Dhruv Parikh

Dhruv Parikh, Jacob Fein-Ashley, Rajgopal Kannan et al.

Large Multimodal Models (LMMs) such as LLaVA are typically trained with an autoregressive language modeling objective, providing only indirect supervision to visual tokens. This often yields weak internal visual representations and brittle behavior under distribution shift. Inspired by recent progress on latent denoising for learning high-quality visual tokenizers, we show that the same principle provides an effective form of visual supervision for improving internal visual feature alignment and multimodal understanding in LMMs. We propose a latent denoising framework that corrupts projected visual tokens using a saliency-aware mixture of masking and Gaussian noising. The LMM is trained to denoise these corrupted tokens by recovering clean teacher patch features from hidden states at a selected intermediate LLM layer using a decoder. To prevent representation collapse, our framework also preserves the teacher's intra-image similarity structure and applies intra-image contrastive patch distillation. During inference, corruption and auxiliary heads are disabled, introducing no additional inference-time overhead. Across a broad suite of standard multimodal benchmarks, our method consistently improves visual understanding and reasoning over strong baselines, and yields clear gains on compositional robustness benchmarks (e.g., NaturalBench). Moreover, under ImageNet-C-style non-adversarial common corruptions applied to benchmark images, our method maintains higher accuracy and exhibits reduced degradation at both moderate and severe corruption levels. Our code is available at https://github.com/dhruvashp/latent-denoising-for-lmms.

Jebacyril Arockiaraj, Dhruv Parikh, Viktor Prasanna

Real-time, energy-efficient inference on edge devices is essential for graph classification across a range of applications. Hyperdimensional Computing (HDC) is a brain-inspired computing paradigm that encodes input features into low-precision, high-dimensional vectors with simple element-wise operations, making it well-suited for resource-constrained edge platforms. Recent work enhances HDC accuracy for graph classification via Nyström kernel approximations. Edge acceleration of such methods faces several challenges: (i) redundancy among (landmark) samples selected via uniform sampling, (ii) storing the Nyström projection matrix under limited on-chip memory, (iii) expensive, contention-prone codebook lookups, and (iv) load imbalance due to irregular sparsity in SpMV. To address these challenges, we propose HyperX, the first end-to-end FPGA accelerator for Nyström-based HDC graph classification at the edge. HyperX integrates four key optimizations: (i) a hybrid landmark selection strategy combining uniform sampling with determinantal point processes (DPPs) to reduce redundancy while improving accuracy; (ii) a streaming architecture for Nyström projection matrix maximizing external memory bandwidth utilization; (iii) a minimal-perfect-hash lookup engine enabling $O(1)$ key-to-index mapping; and (iv) sparsity-aware SpMV engines with static load balancing. Implemented on an AMD Zynq UltraScale+ (ZCU104) FPGA, HyperX achieves $6.85\times$ ($4.32\times$) speedup and $169\times$ ($314\times$) energy efficiency gains over optimized CPU (GPU) baselines, while improving classification accuracy by $3.4\%$ on average across TUDataset benchmarks, a widely used standard for graph classification.

Gangda Deng, Zhaoling Chen, Zhongming Yu et al.

With AI agents increasingly deployed as long-running systems, it becomes essential to autonomously construct and continuously evolve customized software to enable interaction within dynamic environments. Yet, existing benchmarks evaluate agents on isolated, one-off coding tasks, neglecting the temporal dependencies and technical debt inherent in real-world software evolution. To bridge this gap, we introduce DeepCommit, an agentic pipeline that reconstructs verifiable Milestone DAGs from noisy commit logs, where milestones are defined as semantically cohesive development goals. These executable sequences enable EvoClaw, a novel benchmark that requires agents to sustain system integrity and limit error accumulation, dimensions of long-term software evolution largely missing from current benchmarks. Our evaluation of 12 frontier models across 4 agent frameworks reveals a critical vulnerability: overall performance scores drop significantly from $>$80% on isolated tasks to at most 38% in continuous settings, exposing agents' profound struggle with long-term maintenance and error propagation.

Anvitha Ramachandran, Dhruv Parikh, Viktor Prasanna

Vision Graph Neural Networks (ViGs) represent an image as a graph of patch tokens, enabling adaptive, feature-driven neighborhoods. Unlike CNNs with fixed grid biases or Vision Transformers with global token interactions, ViGs rely on dynamic graph convolution: at each layer, a feature-dependent graph is built via k-nearest-neighbor (kNN) search on current patch features, followed by message passing. This per-layer graph construction is the main bottleneck, consuming 50--95\% of graph convolution time on CPUs and GPUs, scaling as $O(N^2)$ with the number of patches $N$, and creating a sequential dependency between graph construction and feature updates. We introduce GraphLeap, a simple reformulation that removes this dependency by decoupling graph construction from feature update across layers. GraphLeap performs the feature update at layer $\ell$ using a graph built from the previous layer's features, while simultaneously using the current layer's features to construct the graph for layer $\ell+1$. This one-layer-lookahead graph construction enables concurrent graph construction and message passing. Although using prior-layer features can introduce minor accuracy degradation, lightweight fine-tuning for a few epochs is sufficient to recover the original accuracy. Building on GraphLeap, we present the first end-to-end FPGA accelerator for Vision GNNs. Our streaming, layer-pipelined design overlaps a kNN graph construction engine with a feature update engine, exploits node- and channel-level parallelism, and enables efficient on-chip dataflow without explicit edge-feature materialization. Evaluated on isotropic and pyramidal ViG models on an Alveo U280 FPGA, GraphLeap achieves up to $95.7\times$ speedup over CPU and $8.5\times$ speedup over GPU baselines, demonstrating the feasibility of real-time Vision GNN inference.

Jebacyril Arockiaraj, Dhruv Parikh, Viktor Prasanna

On-device continual learning (CL) is critical for edge AI systems operating on non-stationary data streams, but most existing methods rely on backpropagation or exemplar-heavy classifiers, incurring substantial compute, memory, and latency overheads. Hyperdimensional computing (HDC) offers a lightweight alternative through fast, non-iterative online updates. Combined with a compact convolutional neural network (CNN) feature extractor, HDC enables efficient on-device adaptation with strong visual representations. However, prior HDC-based CL systems often depend on multi-tier memory hierarchies and complex cluster management, limiting deployability on resource-constrained hardware. We present ImageHD, an FPGA accelerator for on-device continual learning of visual data based on HDC. ImageHD targets streaming CL under strict latency and on-chip memory constraints, avoiding costly iterative optimization. At the algorithmic level, we introduce a hardware-aware CL method that bounds class exemplars through a unified exemplar memory and a hardware-efficient cluster merging strategy, while incorporating a quantized CNN front-end to reduce deployment overhead without sacrificing accuracy. At the system level, ImageHD is implemented as a streaming dataflow architecture on the AMD Zynq ZCU104 FPGA, integrating HDC encoding, similarity search, and bounded cluster management using word-packed binary hypervectors for massively parallel bitwise computation within tight on-chip resource budgets. On CORe50, ImageHD achieves up to 40.4x (4.84x) speedup and 383x (105.1x) energy efficiency over optimized CPU (GPU) baselines, demonstrating the practicality of HDC-enabled continual learning for real-time edge AI.

Dhruv Parikh, Anvitha Ramachandran, Haoyang Fan et al.

Vision state space models inherit the efficiency and long-range modeling ability of Mamba-style selective scans. However, their performance depends critically on the representation of two-dimensional visual features as one-dimensional token sequences. Existing scan operators range from predefined geometric traversals to dynamic coordinate-based samplers that reroute tokens through predicted offsets and interpolation. While effective, these mechanisms primarily adapt paths or sampling locations, rather than explicitly modeling which local patches should exchange information before global state-space mixing. This motivates a simple question: \emph{can graphs help vision state space models see better?} We introduce \textbf{GraphScan}, a graph-induced dynamic scanning operator for Vision SSMs. For each token, GraphScan constructs a spatially bounded local graph, learns feature-conditioned affinities with relative positional bias, and produces the output token by one-step message passing over its semantic neighborhood. The resulting tokens are locally grounded before being processed by the selective SSM for global aggregation. GraphScan preserves token count and linear scaling in image size, while replacing coordinate-conditioned interpolation with feature-conditioned semantic routing. Integrated into a hierarchical backbone, \textbf{GraphScan-Mamba} achieves state-of-the-art performance among Vision SSMs across image classification, object detection, instance segmentation, and semantic segmentation, with modest computational overhead. Our analysis further shows that GraphScan induces interpretable displacement fields over the token lattice, providing a semantic and spatially grounded view of dynamic scanning. These results suggest that future Vision SSMs should treat scanning not merely as geometric serialization, but as learned local semantic routing before global state-space modeling.

Dhruv Parikh, Haoyang Fan, Rajgopal Kannan et al.

Vision-Language Models (VLMs) are expensive because the LLM processes hundreds of largely redundant visual tokens. Existing token reduction methods typically exploit \textit{either} vision-encoder saliency (broad but query-agnostic) \textit{or} LLM cross-attention (query-aware but sparse and costly). We show that neither signal alone is sufficient: fusing them consistently improves performance compared to unimodal visual token selection (ranking). However, making such fusion practical is non-trivial: cross-modal saliency is usually only available \emph{inside} the LLM (too late for efficient pre-LLM pruning), and the two signals are inherently asymmetric, so naive fusion underutilizes their complementary strengths. We propose \textbf{ConsensusDrop}, a training-free framework that derives a \emph{consensus} ranking by reconciling vision encoder saliency with query-aware cross-attention, retaining the most informative tokens while compressing the remainder via encoder-guided token merging. Across LLaVA-1.5/NeXT, Video-LLaVA, and other open-source VLMs, ConsensusDrop consistently outperforms prior pruning methods under identical token budgets and delivers a stronger accuracy-efficiency Pareto frontier -- preserving near-baseline accuracy even at aggressive token reductions while reducing TTFT and KV cache footprint. Our code will be open-sourced.

Jacob Fein-Ashley, Dhruv Parikh, Rajgopal Kannan et al.

Open-source Large Language Models (LLMs) increasingly specialize by domain (e.g., math, code, general reasoning), motivating systems that leverage complementary strengths across models. Prior multi-LLM approaches either (i) route a query to one or a few experts and generate independently, (ii) aggregate outputs from each model via costly multi-turn exchanges, or (iii) fuse weights into a single model-typically requiring architectural homogeneity. We introduce Mixture of Thoughts (MoT), a simple method for latent-level collaboration among heterogeneous experts under a global routing scheme. For each query, a lightweight router selects top-$K$ experts and designates a primary expert; uniformly placed interaction layers project hidden states into a shared latent space where the primary expert performs cross-attention over its active (selected) peers. Pre-trained experts remain frozen; only the router and the lightweight interaction layers are trained with a novel joint training objective that improves both the expert selection and inter-expert collaboration. Across five in-distribution (ID) and three out-of-distribution (OOD) benchmarks, MoT surpasses the current routing and aggregation-based state-of-the-art, Avengers, by $+0.38\%$ and $+2.92\%$, respectively. Further, MoT significantly outperforms the best-performing single model. It achieves this with single-pass inference, runtime comparable to routing baselines, and none of the overheads of iterative aggregation. MoT offers a simple latent-space mechanism for combining heterogeneous LLMs, a practical step toward broader multi-LLM collaboration. Our code is publicly available at https://github.com/jacobfa/mot.

Anish Bhattacharya, Nishanth Rao, Dhruv Parikh et al.

We demonstrate the capabilities of an attention-based end-to-end approach for high-speed vision-based quadrotor obstacle avoidance in dense, cluttered environments, with comparison to various state-of-the-art learning architectures. Quadrotor unmanned aerial vehicles (UAVs) have tremendous maneuverability when flown fast; however, as flight speed increases, traditional model-based approaches to navigation via independent perception, mapping, planning, and control modules breaks down due to increased sensor noise, compounding errors, and increased processing latency. Thus, learning-based, end-to-end vision-to-control networks have shown to have great potential for online control of these fast robots through cluttered environments. We train and compare convolutional, U-Net, and recurrent architectures against vision transformer (ViT) models for depth image-to-control in high-fidelity simulation, observing that ViT models are more effective than others as quadrotor speeds increase and in generalization to unseen environments, while the addition of recurrence further improves performance while reducing quadrotor energy cost across all tested flight speeds. We assess performance at speeds of up to 7m/s in simulation and hardware. To the best of our knowledge, this is the first work to utilize vision transformers for end-to-end vision-based quadrotor control.

Dhruv Parikh, Jacob Fein-Ashley, Tian Ye et al.

Convolutional Neural Networks (CNN) and Vision Transformers (ViT) have dominated the field of Computer Vision (CV). Graph Neural Networks (GNN) have performed remarkably well across diverse domains because they can represent complex relationships via unstructured graphs. However, the applicability of GNNs for visual tasks was unexplored till the introduction of Vision GNNs (ViG). Despite the success of ViGs, their performance is severely bottlenecked due to the expensive $k$-Nearest Neighbors ($k$-NN) based graph construction. Recent works addressing this bottleneck impose constraints on the flexibility of GNNs to build unstructured graphs, undermining their core advantage while introducing additional inefficiencies. To address these issues, in this paper, we propose a novel method called Dynamic Efficient Graph Convolution (DEGC) for designing efficient and globally aware ViGs. DEGC partitions the input image and constructs graphs in parallel for each partition, improving graph construction efficiency. Further, DEGC integrates local intra-graph and global inter-graph feature learning, enabling enhanced global context awareness. Using DEGC as a building block, we propose a novel CNN-GNN architecture, ClusterViG, for CV tasks. Extensive experiments indicate that ClusterViG reduces end-to-end inference latency for vision tasks by up to $5\times$ when compared against a suite of models such as ViG, ViHGNN, PVG, and GreedyViG, with a similar model parameter count. Additionally, ClusterViG reaches state-of-the-art performance on image classification, object detection, and instance segmentation tasks, demonstrating the effectiveness of the proposed globally aware learning strategy. Finally, input partitioning performed by DEGC enables ClusterViG to be trained efficiently on higher-resolution images, underscoring the scalability of our approach.

Dhruv Parikh, Shouyi Li, Bingyi Zhang et al.

Vision Transformers (ViTs) have achieved state-of-the-art accuracy on various computer vision tasks. However, their high computational complexity prevents them from being applied to many real-world applications. Weight and token pruning are two well-known methods for reducing complexity: weight pruning reduces the model size and associated computational demands, while token pruning further dynamically reduces the computation based on the input. Combining these two techniques should significantly reduce computation complexity and model size; however, naively integrating them results in irregular computation patterns, leading to significant accuracy drops and difficulties in hardware acceleration. Addressing the above challenges, we propose a comprehensive algorithm-hardware codesign for accelerating ViT on FPGA through simultaneous pruning -combining static weight pruning and dynamic token pruning. For algorithm design, we systematically combine a hardware-aware structured block-pruning method for pruning model parameters and a dynamic token pruning method for removing unimportant token vectors. Moreover, we design a novel training algorithm to recover the model's accuracy. For hardware design, we develop a novel hardware accelerator for executing the pruned model. The proposed hardware design employs multi-level parallelism with load balancing strategy to efficiently deal with the irregular computation pattern led by the two pruning approaches. Moreover, we develop an efficient hardware mechanism for efficiently executing the on-the-fly token pruning.

Sachini Wickramasinghe, Dhruv Parikh, Bingyi Zhang et al.

Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) is a key technique used in military applications like remote-sensing image recognition. Vision Transformers (ViTs) are the current state-of-the-art in various computer vision applications, outperforming their CNN counterparts. However, using ViTs for SAR ATR applications is challenging due to (1) standard ViTs require extensive training data to generalize well due to their low locality; the standard SAR datasets, however, have a limited number of labeled training data which reduces the learning capability of ViTs; (2) ViTs have a high parameter count and are computation intensive which makes their deployment on resource-constrained SAR platforms difficult. In this work, we develop a lightweight ViT model that can be trained directly on small datasets without any pre-training by utilizing the Shifted Patch Tokenization (SPT) and Locality Self-Attention (LSA) modules. We directly train this model on SAR datasets which have limited training samples to evaluate its effectiveness for SAR ATR applications. We evaluate our proposed model, that we call VTR (ViT for SAR ATR), on three widely used SAR datasets: MSTAR, SynthWakeSAR, and GBSAR. Further, we propose a novel FPGA accelerator for VTR, in order to enable deployment for real-time SAR ATR applications.

Dhruv Parikh, Viktor Prasanna

Hyperdimensional Computing (HDC) is a brain-inspired computing paradigm that represents and manipulates information using high-dimensional vectors, called hypervectors (HV). Traditional HDC methods, while robust to noise and inherently parallel, rely on single-pass, non-parametric training and often suffer from low accuracy. To address this, recent approaches adopt iterative training of base and class HVs, typically accelerated on GPUs. Inference, however, remains lightweight and well-suited for real-time execution. Yet, efficient HDC inference has been studied almost exclusively on specialized hardware such as FPGAs and GPUs, with limited attention to general-purpose multi-core CPUs. To address this gap, we propose ScalableHD for scalable and high-throughput HDC inference on multi-core CPUs. ScalableHD employs a two-stage pipelined execution model, where each stage is parallelized across cores and processes chunks of base and class HVs. Intermediate results are streamed between stages using a producer-consumer mechanism, enabling on-the-fly consumption and improving cache locality. To maximize performance, ScalableHD integrates memory tiling and NUMA-aware worker-to-core binding. Further, it features two execution variants tailored for small and large batch sizes, each designed to exploit compute parallelism based on workload characteristics while mitigating the memory-bound compute pattern that limits HDC inference performance on modern multi-core CPUs. ScalableHD achieves up to 10x speedup in throughput (samples per second) over state-of-the-art baselines such as TorchHD, across a diverse set of tasks ranging from human activity recognition to image classification, while preserving task accuracy. Furthermore, ScalableHD exhibits robust scalability: increasing the number of cores yields near-proportional throughput improvements.

Dhruv Parikh, Jebacyril Arockiaraj, Viktor Prasanna

Hyperdimensional Computing (HDC) represents data using extremely high-dimensional, low-precision vectors, termed hypervectors (HVs), and performs learning and inference through lightweight, noise-tolerant operations. However, the high dimensionality, sparsity, and repeated data movement involved in HDC make these computations difficult to accelerate efficiently on conventional processors. As a result, executing core HDC operations: binding, permutation, bundling, and similarity search: on CPUs or GPUs often leads to suboptimal utilization, memory bottlenecks, and limits on real-time performance. In this paper, our contributions are two-fold. First, we develop an image-encoding algorithm that, similar in spirit to convolutional neural networks, maps local image patches to hypervectors enriched with spatial information. These patch-level hypervectors are then merged into a global representation using the fundamental HDC operations, enabling spatially sensitive and robust image encoding. This encoder achieves 95.67% accuracy on MNIST and 85.14% on Fashion-MNIST, outperforming prior HDC-based image encoders. Second, we design an end-to-end accelerator that implements these compute operations on an FPGA through a pipelined architecture that exploits parallelism both across the hypervector dimensionality and across the set of image patches. Our Alveo U280 implementation delivers 0.09ms inference latency, achieving up to 1300x and 60x speedup over state-of-the-art CPU and GPU baselines, respectively.

Neelesh Gupta, Rakshith Jayanth, Dhruv Parikh et al.

The proliferation of large language models (LLMs) has driven demand for long context inference on resource constrained edge devices. However, deploying these models on Neural Processing Units (NPUs) presents significant challenges due to the architectural mismatch: quadratic complexity of standard attention mechanisms conflicts with memory and compute patterns of edge accelerators. This paper presents a comprehensive performance analysis of various causal inference operators on a modern NPU. We benchmark standard quadratic attention against several sub-quadratic alternatives, including structured state-space and linear attention models. Our analysis reveals that while sub-quadratic methods offer superior scalability, they introduce distinct computational bottlenecks on the NPU's specialized execution units. We identify that quadratic attention becomes severely memory-bound, suffering from cache inefficiency and pipeline stalls exceeding 95% at long contexts. In contrast, sub-quadratic models can become compute-bound on programmable vector cores. These findings provide critical insights for the co-design of hardware-aware models and optimization strategies to enable on-device AI inference with long-contexts.

Mann Patel, Divyajyoti Panda, Hilay Mehta et al.

Parameter-Efficient Fine-Tuning (PEFT) is an efficient alternative to full scale fine-tuning, gaining popularity recently. With pre-trained model sizes growing exponentially, PEFT can be effectively utilized to fine-tune compact modules, Parameter-Efficient Modules (PEMs), trained to be domain experts over diverse domains. In this project, we explore composing such individually fine-tuned PEMs for distribution generalization over the composite domain. To compose PEMs, simple composing functions are used that operate purely on the weight space of the individually fine-tuned PEMs, without requiring any additional fine-tuning. The proposed method is applied to the task of representing the 16 Myers-Briggs Type Indicator (MBTI) composite personalities via 4 building block dichotomies, comprising of 8 individual traits which can be merged (composed) to yield a unique personality. We evaluate the individual trait PEMs and the composed personality PEMs via an online MBTI personality quiz questionnaire, validating the efficacy of PEFT to fine-tune PEMs and merging PEMs without further fine-tuning for domain composition.