Research | SCU CoAI Lab

Efficient AI Systems

Wed, 09 Sep 2026 00:00:00 +0000

On-device LLM inference through adaptive sparsity

We study how to make large language models practical on resource-constrained devices by exploiting activation sparsity—the computations that a model does not need for a particular input. Our work spans the full inference stack: predicting sparsity before expensive operations begin, translating those predictions into skipped memory accesses and computation, and adapting the amount of sparsity at runtime so that efficiency gains do not come at an uncontrolled cost to model quality.

Students: Ethan Lin · Suri Wang

Collaborators: Youngmin Yi · Joon-Sung Yang · Youngpyo Joo · Chae Eun Rhee

Selected publications

Closed-loop control adapts sparsity to measured output quality.

IEEE TCAD · CODES 2026

Adaptive sparsity control

ACAS treats sparsity as a runtime control problem. It monitors a quality signal and adjusts sparse execution online, allowing an on-device system to respond to changing inputs and operating conditions instead of relying on one fixed threshold.

Paper

Offline structure enables efficient prediction and GPU execution.

ASPLOS 2026

Structured FFNs for fast, accurate inference

oFFN jointly organizes activation outliers and hot and cold neurons within the FFN. This structure supports accurate sparsity prediction and selects efficient computation paths as batch size changes, translating sparsity into practical end-to-end acceleration.

Paper

Predict sparsity before loading the corresponding FFN weights.

DAC 2025 · DATE 2025

Efficient activation-sparsity prediction

SparseInfer uses a lightweight, training-free predictor to expose activation sparsity before computation. Grasp extends this direction with group-based, magnitude-aware prediction, improving the balance between prediction overhead and retained model quality.

Paper 1 (SparseInfer)·Paper 2 (Grasp)

Real-Time Intelligent Systems

Wed, 09 Sep 2026 00:00:00 +0000

Intelligent decisions under timing and resource constraints

We design cyber-physical and robotic systems that make useful decisions within explicit timing and resource limits. Our work combines real-time scheduling, executable models of hardware–software interaction, and learning-enabled physical control so that an intelligent system can adapt to uncertainty without giving up predictable behavior.

Student: Tejeswini Jayaramareddy

Collaborator: Hokeun Kim

Selected publications

An RSU prioritizes authentication as traffic and deadlines change.

IEEE ESL 2026

Value-aware scheduling for intelligent transportation systems

At a busy intersection, a roadside unit may not have enough time or compute capacity to authenticate every connected-vehicle request. Our scheduler prioritizes feasible authentication task graphs by their expected contribution to system utility, increasing the value completed before global deadlines under heavy traffic or limited processor capacity.

Paper

Explicit timing models reveal safe interrupt and FIFO configurations.

IEEE ESL 2026

Timing-aware integration of third-party automotive subsystems

Automotive systems often integrate third-party GNSS and other peripherals through finite FIFOs and interrupt-driven microcontroller interfaces. We model the complete hardware-software path and derive safe operating thresholds systematically, because timing choices can affect end-to-end functionality rather than performance alone.

Paper

Onboard sensing and control enable autonomous vertical burrowing.

ICRA 2024

Low-power RL-assisted robotics

Our snake-like robot learns coordinated motions for vertical self-burrowing in granular material, where physical forces are difficult to model. Onboard inertial and magnetic sensing provide feedback for reaching target depths, illustrating how learning and embedded intelligence can enable adaptive physical behavior.

Paper

Across these projects, our goal is adaptive intelligence with explicit timing, safety, and resource awareness—from individual embedded devices to distributed sensing and control pipelines.

Resource-Efficient Computer Vision

Wed, 09 Sep 2026 00:00:00 +0000

Efficient perception under real resource limits

We design computer-vision systems that reduce computation and latency without treating accuracy as an afterthought. Our work simplifies detector pipelines, targets the stages that dominate runtime, and adapts model execution to the structure of each application so that capable perception can move closer to sensors and resource-constrained platforms.

Student: Rikesh Naresh Mehta

Collaborator: Hyungshin Kim

Selected publication

A single-feature detector reduces computation while preserving accuracy.

IEEE Access 2025

Efficient two-stage detection for on-board remote sensing

We simplify feature extraction in two-stage detectors and recover accuracy through better positive-anchor selection and high-pass filtering in the region proposal network. The resulting design lowers computation and improves throughput while retaining competitive detection accuracy on high-resolution satellite imagery.

Paper

Our ongoing work under review extends this direction through joint optimization of DCNN inference and non-maximum suppression for satellite imagery, targeting higher throughput with tightly controlled accuracy loss.

Secure & Reliable Systems

Wed, 09 Sep 2026 00:00:00 +0000

Adaptive protection across changing conditions

We build systems that remain safe and dependable as physical conditions, workloads, and attack surfaces change. Our research connects hardware behavior with runtime monitoring and lightweight control, seeking protection mechanisms that respond to the environment while retaining measurable security guarantees and bounded overhead.

Student: Aziz Alajmi

Selected publication

Hotter DRAM lowers the safe threshold and triggers earlier mitigation.

DRAMSec 2026

Temperature-aware RowHammer protection

Dynamic RowHammer Threshold Management adapts defense thresholds to live temperature and device variation. The study shows how runtime calibration can remove PRAC’s high-temperature staleness breaches and substantially reduce SALT-C breaches with limited latency overhead, while also exposing defenses whose protection is constrained by capacity rather than threshold choice.

Paper

This work reflects our broader direction: cross-layer resilience grounded in explicit evidence about protection, performance cost, and the limits of each mitigation mechanism.