Shristi Das Biswas

Shristi Das Biswas, Adarsh Kosta, Chamika Liyanagedera et al.

Event cameras detect changes in per-pixel intensity to generate asynchronous `event streams'. They offer great potential for accurate semantic map retrieval in real-time autonomous systems owing to their much higher temporal resolution and high dynamic range (HDR) compared to conventional cameras. However, existing implementations for event-based segmentation suffer from sub-optimal performance since these temporally dense events only measure the varying component of a visual signal, limiting their ability to encode dense spatial context compared to frames. To address this issue, we propose a hybrid end-to-end learning framework HALSIE, utilizing three key concepts to reduce inference cost by up to $20\times$ versus prior art while retaining similar performance: First, a simple and efficient cross-domain learning scheme to extract complementary spatio-temporal embeddings from both frames and events. Second, a specially designed dual-encoder scheme with Spiking Neural Network (SNN) and Artificial Neural Network (ANN) branches to minimize latency while retaining cross-domain feature aggregation. Third, a multi-scale cue mixer to model rich representations of the fused embeddings. These qualities of HALSIE allow for a very lightweight architecture achieving state-of-the-art segmentation performance on DDD-17, MVSEC, and DSEC-Semantic datasets with up to $33\times$ higher parameter efficiency and favorable inference cost (17.9mJ per cycle). Our ablation study also brings new insights into effective design choices that can prove beneficial for research across other vision tasks.

Shristi Das Biswas, Kaushik Roy

Instruction tuning of large vision-language models (LVLMs) increasingly depends on massive multimodal corpora, yet these datasets contain samples with substantial redundancy, low visual dependency, and highly imbalanced coverage of multimodal reasoning behaviors. As a result, uniform subsampling or naive score-based selection often yields suboptimal training subsets. We introduce MAGIC, a training-free, forward-only coreset selection method designed to construct compact yet behaviorally faithful subsets for multimodal instruction tuning. MAGIC is built on three intrinsic signals extracted from a pretrained VLM: Multimodal Gain, which measures the likelihood improvement obtained from visual input; Bridging Relevance, which captures the sharpness of answer-token grounding over visual tokens; and Skill-Neuron Signatures, which characterize the functional computation elicited by each sample via top-activated feed-forward neurons. MAGIC combines these signals in a three-stage pipeline: filtering low-gain examples, ranking candidates by a normalized quality objective, and performing bucket-wise budget allocation over discrete neuron signatures to preserve latent multimodal skill coverage. This formulation avoids backpropagation, auxiliary selector training, and expensive clustering in continuous activation spaces, while remaining efficient and easily deployable in existing VLMs. Across LLaVA-665K and Vision-Flan datasets, and transfer settings to large target models, LLaVA-1.5-7B and -13B, MAGIC consistently improves over strong baselines under matched 20% budgets: it achieves 100.3% relative performance to full finetuning on LLaVA-665K and 101.6% relative performance on Vision-Flan-186K, while yielding a 73.7% reduction in wall-clock run time.

Shristi Das Biswas, Matthew Shreve, Xuelu Li et al.

Recent advancements in language-guided diffusion models for image editing are often bottle-necked by cumbersome prompt engineering to precisely articulate desired changes. An intuitive alternative calls on guidance from in-the-wild image exemplars to help users bring their imagined edits to life. Contemporary exemplar-based editing methods shy away from leveraging the rich latent space learnt by pre-existing large text-to-image (TTI) models and fall back on training with curated objective functions to achieve the task. Though somewhat effective, this demands significant computational resources and lacks compatibility with diverse base models and arbitrary exemplar count. On further investigation, we also find that these techniques restrict user control to only applying uniform global changes over the entire edited region. In this paper, we introduce a novel framework for progressive exemplar-driven editing with off-the-shelf diffusion models, dubbed PIXELS, to enable customization by providing granular control over edits, allowing adjustments at the pixel or region level. Our method operates solely during inference to facilitate imitative editing, enabling users to draw inspiration from a dynamic number of reference images, or multimodal prompts, and progressively incorporate all the desired changes without retraining or fine-tuning existing TTI models. This capability of fine-grained control opens up a range of new possibilities, including selective modification of individual objects and specifying gradual spatial changes. We demonstrate that PIXELS delivers high-quality edits efficiently, leading to a notable improvement in quantitative metrics as well as human evaluation. By making high-quality image editing more accessible, PIXELS has the potential to enable professional-grade edits to a wider audience with the ease of using any open-source image generation model.

Shristi Das Biswas, Efstathia Soufleri, Arani Roy et al.

Training robust deep video representations has proven to be computationally challenging due to substantial decoding overheads, the enormous size of raw video streams, and their inherent high temporal redundancy. Different from existing schemes, operating exclusively in the compressed video domain and exploiting all freely available modalities, i.e., I-frames, and P-frames (motion vectors and residuals) offers a compute-efficient alternative. Existing methods approach this task as a naive multi-modality problem, ignoring the temporal correlation and implicit sparsity across P-frames for modeling stronger shared representations for videos of the same action, making training and generalization easier. By revisiting the high-level design of dominant video understanding backbones, we increase inference speed by a factor of $56$ while retaining similar performance. For this, we propose a hybrid end-to-end framework that factorizes learning across three key concepts to reduce inference cost by $330\times$ versus prior art: First, a specially designed dual-encoder scheme with efficient Spiking Temporal Modulators to minimize latency while retaining cross-domain feature aggregation. Second, a unified transformer model to capture inter-modal dependencies using global self-attention to enhance I-frame -- P-frame contextual interactions. Third, a Multi-Modal Mixer Block to model rich representations from the joint spatiotemporal token embeddings. Experiments show that our method results in a lightweight architecture achieving state-of-the-art video recognition performance on UCF-101, HMDB-51, K-400, K-600 and SS-v2 datasets with favorable costs ($0.73$J/V) and fast inference ($16$V/s). Our observations bring new insights into practical design choices for efficient next-generation spatiotemporal learners. Code is available.

Shristi Das Biswas, Yue Zhang, Anwesan Pal et al.

Large Language Models (LLMs) suffer severe catastrophic forgetting when adapted sequentially to new tasks in a continual learning (CL) setting. Existing approaches are fundamentally limited: replay-based methods are impractical and privacy-violating, while strict orthogonality-based methods collapse under scale: each new task is projected onto an orthogonal complement, progressively reducing the residual degrees of freedom and eliminating forward transfer by forbidding overlap in shared representations. In this work, we introduce ELLA, a training framework built on the principle of selective subspace de-correlation. Rather than forbidding all overlap, ELLA explicitly characterizes the structure of past updates and penalizes alignments along their high-energy, task-specific directions, while preserving freedom in the low-energy residual subspaces to enable transfer. Formally, this is realized via a lightweight regularizer on a single aggregated update matrix. We prove this mechanism corresponds to an anisotropic shrinkage operator that bounds interference, yielding a penalty that is both memory- and compute-constant regardless of task sequence length. ELLA requires no data replay, no architectural expansion, and negligible storage. Empirically, it achieves state-of-the-art CL performance on three popular benchmarks, with relative accuracy gains of up to $9.6\%$ and a $35\times$ smaller memory footprint. Further, ELLA scales robustly across architectures and actively enhances the model's zero-shot generalization performance on unseen tasks, establishing a principled and scalable solution for constructive lifelong LLM adaptation.

Arani Roy, Shristi Das Biswas, Kaushik Roy

Text-to-image diffusion models can generate visually stunning images, yet, controlling what appears and how it appears, remains surprisingly difficult, especially when operating solely within the constraints of the text-conditioning space. For example, changing a subject or adjusting an attribute often leads to unintended side effects, such as altered backgrounds or distorted details. This is because most existing text-based control methods treat the embedding space as Euclidean and apply simple linear transformations, which do not reflect how semantic concepts are actually organized. In this work, we take a step back and ask: what is the true geometry of these embeddings? We find that text encoder representations lie on a hypersphere, where concepts are not linear directions but structured, anisotropic distributions better captured by Kent distributions. Building on this insight, we propose HEART, a training-free framework that performs Kent-aware geodesic transformations directly on the hypersphere. By respecting the underlying geometry, HEART enables intuitive and precise edits, such as consistent subject replacement and fine-grained attribute control, while preserving the original scene. Importantly, HEART requires no finetuning, inversion, or optimization, and generalizes across diffusion model architectures. Our results show that a simple shift in perspective, from linear to spherical, can unlock fast, and controllable image generation.

Shristi Das Biswas, Arani Roy, Kaushik Roy

Robust concept removal for text-to-image (T2I) and text-to-video (T2V) models is essential for their safe deployment. Existing methods, however, suffer from costly retraining, inference overhead, or vulnerability to adversarial attacks. Crucially, they rarely model the latent semantic overlap between the target erase concept and surrounding content -- causing collateral damage post-erasure -- and even fewer methods work reliably across both T2I and T2V domains. We introduce Instant Concept Erasure (ICE), a training-free, modality-agnostic, one-shot weight modification approach that achieves precise, persistent unlearning with zero overhead. ICE defines erase and preserve subspaces using anisotropic energy-weighted scaling, then explicitly regularises against their intersection using a unique, closed-form overlap projector. We pose a convex and Lipschitz-bounded Spectral Unlearning Objective, balancing erasure fidelity and intersection preservation, that admits a stable and unique analytical solution. This solution defines a dissociation operator that is translated to the model's text-conditioning layers, making the edit permanent and runtime-free. Across targeted removals of artistic styles, objects, identities, and explicit content, ICE efficiently achieves strong erasure with improved robustness to red-teaming, all while causing only minimal degradation of original generative abilities in both T2I and T2V models.

Arani Roy, Marco P. Apolinario, Shristi Das Biswas et al.

Training deep neural networks (DNNs) with backpropagation (BP) achieves state-of-the-art accuracy but requires global error propagation and full parameterization, leading to substantial memory and computational overhead. Direct Feedback Alignment (DFA) enables local, parallelizable updates with lower memory requirements but is limited by unstructured feedback and poor scalability in deeper architectures, specially convolutional neural networks. To address these limitations, we propose a structured local learning framework that operates directly on low-rank manifolds defined by the Singular Value Decomposition (SVD) of weight matrices. Each layer is trained in its decomposed form, with updates applied to the SVD components using a composite loss that integrates cross-entropy, subspace alignment, and orthogonality regularization. Feedback matrices are constructed to match the SVD structure, ensuring consistent alignment between forward and feedback pathways. Our method reduces the number of trainable parameters relative to the original DFA model, without relying on pruning or post hoc compression. Experiments on CIFAR-10, CIFAR-100, and ImageNet show that our method achieves accuracy comparable to that of BP. Ablation studies confirm the importance of each loss term in the low-rank setting. These results establish local learning on low-rank manifolds as a principled and scalable alternative to full-rank gradient-based training.

Arani Roy, Shristi Das Biswas, Kaushik Roy

Diffusion models (DMs), lauded for their generative performance, are computationally prohibitive due to their billion-scale parameters and iterative denoising dynamics. Existing efficiency techniques, such as quantization, timestep reduction, or pruning, offer savings in compute, memory, or runtime but are strictly bottlenecked by reliance on fine-tuning or retraining to recover performance. In this work, we introduce SlimDiff, an automated activation-informed structural compression framework that reduces both attention and feedforward dimensionalities in DMs, while being entirely gradient-free. SlimDiff reframes DM compression as a spectral approximation task, where activation covariances across denoising timesteps define low-rank subspaces that guide dynamic pruning under a fixed compression budget. This activation-aware formulation mitigates error accumulation across timesteps by applying module-wise decompositions over functional weight groups: query--key interactions, value--output couplings, and feedforward projections, rather than isolated matrix factorizations, while adaptively allocating sparsity across modules to respect the non-uniform geometry of diffusion trajectories. SlimDiff achieves up to 35\% acceleration and $\sim$100M parameter reduction over baselines, with generation quality on par with uncompressed models without any backpropagation. Crucially, our approach requires only about 500 calibration samples, over 70$\times$ fewer than prior methods. To our knowledge, this is the first closed-form, activation-guided structural compression of DMs that is entirely training-free, providing both theoretical clarity and practical efficiency.

Shristi Das Biswas, Arani Roy, Kaushik Roy

As Text-to-Image models continue to evolve, so does the risk of generating unsafe, copyrighted, or privacy-violating content. Existing safety interventions - ranging from training data curation and model fine-tuning to inference-time filtering and guidance - often suffer from incomplete concept removal, susceptibility to jail-breaking, computational inefficiency, or collateral damage to unrelated capabilities. In this paper, we introduce CURE, a training-free concept unlearning framework that operates directly in the weight space of pre-trained diffusion models, enabling fast, interpretable, and highly specific suppression of undesired concepts. At the core of our method is the Spectral Eraser, a closed-form, orthogonal projection module that identifies discriminative subspaces using Singular Value Decomposition over token embeddings associated with the concepts to forget and retain. Intuitively, the Spectral Eraser identifies and isolates features unique to the undesired concept while preserving safe attributes. This operator is then applied in a single step update to yield an edited model in which the target concept is effectively unlearned - without retraining, supervision, or iterative optimization. To balance the trade-off between filtering toxicity and preserving unrelated concepts, we further introduce an Expansion Mechanism for spectral regularization which selectively modulates singular vectors based on their relative significance to control the strength of forgetting. All the processes above are in closed-form, guaranteeing extremely efficient erasure in only $2$ seconds. Benchmarking against prior approaches, CURE achieves a more efficient and thorough removal for targeted artistic styles, objects, identities, or explicit content, with minor damage to original generation ability and demonstrates enhanced robustness against red-teaming.