About Our Research

The Ko Lab designs the hardware and the algorithms together. The question running through all of it is the same one: how do you get modern AI to run on hardware that is small, cheap and power-constrained — and how do you know you can trust the answer it gives?

Our work spans five areas, from the arithmetic circuits inside a processor to deployed systems in hospitals and on farms.

Hardware Design for Deep Learning

About the Area

Block diagram of an accelerator processing element with input, filter and partial-sum scratchpads.
A processing element with early zero detection.

Arithmetic units and accelerator architectures that make deep learning cheaper to compute — multiple-precision fused multiply-add units, approximate multipliers and dividers, multiplier-free activation functions, sparse-network microarchitecture and spiking neural network processing.

Explore hardware design for deep learning

Efficient and Compressed AI Models

About the Area

Diagram comparing a twelve-layer BERT model with a compressed six-layer version.
Twelve-layer BERT beside its compressed six-layer counterpart.

Large models do not fit on edge hardware. We shrink them through pruning, quantization and architecture search — including a short-training proxy that makes exhaustive transformer pruning search affordable, and a more stable route to ternary quantization.

Explore efficient and compressed AI models

Security, Privacy and Cryptography

About the Area

Table comparing logic cells, accuracy and training throughput for differential privacy accelerator designs.
Differential privacy accelerator results on ZCU102.

Efficient AI is not enough if it cannot be trusted. We build differential privacy accelerators that generate Gaussian noise in hardware, lightweight intrusion detection small enough for constrained devices, and cryptographic circuits for IoT and post-quantum settings.

Explore security, privacy and cryptography

Medical Imaging and Biomedical AI

About the Area

Pipeline diagram showing chest X-ray region extraction feeding a DenseNet-121 backbone and classifier.
The COVID-CXNet chest radiograph pipeline.

Working with radiologists and clinical collaborators, we build systems that read medical images and biomedical signals — early detection of ankylosing spondylitis, rib fracture classification, chest radiographs, breast ultrasound, dental caries and EEG — and we test them against the specialists they are meant to support.

Explore medical imaging and biomedical AI

Computer Vision and Applied AI

About the Area

Overhead camera view of pigs in a barn pen, used as input for pose estimation.
Barn camera footage used for pose estimation and behaviour analysis.

The same methods solve problems well outside the clinic: livestock welfare monitoring, crop yield prediction from drone hyperspectral imagery, licence plate and commercial vehicle inspection, super-resolution, and document and text understanding.

Explore computer vision and applied AI