Istanbul Technical University

Signal Processing for Computational Intelligence Group

Open Source Code

This page gives the source code of selected works of our group. These works study parameter-efficient deep learning. They share weights between layers to decrease model size, computation and memory, so that models can run on devices with limited resources.

All repositories are on the SP4CING GitHub page.

HaLViT: Half of the Weights are Enough

CVPR 2024 Workshops (ELVM)

Koyun, O.C., Töreyin, B.U.

HaLViT uses one weight matrix and its transpose in place of two separate matrices for consecutive linear transformations. The method is applied to the bottleneck and attention layers of CNNs and Vision Transformers. On ImageNet and COCO, it decreases the number of parameters with a small effect on accuracy.

Code Paper Project page

HalF&D: A Parameter Efficient Small Object Detection Approach

IEEE ISCAS 2026 Builds on HaLViT

Yılmaz, A.N., Koyun, O.C., Töreyin, B.U.

HalF&D applies the HaLViT weight sharing to both stages of the Focus&Detect small object detection framework: ResNet-50 in the Focus stage and ResNeXt-101-32x8d in the Detect stage. On the VisDrone aerial image dataset, the number of parameters decreases by 47.3% in the Focus stage and by 70.9% in the Detect stage. The total model size decreases by more than 60%.

Code Paper details Focus&Detect

HALSP-Net: A Shared Projection Architecture with Dynamic Channel Selection

Ongoing work Builds on HaLViT

Hierarchical Active channel Latent Shared Projection Network

HALSP-Net is a compact CNN. In each stage, one weight matrix does the entry projection, the inner channel mixing and the exit projection. Each block also has its own depthwise spatial filter. During training, only a dynamic subset of channels is active, which decreases memory use. Inference uses the full dense matrix. On CIFAR-100, it gives accuracy comparable to deep and wide residual networks with fewer than two million parameters.

Code