arXiv subjects

Brijesh Soni

Publications and source records attributed to Brijesh Soni.

4 recordsLinked to original sources

arXiv2026

Large Models for Small Devices: Recent Advances and Empirical Analysis of Edge AI Deployment

Running large AI models on resource-constrained edge devices requires model compression to reduce model size and computation. What compresses well, however, need not deploy well. We survey dozens of recent works that report compression results on real hardware and extract practical deployment guidelines from them. Following these guidelines, we deploy compact language and image models on GPU, CPU, and Raspberry Pi platforms across question answering and image segmentation. No single technique wins across tasks. For question answering, Qwen3.5 0.8B reaches 93.85 SQuAD F1 and 92 EM under Q5_K_M GGUF quantization, while structured pruning at the same precision costs 16 F1 at a 1% ratio. For segmentation, the ranking reverses: default quantization leaves parameters and MACs unchanged, whereas pruning cuts model size by nearly 80% at near-constant mIoU. Pruning can even inflate the deployed artifact by 21-49% by breaking k-quant super-block alignment; combined with longer, less format-compliant outputs, this raises Raspberry Pi latency up to 3.4x. Compression can also manufacture the appearance of competence rather than destroy it visibly: one LoRA-recovered variant stays fully parseable and holds 71% strict BoolQ accuracy while sending 97 of 100 predictions to a single class, at 52.6% balanced accuracy. We explain these effects through neural-flow graph analysis and prefill-decode-level latency decomposition, and condense them into task-specific deployment research directions. The right technique depends on the task, the model, and the hardware. Our experiment code and artifacts are open-sourced at https://github.com/Arnavvvkumar/deployment

cs.AI↗

arXiv2025

Vision Transformer Based User Equipment Positioning

Recently, Deep Learning (DL) techniques have been used for User Equipment (UE) positioning. However, the key shortcomings of such models is that: i) they weigh the same attention to the entire input; ii) they are not well suited for the non-sequential data e.g., when only instantaneous Channel State Information (CSI) is available. In this context, we propose an attention-based Vision Transformer (ViT) architecture that focuses on the Angle Delay Profile (ADP) from CSI matrix. Our approach, validated on the `DeepMIMO' and `ViWi' ray-tracing datasets, achieves an Root Mean Squared Error (RMSE) of 0.55m indoors, 13.59m outdoors in DeepMIMO, and 3.45m in ViWi's outdoor blockage scenario. The proposed scheme outperforms state-of-the-art schemes by $\sim$ 38\%. It also performs substantially better than other approaches that we have considered in terms of the distribution of error distance.

cs.CV↗

arXiv2025

Optimizing Multi-UAV 3D Deployment for Energy-Efficient Sensing over Uneven Terrains

In this work, we consider a multi-unmanned aerial vehicle (UAV) cooperative sensing system where UAVs are deployed to sense multiple targets in terrain-aware line of sight (LoS) conditions in uneven terrain equipped with directional antennas. To mitigate terrain-induced LoS blockages that degrade detection performance, we incorporate a binary LoS indicator and propose a bounding volume hierarchy (BHV)-based adaptive scheme for efficient LoS evaluation. We formulate a bi-objective problem that maximizes the probability of cooperative detection with minimal hover energy constraints governing spatial, orientational, and safety constraints. To address the problem, which is inherently non-convex, we propose a hierarchical heuristic framework that combines exploration through a genetic algorithm (GA) with per-UAV refinement via particle swarm optimization (PSO), where a penalty-based fitness evaluation guides solutions toward feasibility, bounded within constraints. The proposed methodology is an effective trade-off method of traversing through a complex search space and maintaining terrain-aware LoS connectivity and energy aware deployment. Monte Carlo simulations on real-world terrain data show that the proposed GA+PSO framework improves detection probability by 37.02% and 36.5% for 2 and 3 UAVs, respectively, while reducing average excess hover energy by 45.0% and 48.9% compared to the PSO-only baseline. Relative to the non-optimized scheme, it further achieves 59.5% and 54.2% higher detection probability with 59.8% and 65.9% lower excess hover energy, thereby showing its effectiveness with a small number of UAVs over uneven terrain.

cs.NI↗

arXiv2023

Forecasting Aided Energy Aware Band Assignment in Multiband Networks

The high frequency communication bands (mmWave and sub-THz) promise tremendous data rates. However, they have very high power consumption which is particularly significant for battery-powered user-equipment (UE), and are prone to blockage. In this context, we design an energy-aware band-assignment system which reduces power consumption while aiming to achieve a target sum rate of M bits/sec in T time-slots. We do this by using 1) Rate forecaster(s); 2) Channel forecaster(s) which forecast either the data rate or the channel for T subsequent time slots, utilizing either a stacked Long-short-term memory (LSTM) or transformer architecture. These forecasts are used to select frequency bands using an iterative algorithm. The proposed approach is validated on the publicly available `DeepMIMO', and `NYUSIM' datasets for both outdoor and indoor scenarios, and using a multiband empirical system. Moreover, we also propose a simple blockage segment generation algorithm such that the channel realizations inherently capture the effects of blockage. We find that the rate-forecaster-based approach outperforms the channel forecaster. Further, our approach consumes ~300 mW lower power compared to a greedy band assignment at a 1.5 Gb/s target rate for outdoor scenarios, and up to 600 mW for indoor scenarios at 2.5 Gb/s target rate, indicating that this is a promising method to reduce UE power consumption in multiband systems.

eess.SP↗