Unsupervised Discriminative Feature Alignment for Domain Adaptation in WiFi-Based HAR

Generated from prompt:

Create a professional 5-slide presentation summarizing the uploaded paper “Unsupervised Discriminative Feature Alignment for Domain Adaptation in WiFi-Based Human Activity Recognition” by Amany Elkelany, Robert Ross, and Susan McKeever. Use a clean academic IEEE-style theme with blue/white colors, concise bullets, and clear visuals. Exactly 5 slides. Slide 1 — Title & Motivation Title: Unsupervised Discriminative Feature Alignment for Domain Adaptation in WiFi-Based HAR Subtitle: FA-DANN for robust cross-environment activity recognition using WiFi CSI signals Key points: WiFi-based HAR is privacy-preserving, device-free, and cost-effective; but models trained in one environment often fail in another because CSI patterns change with room layout, obstacles, furniture, subjects, and hardware. Visual: WiFi sensing in two different rooms showing domain shift. Slide 2 — Problem & Research Gap Main message: Cross-domain WiFi HAR is difficult without labeled target data. Bullets: Same-domain HAR performs well, but cross-domain performance drops sharply; target labels are expensive and time-consuming; DANN-style methods align domains globally but can lose class-discriminative activity structure; Gaussian feature alignment can oversimplify complex non-Gaussian CSI distributions. Visual: source-domain and target-domain feature distributions before adaptation. Slide 3 — Proposed Method: FA-DANN Main message: FA-DANN combines feature alignment with adversarial domain adaptation. Include: CNN-ABiLSTM activity classifier; shared encoder-decoder autoencoder; domain classifier with gradient reversal layer; unpaired L1 loss aligns reconstructed source and target features; trained with labeled source data and unlabeled target data only. Visual: architecture flow: source/target CSI → encoder → decoder → activity classifier + domain classifier/GRL. Slide 4 — Experiments & Key Results Main message: FA-DANN substantially improves cross-environment HAR. Include datasets: GJWiFi, OPERAnet, SHARP-2. Include preprocessing: CSI amplitude extraction, Hampel denoising, Z-score normalization, 1-second windows. Include key findings: no adaptation average F1-score = 14.34%; FA-DANN average F1-score = 85.57%; average improvement over state-of-the-art = 19.98%. Mention 5-fold cross-validation and cross-domain evaluation. Visual: bar chart comparing No DA vs FA-DANN, plus dataset labels. Slide 5 — Takeaways & Deployment Value Main message: FA-DANN enables scalable WiFi HAR in unseen environments without labeled target data. Bullets: preserves discriminative activity semantics; avoids Gaussian prior assumptions; robust under large domain shifts and noisy CSI; computationally practical for edge inference, around 4.27M–6.37M FLOPs per forward pass; ablation studies confirm the importance of both decoder and domain classifier. Future work: broader deployments, more users/devices, and model compression. Visual: deployment pipeline from WiFi access point to edge device to activity label. Make slides polished and presentation-ready, with speaker-friendly wording and minimal dense text.