptability to different video frame rates and enhanced training speed when working with longer sequence of keypoints. Moreover, the proposed model supports both offline and real-time applications. For real-time sequential prediction, our model is both memory efficient and several times faster than transformer-based approaches while maintaining their high accuracy.

Kusha Sareen, Daniel Levy, Arnab Kumar Mondal et al.

Generative modeling of symmetric densities has a range of applications in AI for science, from drug discovery to physics simulations. The existing generative modeling paradigm for invariant densities combines an invariant prior with an equivariant generative process. However, we observe that this technique is not necessary and has several drawbacks resulting from the limitations of equivariant networks. Instead, we propose to model a learned slice of the density so that only one representative element per orbit is learned. To accomplish this, we learn a group-equivariant canonicalization network that maps training samples to a canonical pose and train a non-equivariant generative model over these canonicalized samples. We implement this idea in the context of diffusion models. Our preliminary experimental results on molecular modeling are promising, demonstrating improved sample quality and faster inference time.

Sahar Dastani, Ali Bahri, Moslem Yazdanpanah et al.

State Space Models (SSMs) have recently emerged as an alternative to Vision Transformers (ViTs) due to their unique ability of modeling global relationships with linear complexity. SSMs are specifically designed to capture spatially proximate relationships of image patches. However, they fail to identify relationships between conceptually related yet not adjacent patches. This limitation arises from the non-causal nature of image data, which lacks inherent directional relationships. Additionally, current vision-based SSMs are highly sensitive to transformations such as rotation. Their predefined scanning directions depend on the original image orientation, which can cause the model to produce inconsistent patch-processing sequences after rotation. To address these limitations, we introduce Spectral VMamba, a novel approach that effectively captures the global structure within an image by leveraging spectral information derived from the graph Laplacian of image patches. Through spectral decomposition, our approach encodes patch relationships independently of image orientation, achieving rotation invariance with the aid of our Rotational Feature Normalizer (RFN) module. Our experiments on classification tasks show that Spectral VMamba outperforms the leading SSM models in vision, such as VMamba, while maintaining invariance to rotations and a providing a similar runtime efficiency.

Harman Singh, Poorva Garg, Mohit Gupta et al.

We are interested in image manipulation via natural language text -- a task that is useful for multiple AI applications but requires complex reasoning over multi-modal spaces. We extend recently proposed Neuro Symbolic Concept Learning (NSCL), which has been quite effective for the task of Visual Question Answering (VQA), for the task of image manipulation. Our system referred to as NeuroSIM can perform complex multi-hop reasoning over multi-object scenes and only requires weak supervision in the form of annotated data for VQA. NeuroSIM parses an instruction into a symbolic program, based on a Domain Specific Language (DSL) comprising of object attributes and manipulation operations, that guides its execution. We create a new dataset for the task, and extensive experiments demonstrate that NeuroSIM is highly competitive with or beats SOTA baselines that make use of supervised data for manipulation.

Mehran Shakerinava, Arnab Kumar Mondal, Siamak Ravanbakhsh

We present a simple non-generative approach to deep representation learning that seeks equivariant deep embedding through simple objectives. In contrast to existing equivariant networks, our transformation coding approach does not constrain the choice of the feed-forward layer or the architecture and allows for an unknown group action on the input space. We introduce several such transformation coding objectives for different Lie groups such as the Euclidean, Orthogonal and the Unitary groups. When using product groups, the representation is decomposed and disentangled. We show that the presence of additional information on different transformations improves disentanglement in transformation coding. We evaluate the representations learnt by transformation coding both qualitatively and quantitatively on downstream tasks, including reinforcement learning.

Arna Ghosh, Arnab Kumar Mondal, Kumar Krishna Agrawal et al.

Representation learning that leverages large-scale labelled datasets, is central to recent progress in machine learning. Access to task relevant labels at scale is often scarce or expensive, motivating the need to learn from unlabelled datasets with self-supervised learning (SSL). Such large unlabelled datasets (with data augmentations) often provide a good coverage of the underlying input distribution. However evaluating the representations learned by SSL algorithms still requires task-specific labelled samples in the training pipeline. Additionally, the generalization of task-specific encoding is often sensitive to potential distribution shift. Inspired by recent advances in theoretical machine learning and vision neuroscience, we observe that the eigenspectrum of the empirical feature covariance matrix often follows a power law. For visual representations, we estimate the coefficient of the power law, $α$, across three key attributes which influence representation learning: learning objective (supervised, SimCLR, Barlow Twins and BYOL), network architecture (VGG, ResNet and Vision Transformer), and tasks (object and scene recognition). We observe that under mild conditions, proximity of $α$ to 1, is strongly correlated to the downstream generalization performance. Furthermore, $α\approx 1$ is a strong indicator of robustness to label noise during fine-tuning. Notably, $α$ is computable from the representations without knowledge of any labels, thereby offering a framework to evaluate the quality of representations in unlabelled datasets.

Arnab Kumar Mondal, Himanshu Asnani, Parag Singla et al.

Clustering single-cell RNA sequence (scRNA-seq) data poses statistical and computational challenges due to their high-dimensionality and data-sparsity, also known as `dropout' events. Recently, Regularized Auto-Encoder (RAE) based deep neural network models have achieved remarkable success in learning robust low-dimensional representations. The basic idea in RAEs is to learn a non-linear mapping from the high-dimensional data space to a low-dimensional latent space and vice-versa, simultaneously imposing a distributional prior on the latent space, which brings in a regularization effect. This paper argues that RAEs suffer from the infamous problem of bias-variance trade-off in their naive formulation. While a simple AE without a latent regularization results in data over-fitting, a very strong prior leads to under-representation and thus bad clustering. To address the above issues, we propose a modified RAE framework (called the scRAE) for effective clustering of the single-cell RNA sequencing data. scRAE consists of deterministic AE with a flexibly learnable prior generator network, which is jointly trained with the AE. This facilitates scRAE to trade-off better between the bias and variance in the latent space. We demonstrate the efficacy of the proposed method through extensive experimentation on several real-world single-cell Gene expression datasets.

Arnab Kumar Mondal, Vineet Jain, Kaleem Siddiqi

Current deep learning models for classification tasks in computer vision are trained using mini-batches. In the present article, we take advantage of the relationships between samples in a mini-batch, using graph neural networks to aggregate information from similar images. This helps mitigate the adverse effects of alterations to the input images on classification performance. Diverse experiments on image-based object and scene classification show that this approach not only improves a classifier's performance but also increases its robustness to image perturbations and adversarial attacks. Further, we also show that mini-batch graph neural networks can help to alleviate the problem of mode collapse in Generative Adversarial Networks.

Arnab Kumar Mondal, Prathosh A. P

Voice Activity Detection (VAD) refers to the task of identification of regions of human speech in digital signals such as audio and video. While VAD is a necessary first step in many speech processing systems, it poses challenges when there are high levels of ambient noise during the audio recording. To improve the performance of VAD in such conditions, several methods utilizing the visual information extracted from the region surrounding the mouth/lip region of the speakers' video recording have been proposed. Even though these provide advantages over audio-only methods, they depend on faithful extraction of lip/mouth regions. Motivated by these, a new paradigm for VAD based on the fact that respiration forms the primary source of energy for speech production is proposed. Specifically, an audio-independent VAD technique using the respiration pattern extracted from the speakers' video is developed. The Respiration Pattern is first extracted from the video focusing on the abdominal-thoracic region of a speaker using an optical flow based method. Subsequently, voice activity is detected from the respiration pattern signal using neural sequence-to-sequence prediction models. The efficacy of the proposed method is demonstrated through experiments on a challenging dataset recorded in real acoustic environments and compared with four previous methods based on audio and visual cues.

Arnab Kumar Mondal, Pratheeksha Nair, Kaleem Siddiqi

In Reinforcement Learning (RL), Convolutional Neural Networks(CNNs) have been successfully applied as function approximators in Deep Q-Learning algorithms, which seek to learn action-value functions and policies in various environments. However, to date, there has been little work on the learning of symmetry-transformation equivariant representations of the input environment state. In this paper, we propose the use of Equivariant CNNs to train RL agents and study their inductive bias for transformation equivariant Q-value approximation. We demonstrate that equivariant architectures can dramatically enhance the performance and sample efficiency of RL agents in a highly symmetric environment while requiring fewer parameters. Additionally, we show that they are robust to changes in the environment caused by affine transformations.

Arnab Kumar Mondal, Himanshu Asnani, Parag Singla et al.

Regularized Auto-Encoders (RAEs) form a rich class of neural generative models. They effectively model the joint-distribution between the data and the latent space using an Encoder-Decoder combination, with regularization imposed in terms of a prior over the latent space. Despite their advantages, such as stability in training, the performance of AE based models has not reached the superior standards as that of the other generative models such as Generative Adversarial Networks (GANs). Motivated by this, we examine the effect of the latent prior on the generation quality of deterministic AE models in this paper. Specifically, we consider the class of RAEs with deterministic Encoder-Decoder pairs, Wasserstein Auto-Encoders (WAE), and show that having a fixed prior distribution, \textit{a priori}, oblivious to the dimensionality of the `true' latent space, will lead to the infeasibility of the optimization problem considered. Further, we show that, in the finite data regime, despite knowing the correct latent dimensionality, there exists a bias-variance trade-off with any arbitrary prior imposition. As a remedy to both the issues mentioned above, we introduce an additional state space in the form of flexibly learnable latent priors, in the optimization objective of the WAEs. We implicitly learn the distribution of the latent prior jointly with the AE training, which not only makes the learning objective feasible but also facilitates operation on different points of the bias-variance curve. We show the efficacy of our model, called FlexAE, through several experiments on multiple datasets, and demonstrate that it is the new state-of-the-art for the AE based generative models.

Arnab Kumar Mondal, Arnab Bhattacharya, Sudipto Mukherjee et al.

Estimation of information theoretic quantities such as mutual information and its conditional variant has drawn interest in recent times owing to their multifaceted applications. Newly proposed neural estimators for these quantities have overcome severe drawbacks of classical $k$NN-based estimators in high dimensions. In this work, we focus on conditional mutual information (CMI) estimation by utilizing its formulation as a minmax optimization problem. Such a formulation leads to a joint training procedure similar to that of generative adversarial networks. We find that our proposed estimator provides better estimates than the existing approaches on a variety of simulated data sets comprising linear and non-linear relations between variables. As an application of CMI estimation, we deploy our estimator for conditional independence (CI) testing on real data and obtain better results than state-of-the-art CI testers.

Arnab Kumar Mondal, Sankalan Pal Chowdhury, Aravind Jayendran et al.

The field of neural generative models is dominated by the highly successful Generative Adversarial Networks (GANs) despite their challenges, such as training instability and mode collapse. Auto-Encoders (AE) with regularized latent space provide an alternative framework for generative models, albeit their performance levels have not reached that of GANs. In this work, we hypothesise that the dimensionality of the AE model's latent space has a critical effect on the quality of generated data. Under the assumption that nature generates data by sampling from a "true" generative latent space followed by a deterministic function, we show that the optimal performance is obtained when the dimensionality of the latent space of the AE-model matches with that of the "true" generative latent space. Further, we propose an algorithm called the Mask Adversarial Auto-Encoder (MaskAAE), in which the dimensionality of the latent space of an adversarial auto encoder is brought closer to that of the "true" generative latent space, via a procedure to mask the spurious latent dimensions. We demonstrate through experiments on synthetic and several real-world datasets that the proposed formulation yields betterment in the generation quality.

Arnab Kumar Mondal, Aniket Agarwal, Jose Dolz et al.

In this work, we study the problem of training deep networks for semantic image segmentation using only a fraction of annotated images, which may significantly reduce human annotation efforts. Particularly, we propose a strategy that exploits the unpaired image style transfer capabilities of CycleGAN in semi-supervised segmentation. Unlike recent works using adversarial learning for semi-supervised segmentation, we enforce cycle consistency to learn a bidirectional mapping between unpaired images and segmentation masks. This adds an unsupervised regularization effect that boosts the segmentation performance when annotated data is limited. Experiments on three different public segmentation benchmarks (PASCAL VOC 2012, Cityscapes and ACDC) demonstrate the effectiveness of the proposed method. The proposed model achieves 2-4% of improvement with respect to the baseline and outperforms recent approaches for this task, particularly in low labeled data regime.