arXiv subjects

Mohit Goyal

Publications and source records attributed to Mohit Goyal.

11 recordsLinked to original sources

arXiv2026

GRIT -- Geometry-Aware PEFT with K-FACPreconditioning, Fisher-Guided Reprojection, andDynamic Rank Adaptation

Parameter-efficient fine-tuning (PEFT) is the default way to adapt LLMs, but widely used LoRA and QLoRA are largely geometry-agnostic: they optimize in fixed, randomly oriented low-rank subspaces with first-order descent, mostly ignoring local loss curvature. This can inflate the effective update budget and amplify drift along weakly constrained directions. We introduce GRIT, a dynamic, curvature-aware LoRA procedure that preserves the LoRA parameterization but: (1) preconditions gradients in rank space using K-FAC as a natural-gradient proxy; (2) periodically reprojects the low-rank basis onto dominant Fisher eigendirections to suppress drift; and (3) adapts the effective rank from the spectrum so capacity concentrates where signal resides. Across instruction-following, comprehension, and reasoning benchmarks on LLaMA backbones, GRIT matches or surpasses LoRA and QLoRA while reducing trainable parameters by 46% on average (25--80% across tasks), without practical quality loss across prompt styles and data mixes. To model forgetting, we fit a curvature-modulated power law. Empirically, GRIT yields lower drift and a better updates-vs-retention frontier than strong PEFT-optimizer baselines (Orthogonal-LoRA, IA3, DoRA, Eff-FT, Shampoo).

cs.LG↗

arXiv2025

MapTrace: Scalable Data Generation for Route Tracing on Maps

While Multimodal Large Language Models have achieved human-like performance on many visual and textual reasoning tasks, their proficiency in fine-grained spatial understanding, such as route tracing on maps remains limited. Unlike humans, who can quickly learn to parse and navigate maps, current models often fail to respect fundamental path constraints, in part due to the prohibitive cost and difficulty of collecting large-scale, pixel-accurate path annotations. To address this, we introduce a scalable synthetic data generation pipeline that leverages synthetic map images and pixel-level parsing to automatically produce precise annotations for this challenging task. Using this pipeline, we construct a fine-tuning dataset of 23k path samples across 4k maps, enabling models to acquire more human-like spatial capabilities. Using this dataset, we fine-tune both open-source and proprietary MLLMs. Results on MapBench show that finetuning substantially improves robustness, raising success rates by up to 6.4 points, while also reducing path-tracing error (NDTW). These gains highlight that fine-grained spatial reasoning, absent in pretrained models, can be explicitly taught with synthetic supervision.

cs.CV↗

arXiv2025

Geometry Aware Passthrough Mitigates Cybersickness

Virtual Reality headsets isolate users from the real-world by restricting their perception to the virtual-world. Video See-Through (VST) headsets address this by utilizing world-facing cameras to create Augmented Reality experiences. However, directly displaying camera feeds causes visual discomfort and cybersickness due to the inaccurate perception of scale and exaggerated motion parallax. This paper demonstrates the potential of geometry aware passthrough systems in mitigating cybersickness through accurate depth perception. We first present a methodology to benchmark and compare passthrough algorithms. Furthermore, we design a protocol to quantitatively measure cybersickness experienced by users in VST headsets. Using this protocol, we conduct a user study to compare direct passthrough and geometry aware passthrough systems. To the best of our knowledge, our study is the first one to reveal significantly reduced nausea, disorientation, and total scores of cybersickness with geometry aware passthrough. It also uncovers several potential avenues to further mitigate visually-induced discomfort.

cs.HC↗

arXiv2023

Roaming Performance Analysis and Comparison between Wi-Fi and Private Cellular Network

Private network deployment is gaining momentum in warehouses, retail, automation, health care, and many such use cases to guarantee mission-critical services with less latency. Guaranteeing the delay-sensitive application in Wi-Fi is always challenging due to the nature of unlicensed spectrum. As the device ecosystem keeps growing and expanding, all the current and future devices can support both Wi-Fi and Private Cellular Network (CBRS is the primary spectrum in the US for private network deployment). However, due to the existing infrastructure and huge investment in the dense Wi-Fi network, consumers prefer two deployment models. The first scenario is deploying the private network outdoors and using the existing Wi-Fi indoors. The second scenario is to use the existing Wi-Fi network as a backup for offloading the traffic indoors and parallely utilizes the private network deployment for less latency applications. Hence, we expect, in both scenarios, a roaming between two technologies \emph{i.e.,} Wi-Fi and Private Cellular Network. In this work, we would like to quantify the roaming performance or service interruption time when the device moves from Wi-Fi to Private Network (CBRS) and vice-versa.

cs.NI↗

arXiv2022

Human Hands as Probes for Interactive Object Understanding

Interactive object understanding, or what we can do to objects and how is a long-standing goal of computer vision. In this paper, we tackle this problem through observation of human hands in in-the-wild egocentric videos. We demonstrate that observation of what human hands interact with and how can provide both the relevant data and the necessary supervision. Attending to hands, readily localizes and stabilizes active objects for learning and reveals places where interactions with objects occur. Analyzing the hands shows what we can do to objects and how. We apply these basic principles on the EPIC-KITCHENS dataset, and successfully learn state-sensitive features, and object affordances (regions of interaction and afforded grasps), purely by observing hands in egocentric videos.

cs.CV↗

arXiv2022

A Note on the Stochastic Ruler Method for Discrete Simulation Optimization

In this paper, we propose a relaxation to the stochastic ruler method originally described by Yan and Mukai in 1992 for asymptotically determining the global optima of discrete simulation optimization problems. The `original' version of the stochastic ruler and its variants require that a candidate for the next estimate of the optimal solution pass a certain number of tests with respect to the stochastic ruler to be selected as the next estimate of the optimal solution. This requirement - that all tests need to be passed - can lead to promising candidate solutions being rejected and can slow down the convergence of the algorithm. Our proposed modification to the stochastic ruler algorithm relaxes this requirement, and we show analytically that our proposed variant of the stochastic ruler method incurs lesser computational overhead when a new solution in the neighborhood of the current solution is a `successful' candidate for the next estimate of the current solution. We then show numerically that this can yield accelerated convergence to the optimal solution via multiple numerical examples. We also provide the theoretical grounding for the asymptotic convergence in probability of the variant to the global optimal solution under the same set of assumptions as those underlying the original stochastic ruler method.

math.OC↗

arXiv2020

Learning Activation Functions: A new paradigm for understanding Neural Networks

The scope of research in the domain of activation functions remains limited and centered around improving the ease of optimization or generalization quality of neural networks (NNs). However, to develop a deeper understanding of deep learning, it becomes important to look at the non linear component of NNs more carefully. In this paper, we aim to provide a generic form of activation function along with appropriate mathematical grounding so as to allow for insights into the working of NNs in future. We propose "Self-Learnable Activation Functions" (SLAF), which are learned during training and are capable of approximating most of the existing activation functions. SLAF is given as a weighted sum of pre-defined basis elements which can serve for a good approximation of the optimal activation function. The coefficients for these basis elements allow a search in the entire space of continuous functions (consisting of all the conventional activations). We propose various training routines which can be used to achieve performance with SLAF equipped neural networks (SLNNs). We prove that SLNNs can approximate any neural network with lipschitz continuous activations, to any arbitrary error highlighting their capacity and possible equivalence with standard NNs. Also, SLNNs can be completely represented as a collections of finite degree polynomial upto the very last layer obviating several hyper parameters like width and depth. Since the optimization of SLNNs is still a challenge, we show that using SLAF along with standard activations (like ReLU) can provide performance improvements with only a small increase in number of parameters.

cs.LG↗

arXiv2020

DZip: improved general-purpose lossless compression based on novel neural network modeling

We consider lossless compression based on statistical data modeling followed by prediction-based encoding, where an accurate statistical model for the input data leads to substantial improvements in compression. We propose DZip, a general-purpose compressor for sequential data that exploits the well-known modeling capabilities of neural networks (NNs) for prediction, followed by arithmetic coding. Dzip uses a novel hybrid architecture based on adaptive and semi-adaptive training. Unlike most NN based compressors, DZip does not require additional training data and is not restricted to specific data types, only needing the alphabet size of the input data. The proposed compressor outperforms general-purpose compressors such as Gzip (on average 26% reduction) on a variety of real datasets, achieves near-optimal compression on synthetic datasets, and performs close to specialized compressors for large sequence lengths, without any human input. The main limitation of DZip in its current implementation is the encoding/decoding time, which limits its practicality. Nevertheless, the results showcase the potential of developing improved general-purpose compressors based on neural networks and hybrid modeling.

cs.LG↗

arXiv2019

Detection of Glottal Closure Instants from Raw Speech using Convolutional Neural Networks

Glottal Closure Instants (GCIs) correspond to the temporal locations of significant excitation to the vocal tract occurring during the production of voiced speech. GCI detection from speech signals is a well-studied problem given its importance in speech processing. Most of the existing approaches for GCI detection adopt a two-stage approach (i) Transformation of speech signal into a representative signal where GCIs are localized better, (ii) extraction of GCIs using the representative signal obtained in first stage. The former stage is accomplished using signal processing techniques based on the principles of speech production and the latter with heuristic-algorithms such as dynamic-programming and peak-picking. These methods are thus task-specific and rely on the methods used for representative signal extraction. However, in this paper, we formulate the GCI detection problem from a representation learning perspective where appropriate representation is implicitly learned from the raw-speech data samples. Specifically, GCI detection is cast as a supervised multi-task learning problem solved using a deep convolutional neural network jointly optimizing a classification and regression cost. The learning capability is demonstrated with several experiments on standard datasets. The results compare well with the state-of-the-art algorithms while performing better in the case of presence of real-world non-stationary noise.

cs.SD↗

arXiv2019

Achievable Rates of Attack Detection Strategies in Echo-Assisted Communication

We consider an echo-assisted communication model wherein block-coded messages, when transmitted across several frames, reach the destination as multiple noisy copies. We address adversarial attacks on such models wherein a subset of the noisy copies are vulnerable to manipulation by an adversary. Particularly, we study a non-persistent attack model with the adversary attacking 50% of the frames on the vulnerable copies in an i.i.d. fashion. We show that this adversarial model drives the destination to detect the attack locally within every frame, thereby resulting in degraded performance due to false-positives and miss-detection. Our main objective is to characterize the mutual information of this adversarial echo-assisted channel by incorporating the performance of attack-detection strategies. With the use of an imperfect detector, we show that the compound channel comprising the adversarial echo-assisted channel and the attack detector exhibits memory-property, and as a result, obtaining closed-form expressions on mutual information is intractable. To circumvent this problem, we present a new framework to approximate the mutual information by deriving sufficient conditions on the channel parameters and also the performance of the attack detectors. Finally, we propose two attack-detectors, which are inspired by traditional as well as neural-network ideas, and show that the mutual information offered by these detectors is close to that of the Genie detector for short frame-lengths.

cs.IT↗

arXiv2018

DeepZip: Lossless Data Compression using Recurrent Neural Networks

Sequential data is being generated at an unprecedented pace in various forms, including text and genomic data. This creates the need for efficient compression mechanisms to enable better storage, transmission and processing of such data. To solve this problem, many of the existing compressors attempt to learn models for the data and perform prediction-based compression. Since neural networks are known as universal function approximators with the capability to learn arbitrarily complex mappings, and in practice show excellent performance in prediction tasks, we explore and devise methods to compress sequential data using neural network predictors. We combine recurrent neural network predictors with an arithmetic coder and losslessly compress a variety of synthetic, text and genomic datasets. The proposed compressor outperforms Gzip on the real datasets and achieves near-optimal compression for the synthetic datasets. The results also help understand why and where neural networks are good alternatives for traditional finite context models

cs.CL↗