Hit-reco: ProtoDUNE Raw Data Denoising with DL models
summary
Uses 10 simulated ProtoDUNE events split into train/test/val with 24,000 crops of 32×32 sampled from planes sized 960×6000, trains two architectures: a ~1e5‑parameter ROI net and a ~1e6‑parameter denoiser, compares convolutional nets to graph convolutional nets (GCNN). GCNN captures nonlocal correlations and yields higher denoising SSIM while CNNs train 2.5–4× faster and deliver better ROI scores, achieving sensitivity 0.85 and ROC AUC 0.996; GCNN denoiser reaches PSNR ≈70.7.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:11DUNE Experiment
- 1:20Detector - LATTPC
- 3:02Raw Data
- 3:27Offline Data Processing
- 5:09Denoising
- 6:36Dataset
- 7:43Proposed Model
- 9:17Model overview - ROI
- 14:33Working Environment
- 14:54Training - DN
- 17:44Benchmark - ROI
- 20:28Results - ROI
- 21:24Benchmark - DN
- 22:17Results - DN
- 24:52Summary
related talks
39mML Denoising With Tensor Flow
33mLow to High Resolution
59mDeep Neural Networks for Video Applications
01h 30mDeep Learning Techniques for Monocular Depth Estimation
47mInterpreting Deep Neural Networks (DNNs)
57mTensorflow, deep learning and modern convolutional neural nets, without a PhD
29mHow AI boosted audio processing?
27mU-Net is all you need: applying the same AI approach for many different business domains