[WACV 2025 Oral] Calib3D: Calibrating Model Preferences for Reliable 3D Scene Understanding

PyTorch implementation for "Long Horizon Temperature Scaling", ICML 2023

Implemented GPT from scratch

WeldFusionNet , a weld defect classification project using sensor, audio, and video data, with feature pipelines, interactive analytics, and inference for manufacturing quality control (F1: 0.9567).

Newton–Puiseux for CVNNs: complete toolkit for uncertainty mining, confidence calibration and local symbolic-numeric analysis on ECG (MIT-BIH) and wireless IQ data (RadioML 2016.10A).

A library for calibrating and visualizing the Softmax function for probabilistic interpretation.

post-hoc calibration on GPU

Code for enhancing Conformal Prediction using Temperature Scaling. Explore more of our work at:

Calibrated lightweight chest X-ray triage for tuberculosis in resource-constrained settings. Reproducible code and reference deployment.

DermaSense - CNN-based skin lesion classification (HAM10000, 7 classes) with calibrated 4-model ensemble, Grad-CAM explainability, and a Flutter clinical-style UI. UTS 42028 coursework. Research/educational use only.

Official code for the TMLR 2025 paper: "On Joint Regularization and Calibration in Deep Ensembles"

Official repository for the paper "A Real-World Evaluation of Failure Detection for Liver CT Segmentation"

Advanced Deepfake image detection framework using Transfer Learning (ResNet50), optimized with TTA, architectural tuning, and confidence calibration.

Experiments with Temperature Scaling. Based on the original paper.

Fine-Grained Image Classification of World Architecture: An EfficientNetV2 Transfer Learning Approach with Layered Regularization

Self-evaluation framework for LLM confidence calibration. Extracts True/False logits on TriviaQA; temperature scaling reduces ECE from 0.217→0.132 (Qwen-2.5-1.5B) without model modification.

A speaker identification pipeline built with TensorFlow and trained on the LibriSpeech corpus. The model classifies 10-second audio clips as belonging to one of 317 speakers using MFCC features and a transformer-inspired architecture.

Confidence calibration toolkit for LLM verbalized-probability outputs. Real benchmark on 998 BoolQ questions with Llama-3.1-8B: ECE 0.148 -> 0.030, log-loss 3.9 -> 0.41.