prunning

Tensorflow and Keras implementation of the state of the art researches in Dialog System NLU

Single shot neural network pruning before training the model, based on connection sensitivity

The goal of this project was to make a User vs AI chess game, such that user makes whichever (legal) move they like and in turn, the AI performs the best move it can find. The game ends when either the user or the AI enters checkmate or stalemate state - or if you quit forcibly :)

Code for paper "Deep Reinforcement Learning based Multi-task Automated Channel Pruning for DNNs"

Dynamic Sparsity Is Channel-Level Sparsity Learner [Neurips 2023]

QDP Studio is a unified framework for deep learning model compression. It combines quantization, pruning, and decomposition to reduce model size, improve inference speed, and maintain accuracy. Its streamlined pipeline for training, compressing, and evaluating models optimizes deployments in resource-constrained environments.

SparK: Query-Aware Unstructured Sparsity with Recoverable KV Cache Channel Pruning

This project implements a Decision Tree Classifier and demonstrates how pre-pruning parameters (such as max depth and minimum samples) affect model complexity and accuracy

Practical experiments and research notes on compressing large language models — quantization (INT4/AWQ), structured pruning, MoE expert pruning, and KV-cache optimization — with a fixed benchmarking methodology and reproducible experiment structure.

Statistical machine learning

MiniMax Python Othello with Alpha–beta pruning

Federated Learning with client selection strategies (Random, Loss-based, Power-of-Choice), domain generalisation and model pruning. Applied to image classification (FEMNIST) and semantic segmentation (IDDA). PyTorch.

Termal ve RGB havadan görüntülerde nesne tespiti — mimari karşılaştırması, sınıf dengesizliği deneyleri, budama, kuantizasyon ve Jetson/Raspberry Pi uç cihaz ölçümleri (SAYZEK ATP ekip projesi). Katkım: DETR/Faster R-CNN eğitimi ve kayıp fonksiyonu deneyleri.

Use decision trees to prepare a model on fraud data. Treating those who have taxable income <= 30000 as "Risky" and others are "Good" and A cloth manufacturing company is interested to know about the segment or attributes causes high sale.

Appyling pNML for deep nerual networks by purnning