Emirhan Bilgiç

First Open-Source Turkish Text-to-Speech Model

• Developed the first open-source Turkish TTS model using SpeechT5. Fine-tuned the transformer-based architecture on a Turkish Voice Dataset for speech synthesis.
• Created a Hugging Face demo and YouTube tutorial, providing the full dataset, training codes, and model weights. Engineered phoneme tokenization for improved speech quality.

Neural Network Design for Solving the Data Class Imbalance Problem - Final Thesis

• Designed a novel neural network to tackle class imbalance in datasets, incorporating dynamic data adjustment.
• Enhanced model performance by implementing automatic hyperparameter optimization and ensemble learning techniques, significantly improving accuracy in imbalanced datasets.

LLM From Scratch and Instruction Finetuning with LoRA

• Developed a GPT model from scratch, leveraging attention mechanisms and various byte pair encoder implementations to optimize text data processing.
• Performed instruction finetuning with LoRA, followed by a benchmark evaluation using MMLU-style Q&A, LLM-based scoring, and human ratings.
• Evaluated model efficiency through various attention mechanism implementations and examined PyTorch buffers for causal attention.

Deepfake Detection Survey and a VLP Approach

• Conducted a comprehensive survey of deepfake detection techniques, comparing methods in cross-datasets.
• Introduced a Vision-Language Pre-training (VLP) approach to enhance detection by analyzing image manipulations through targeted questioning. Addressed cross-dataset generalization challenges.

Predicting the Next Frames of a Video with AI | Next-Frame Prediction

• Addressed the limitations of CNNs in predicting future video frames by implementing advanced models like CNN-LSTM and PredRNN, and conducted a comparative analysis of their effectiveness.
• Solved a real-life problem by using cloudiness satellite image data of EUMETSAT on the weather forecast.

HairCLIPv2: Unifying Hair Editing via Proxy Feature Blending

• Implemented the paper HairCLIPv2: Unifying Hair Editing via Proxy Feature Blending presented at ICCV 2023 as a Replicate demo, making it publicly accessible for hair editing through proxy feature blending.

Brain Tumor Segmentation

• Benchmarked success using various techniques to solve the brain tumor segmentation problem.
• In addition to deep learning techniques such as SegNet and UNet; used machine learning techniques such as Adaptive Thresholding, Otsu's Method, K-Means, and Fuzzy C-Means.
• Using UNet, trained a model and created a Hugging Face demo for hair segmentation.

Face Morphing to Identify Future Baby or Person’s Future State

• Used facial landmark detection algorithms, IP-Adapter, and Face ID to embed facial features for diffusion model output generation based on prompts.
• Created a Hugging Face demo to demonstrate the process.

Creating Images with AI | DCGAN from Scratch

• Developed a DCGAN architecture from scratch to generate images resembling a given dataset.
• Enhanced the project by integrating DALL-E and Google Translate API to compare outputs and enable language-agnostic image generation.

Creating a Neural Network | Flexible Neural Nets from Scratch and Implementation at System Recognition

• Wrote a neural network in Python with adjustable layers and neurons using object-oriented programming (OOP) approach.
• Coded the mathematical infrastructure using only basic libraries like NumPy and Pandas, and implemented a static version in MATLAB.

PDF Reader

• Created a PDF reader Hugging Face demo using Parler TTS model.
• Engineered output combination due to token limitations and enabled translation using Helsinki NLP open-source models.

Hairstyler

• Created a Hugging Face demo to style hair using an inpainting method with ControlNet to protect other details in images.
• Segmented the hair with a custom hair segmentation model (UNet) and inpainted the hair area using ControlNet inpainting.