-
ROCS: Request-Oriented Compute Sharing for Efficient Large-Scale Recommendation
Authors:
Yuxin Chen,
Liang Luo,
Buyun Zhang,
Jian Jiao,
Boda Li,
Haoyu Wang,
Tongyi Tang,
Ao Cai,
Zijian Shen,
Zhengkai Zhang,
Wenyi Xie,
Ryan Dick,
Han Liu,
Neng Shi,
Bin Yu,
Jianbo Xiao,
Shuyao Bi,
Hongtao Yu,
Yuanwei Fang,
Zhuoran Zhao,
Sijia Chen,
Yang Chen,
Shuqi Yang,
Qianru Li,
Zikun Liu
, et al. (22 additional authors not shown)
Abstract:
Modern recommendation models gain prediction quality by scaling feature-interaction and sequence modules, but production cost constraints cap how far systems can scale.
In this work, we propose Request-Oriented Compute Sharing (ROCS), a modeling and inference paradigm that exploits a unique property of recommendation inference: each user request is evaluated against many candidates, while reques…
▽ More
Modern recommendation models gain prediction quality by scaling feature-interaction and sequence modules, but production cost constraints cap how far systems can scale.
In this work, we propose Request-Oriented Compute Sharing (ROCS), a modeling and inference paradigm that exploits a unique property of recommendation inference: each user request is evaluated against many candidates, while request-side features are shared across candidates. ROCS defers request-candidate interactions as late as possible, isolates candidate-dependent representations, and evaluates substantial portions of the model once per request rather than once per candidate, significantly improving inference efficiency while maintaining or improving prediction quality. To realize this paradigm, we develop Generalized Layer Masking (GLM) to enforce candidate isolation in feature-interaction architectures, and Deep Cross Attention (DCA) to extend request-oriented sharing to sequence architectures. To support efficient GPU deployment, we co-design In-Kernel Broadcast Optimization (IKBO) that significantly accelerates ROCS model execution.
Experiments on public benchmarks show that ROCS consistently improves the quality-efficiency tradeoff across recommendation backbones. On production-scale workloads, ROCS achieves up to a 3x QPS improvement on retrieval models without quality degradation and a 0.5% relative LogLoss improvement with a 50% QPS gain on a short-form video ranking model. ROCS has been deployed across large-scale recommendation systems spanning ads and organic surfaces, retrieval and ranking stages, and more than two orders of magnitude in inference complexity, delivering significant online gains at reduced infrastructure cost.
△ Less
Submitted 30 July, 2026;
originally announced July 2026.
-
PermDoRA -- Understanding Adapter Interference in Language Models: Limits of Parameter-Space Geometry
Authors:
Gowtham Sivaramakrishnan,
Sarvesha Kumar Kombaiah Seetha,
Kishan Gupta Balaji,
Santhosh Baradwaj Vaduvur Ranganathan
Abstract:
Access control in large language models (LLMs) requires modular mechanisms to enable domain-specific behavior without retraining or cross-domain interference. A common hypothesis is that interference during adapter composition arises from overlap in linear parameter updates, suggesting that enforcing orthogonality or directional independence should improve multi-domain performance. We test this hy…
▽ More
Access control in large language models (LLMs) requires modular mechanisms to enable domain-specific behavior without retraining or cross-domain interference. A common hypothesis is that interference during adapter composition arises from overlap in linear parameter updates, suggesting that enforcing orthogonality or directional independence should improve multi-domain performance. We test this hypothesis using DoRA-RBAC, a hierarchical adapter composition framework based on weight-decomposed low-rank adaptation. We compare conventional Euclidean merging with a geometry-aware Riemannian-inspired merging strategy that approximates the Frechet mean via normalized directional averaging across multiple QA benchmarks (GPQA, PubMedQA, SimpleQA, WMDP) on LLaMA-3.1-8B and Mistral-7B. Our results show that while single-domain performance matches LoRA, geometry-aware merging provides no consistent advantage over standard averaging in multi-domain settings.Diagnostic analysis further reveals that angular alignment and orthogonality of adapter updates are weak predictors of composition performance. These findings suggest that adapter interference is not governed primarily by parameter-space geometry, but is instead consistent with interactions in shared nonlinear representations.
△ Less
Submitted 8 June, 2026;
originally announced June 2026.
-
LoKA: Low-precision Kernel Applications for Recommendation Models At Scale
Authors:
Liang Luo,
Yinbin Ma,
Quanyu Zhu,
Vasiliy Kuznetsov,
Yuxin Chen,
Neng Shi,
Jian Jiao,
Jiecao Yu,
Buyun Zhang,
Tongyi Tang,
Xiaohan Wei,
Yanli Zhao,
Zeliang Chen,
Yuchen Hao,
Venkatesh Ranganathan,
Sandeep Parab,
Yantao Yao,
Maxim Naumov,
Chunzhi Yang,
Shen Li,
Ellie Wen,
Wenlin Chen,
Santanu Kolay,
Chunqiang Tang
Abstract:
Recent GPU generations deliver significantly higher FLOPs using lower-precision arithmetic, such as FP8. While successfully applied to large language models (LLMs), its adoption in large recommendation models (LRMs) has been limited. This is because LRMs are numerically sensitive, dominated by small matrix multiplications (GEMMs) followed by normalization, and trained in communication-intensive en…
▽ More
Recent GPU generations deliver significantly higher FLOPs using lower-precision arithmetic, such as FP8. While successfully applied to large language models (LLMs), its adoption in large recommendation models (LRMs) has been limited. This is because LRMs are numerically sensitive, dominated by small matrix multiplications (GEMMs) followed by normalization, and trained in communication-intensive environments. Applying FP8 directly to LRMs often degrades model quality and prolongs training time. These challenges are inherent to LRM workloads and cannot be resolved merely by introducing better FP8 kernels. Instead, a system-model co-design approach is needed to successfully integrate FP8. We present LoKA (Low-precision Kernel Applications), a framework that makes FP8 practical for LRMs through three principles: profile under realistic distributions to know where low precision is safe, co-design model components with hardware to expand where it is safe, and orchestrate across kernel libraries to maximize the gains. Concretely, LoKA Probe is a statistically grounded, online benchmarking method that learns activation and weight statistics, and quantifies per-layer errors. This process pinpoints safe and unsafe, fast and slow sites for FP8 adoption. LoKA Mods is a set of reusable model adaptations that improve both numerical stability and execution efficiency with FP8. LoKA Dispatch is a runtime that leverages the statistical insights from LoKA Probe to select the fastest FP8 kernel that satisfies the accuracy requirements.
△ Less
Submitted 8 July, 2026; v1 submitted 11 May, 2026;
originally announced May 2026.
-
Listening Alone, Understanding Together: Collaborative Context Recovery for Privacy-Aware AI
Authors:
Tanmay Srivastava,
Amartya Basu,
Shubham Jain,
Vaishnavi Ranganathan
Abstract:
We introduce CONCORD, a privacy-aware asynchronous assistant-to-assistant (A2A) framework that leverages collaboration between proactive speech-based AI. As agents evolve from reactive to always-listening assistants, they face a core privacy risk (of capturing non-consenting speakers), which makes their social deployment a challenge. To overcome this, we implement CONCORD, which enforces owner-onl…
▽ More
We introduce CONCORD, a privacy-aware asynchronous assistant-to-assistant (A2A) framework that leverages collaboration between proactive speech-based AI. As agents evolve from reactive to always-listening assistants, they face a core privacy risk (of capturing non-consenting speakers), which makes their social deployment a challenge. To overcome this, we implement CONCORD, which enforces owner-only speech capture via real-time speaker verification, producing a one-sided transcript that incurs missing context but preserves privacy. We demonstrate that CONCORD can safely recover necessary context through (1) spatio-temporal context resolution, (2) information gap detection, and (3) minimal A2A queries governed by a relationship-aware disclosure. Instead of hallucination-prone inferring, CONCORD treats context recovery as a negotiated safe exchange between assistants. Across a multi-domain dialogue dataset, CONCORD achieves 91.4% recall in gap detection, 96% relationship classification accuracy, and 97% true negative rate in privacy-sensitive disclosure decisions. By reframing always-listening AI as a coordination problem between privacy-preserving agents, CONCORD offers a practical path toward socially deployable proactive conversational agents.
△ Less
Submitted 14 April, 2026;
originally announced April 2026.
-
A Single-Chain Backscatter Tag for Multi-Sensor Multiplexing
Authors:
Yijie Li,
Weichong Ling,
Taiting Lu,
Bao Dao,
Yi-Chao Chen,
Vaishnavi Ranganathan,
Lili Qiu,
Jingxian Wang
Abstract:
Many real-world sensing tasks require co-located, multi-modal measurements at a single site, typically a bundle of two to five sensors, for example, in plant stress sensing and blood pressure estimation. RF-backscatter devices have emerged as a low-power solution for sensing, yet existing backscatter tags support a single sensor. Placing several single-sensor tags at one site increases attachment…
▽ More
Many real-world sensing tasks require co-located, multi-modal measurements at a single site, typically a bundle of two to five sensors, for example, in plant stress sensing and blood pressure estimation. RF-backscatter devices have emerged as a low-power solution for sensing, yet existing backscatter tags support a single sensor. Placing several single-sensor tags at one site increases attachment footprint and induces mutual coupling between nearby tag antennas, thereby limiting practical deployment.
We present MATRIX, a single-chain multi-sensor backscatter tag that concurrently supports multiple onboard sensors and multiplexes them as a composite voltage, then backscatters it through one analog modulation chain. Rather than time-division polling, which introduces inter-sensor sampling offsets, or frequency-division, which requires independent per-sensor modulation chains, MATRIX introduces a voltage-division multiplexing architecture in which each sensor value is encoded as a PWM waveform, carrying the measurement in its duty cycle and reserving the amplitude for multiplexing. To support reliable demultiplexing, MATRIX selects the voltage-division weights in a binary-weighted geometric progression so that every active-sensor set maps to a uniquely invertible, well-spaced composite voltage. The composite voltage is then converted into backscatter frequency shifts through a single modulation chain. At the receiver, MATRIX formulates demultiplexing as a Hidden Markov Model to recover per-sensor readings while tolerating analog hardware imperfections and multipath. MATRIX's ASIC design consumes 25.56uW. Detailed evaluation shows that the prototype, multiplexing five sensors, achieves 20 dB average signal reconstruction SNR at a 30 kHz sampling frequency; we further validate MATRIX with case studies in plant sensing, health monitoring, and microphone-based direction finding.
△ Less
Submitted 14 November, 2025; v1 submitted 2 July, 2025;
originally announced July 2025.
-
TerraTrace: Temporal Signature Land Use Mapping System
Authors:
Angela Busheska,
Vikram Iyer,
Bruno Silva,
Peder Olsen,
Ranveer Chandra,
Vaishnavi Ranganathan
Abstract:
Understanding land use over time is critical to tracking events related to climate change, like deforestation. However, satellite-based remote sensing tools which are used for monitoring struggle to differentiate vegetation types in farms and orchards from forests. We observe that metrics such as the Normalized Difference Vegetation Index (NDVI), based on plant photosynthesis, have unique temporal…
▽ More
Understanding land use over time is critical to tracking events related to climate change, like deforestation. However, satellite-based remote sensing tools which are used for monitoring struggle to differentiate vegetation types in farms and orchards from forests. We observe that metrics such as the Normalized Difference Vegetation Index (NDVI), based on plant photosynthesis, have unique temporal signatures that reflect agricultural practices and seasonal cycles. We analyze yearly NDVI changes on 20 farms for 10 unique crops. Initial results show that NDVI curves are coherent with agricultural practices, are unique to each crop, consistent globally, and can differentiate farms from forests. We develop a novel longitudinal NDVI dataset for the state of California from 2020-2023 with 500~m resolution and over 70 million points. We use this to develop the TerraTrace platform, an end-to-end analytic tool that classifies land use using NDVI signatures and allows users to query the system through an LLM chatbot and graphical interface.
△ Less
Submitted 25 February, 2025;
originally announced February 2025.
-
Brain Tumor Classification From MRI Images Using Machine Learning
Authors:
Vidhyapriya Ranganathan,
Celshiya Udaiyar,
Jaisree Jayanth,
Meghaa P V,
Srija B,
Uthra S
Abstract:
Brain tumor is a life-threatening problem and hampers the normal functioning of the human body. The average five-year relative survival rate for malignant brain tumors is 35.6 percent. For proper diagnosis and efficient treatment planning, it is necessary to detect the brain tumor in early stages. Due to advancement in medical imaging technology, the brain images are taken in different modalities.…
▽ More
Brain tumor is a life-threatening problem and hampers the normal functioning of the human body. The average five-year relative survival rate for malignant brain tumors is 35.6 percent. For proper diagnosis and efficient treatment planning, it is necessary to detect the brain tumor in early stages. Due to advancement in medical imaging technology, the brain images are taken in different modalities. The ability to extract relevant characteristics from magnetic resonance imaging (MRI) scans is a crucial step for brain tumor classifiers. Several studies have proposed various strategies to extract relevant features from different modalities of MRI to predict the growth of abnormal tumors. Most techniques used conventional methods of image processing for feature extraction and machine learning for classification. More recently, the use of deep learning algorithms in medical imaging has resulted in significant improvements in the classification and diagnosis of brain tumors. Since tumors are located at different regions of the brain, localizing the tumor and classifying it to a particular category is a challenging task. The objective of this project is to develop a predictive system for brain tumor detection using machine learning(ensembling).
△ Less
Submitted 15 July, 2024;
originally announced July 2024.
-
Cellular LTE and Solar Energy Harvesting for Long-Term, Reliable Urban Sensor Networks: Challenges and Opportunities
Authors:
Alex Cabral,
Vaishnavi Ranganathan,
Jim Waldo
Abstract:
In a world driven by data, cities are increasingly interested in deploying networks of smart city devices for urban and environmental monitoring. To be successful, these networks must be reliable, scalable, real-time, low-cost, and easy to install and maintain -- criteria that are all significantly affected by the design choices around connectivity and power. LTE networks and solar energy can seem…
▽ More
In a world driven by data, cities are increasingly interested in deploying networks of smart city devices for urban and environmental monitoring. To be successful, these networks must be reliable, scalable, real-time, low-cost, and easy to install and maintain -- criteria that are all significantly affected by the design choices around connectivity and power. LTE networks and solar energy can seemingly both satisfy the necessary criteria and are often used in real-world sensor network deployments. However, there have not been extensive real-world studies to examine how well such networks perform and the challenges they encounter in urban settings over long periods. In this work, we analyze the performance of a stationary 118-node LTE-connected, solar-powered sensor network over one year in Chicago. Results show the promise of LTE networks and solar panels for city-wide IoT deployments, but also reveal areas for improvement. Notably, we find 11 sites with inadequate RSS to support sensing nodes and over 33,000 hours of data loss due to solar energy availability issues between October and March. Furthermore, we discover that the neighborhoods most affected by connectivity and charging issues are socioeconomically disadvantaged areas with a majority Black and Latine residents. This work presents observations from a networking and powering perspective of the urban sensor network to help drive reliable, scalable future smart city deployments. The work also analyzes the impact of land use, adaptive energy harvesting management strategies, and shortcomings of open data, to support the need for increased real-world deployments that ensure the design of equitable smart city networks.
△ Less
Submitted 8 July, 2023;
originally announced July 2023.
-
Towards a Cost vs. Quality Sweet Spot for Monitoring Networks
Authors:
Nofel Yaseen,
Behnaz Arzani,
Krishna Chintalapudi,
Vaishnavi Ranganathan,
Felipe Frujeri,
Kevin Hsieh,
Daniel Berger,
Vincent Liu,
Srikanth Kandula
Abstract:
Continuously monitoring a wide variety of performance and fault metrics has become a crucial part of operating large-scale datacenter networks. In this work, we ask whether we can reduce the costs to monitor -- in terms of collection, storage and analysis -- by judiciously controlling how much and which measurements we collect. By positing that we can treat almost all measured signals as sampled t…
▽ More
Continuously monitoring a wide variety of performance and fault metrics has become a crucial part of operating large-scale datacenter networks. In this work, we ask whether we can reduce the costs to monitor -- in terms of collection, storage and analysis -- by judiciously controlling how much and which measurements we collect. By positing that we can treat almost all measured signals as sampled time-series, we show that we can use signal processing techniques such as the Nyquist-Shannon theorem to avoid wasteful data collection. We show that large savings appear possible by analyzing tens of popular measurements from a production datacenter network. We also discuss the technical challenges that must be solved when applying these techniques in practice.
△ Less
Submitted 11 October, 2021;
originally announced October 2021.
-
ZORB: A Derivative-Free Backpropagation Algorithm for Neural Networks
Authors:
Varun Ranganathan,
Alex Lewandowski
Abstract:
Gradient descent and backpropagation have enabled neural networks to achieve remarkable results in many real-world applications. Despite ongoing success, training a neural network with gradient descent can be a slow and strenuous affair. We present a simple yet faster training algorithm called Zeroth-Order Relaxed Backpropagation (ZORB). Instead of calculating gradients, ZORB uses the pseudoinvers…
▽ More
Gradient descent and backpropagation have enabled neural networks to achieve remarkable results in many real-world applications. Despite ongoing success, training a neural network with gradient descent can be a slow and strenuous affair. We present a simple yet faster training algorithm called Zeroth-Order Relaxed Backpropagation (ZORB). Instead of calculating gradients, ZORB uses the pseudoinverse of targets to backpropagate information. ZORB is designed to reduce the time required to train deep neural networks without penalizing performance. To illustrate the speed up, we trained a feed-forward neural network with 11 layers on MNIST and observed that ZORB converged 300 times faster than Adam while achieving a comparable error rate, without any hyperparameter tuning. We also broaden the scope of ZORB to convolutional neural networks, and apply it to subsamples of the CIFAR-10 dataset. Experiments on standard classification and regression benchmarks demonstrate ZORB's advantage over traditional backpropagation with Gradient Descent.
△ Less
Submitted 17 November, 2020;
originally announced November 2020.
-
A New Backpropagation Algorithm without Gradient Descent
Authors:
Varun Ranganathan,
S. Natarajan
Abstract:
The backpropagation algorithm, which had been originally introduced in the 1970s, is the workhorse of learning in neural networks. This backpropagation algorithm makes use of the famous machine learning algorithm known as Gradient Descent, which is a first-order iterative optimization algorithm for finding the minimum of a function. To find a local minimum of a function using gradient descent, one…
▽ More
The backpropagation algorithm, which had been originally introduced in the 1970s, is the workhorse of learning in neural networks. This backpropagation algorithm makes use of the famous machine learning algorithm known as Gradient Descent, which is a first-order iterative optimization algorithm for finding the minimum of a function. To find a local minimum of a function using gradient descent, one takes steps proportional to the negative of the gradient (or of the approximate gradient) of the function at the current point. In this paper, we develop an alternative to the backpropagation without the use of the Gradient Descent Algorithm, but instead we are going to devise a new algorithm to find the error in the weights and biases of an artificial neuron using Moore-Penrose Pseudo Inverse. The numerical studies and the experiments performed on various datasets are used to verify the working of this alternative algorithm.
△ Less
Submitted 25 January, 2018;
originally announced February 2018.
-
Design of Improved Quasi-Cyclic Protograph-Based Raptor-Like LDPC Codes for Short Block-Lengths
Authors:
Sudarsan V. S. Ranganathan,
Dariush Divsalar,
Richard D. Wesel
Abstract:
Protograph-based Raptor-like low-density parity-check codes (PBRL codes) are a recently proposed family of easily encodable and decodable rate-compatible LDPC (RC-LDPC) codes. These codes have an excellent iterative decoding threshold and performance across all design rates. PBRL codes designed thus far, for both long and short block-lengths, have been based on optimizing the iterative decoding th…
▽ More
Protograph-based Raptor-like low-density parity-check codes (PBRL codes) are a recently proposed family of easily encodable and decodable rate-compatible LDPC (RC-LDPC) codes. These codes have an excellent iterative decoding threshold and performance across all design rates. PBRL codes designed thus far, for both long and short block-lengths, have been based on optimizing the iterative decoding threshold of the protograph of the RC code family at various design rates.
In this work, we propose a design method to obtain better quasi-cyclic (QC) RC-LDPC codes with PBRL structure for short block-lengths (of a few hundred bits). We achieve this by maximizing an upper bound on the minimum distance of any QC-LDPC code that can be obtained from the protograph of a PBRL ensemble. The obtained codes outperform the original PBRL codes at short block-lengths by significantly improving the error floor behavior at all design rates. Furthermore, we identify a reduction in complexity of the design procedure, facilitated by the general structure of a PBRL ensemble.
△ Less
Submitted 6 June, 2017; v1 submitted 25 January, 2017;
originally announced January 2017.
-
Optimizing Transmission Lengths for Limited Feedback with Non-Binary LDPC Examples
Authors:
Kasra Vakilinia,
Sudarsan V. S. Ranganathan,
Dariush Divsalar,
Richard D. Wesel
Abstract:
This paper presents a general approach for optimizing the number of symbols in increments (packets of incremental redundancy) in a feedback communication system with a limited number of increments. This approach is based on a tight normal approximation on the rate for successful decoding. Applying this approach to a variety of feedback systems using non-binary (NB) low-density parity-check (LDPC)…
▽ More
This paper presents a general approach for optimizing the number of symbols in increments (packets of incremental redundancy) in a feedback communication system with a limited number of increments. This approach is based on a tight normal approximation on the rate for successful decoding. Applying this approach to a variety of feedback systems using non-binary (NB) low-density parity-check (LDPC) codes shows that greater than 90% of capacity can be achieved with average blocklengths fewer than 500 transmitted bits. One result is that the performance with ten increments closely approaches the performance with an infinite number of increments. The paper focuses on binary- input additive-white Gaussian noise (BI-AWGN) channels but also demonstrates that the normal approximation works well on examples of fading channels as well as high-SNR AWGN channels that require larger QAM constellations. The paper explores both variable-length feedback codes with termination (VLFT) and the more practical variable length feedback (VLF) codes without termination that require no assumption of noiseless transmitter confirmation. For VLF we consider both a two-phase scheme and CRC-based scheme.
△ Less
Submitted 16 February, 2016;
originally announced February 2016.
-
Optimality and Rate-Compatibility for Erasure-Coded Packet Transmissions when Fading Channel Diversity Increases with Packet Length
Authors:
Sudarsan V. S. Ranganathan,
Tong Mu,
Richard D. Wesel
Abstract:
A message composed of packets is transmitted using erasure and channel coding over a fading channel with no feedback. For this scenario, the paper explores the trade-off between the redundancies allocated to the packet-level erasure code and the channel code, along with an objective of a low probability of failure to recover the message.
To this end, we consider a fading model that we term propo…
▽ More
A message composed of packets is transmitted using erasure and channel coding over a fading channel with no feedback. For this scenario, the paper explores the trade-off between the redundancies allocated to the packet-level erasure code and the channel code, along with an objective of a low probability of failure to recover the message.
To this end, we consider a fading model that we term proportional-diversity block fading (PD block fading). For a fixed overall code rate and transmit power, we formulate an optimization problem to numerically find the optimal channel-coding rate (and thus the optimal erasure-coding rate) that minimizes the probability of failure for various approximations of the problem.
Furthermore, an interpretation of the results from an incremental redundancy point of view shows how rate-compatibility affects the possible trajectories of the failure probability as a function of the overall code rate. Our numerical results suggest that an optimal, rateless, hybrid coding scheme for a single-user wireless system over the PD block-fading channel should have the rate of the erasure code approach one.
△ Less
Submitted 1 February, 2016;
originally announced February 2016.
-
On the Girth of (3,L) Quasi-Cyclic LDPC Codes based on Complete Protographs
Authors:
Sudarsan V. S. Ranganathan,
Dariush Divsalar,
Richard D. Wesel
Abstract:
We consider the problem of constructing $(3,L)$ quasi-cyclic low-density parity-check (LDPC) codes from complete protographs. A complete protograph is a small bipartite graph with two disjoint vertex sets such that every vertex in the variable-node set is connected to every vertex in the check-node set by a unique edge. This paper analyzes the required lifting factor for achieving girths of six or…
▽ More
We consider the problem of constructing $(3,L)$ quasi-cyclic low-density parity-check (LDPC) codes from complete protographs. A complete protograph is a small bipartite graph with two disjoint vertex sets such that every vertex in the variable-node set is connected to every vertex in the check-node set by a unique edge. This paper analyzes the required lifting factor for achieving girths of six or eight in the resulting quasi-cyclic codes with constraints on lifting. The required lifting factors provide lower bounds on the block-length of such codes.
△ Less
Submitted 29 May, 2015; v1 submitted 20 April, 2015;
originally announced April 2015.