Applications — Denoising & Dimensionality Reduction

अब हम Autoencoders के दो प्रायोगिक उपयोगों (applications) को विस्तार से समझेंगे —
🔹 Denoising
🔹 Dimensionality Reduction

ये दोनों real-world problems में बहुत उपयोगी हैं और deep learning की शक्ति को बख़ूबी दर्शाते हैं।

---

🔶 1. Application 1: Denoising Autoencoder

❓ What is it?

Denoising Autoencoder (DAE) एक ऐसा Autoencoder है जो noisy input को clean output में बदलना सीखता है।

🎯 “Input को जानबूझकर corrupt किया जाता है और model को सिखाया जाता है कि वह clean version reconstruct करे।”

---

📦 Working:

   Original Image (x)
         ↓ Add Noise
     Noisy Input (x̃)
         ↓
     [Encoder + Decoder]
         ↓
   Clean Output (x̂)

Model learns to minimize:

---

🔧 Example in PyTorch:

def add_noise(imgs, noise_factor=0.3):
    noisy_imgs = imgs + noise_factor * torch.randn_like(imgs)
    return torch.clip(noisy_imgs, 0., 1.)

You then train autoencoder with (noisy_img, original_img) pairs.

---

🧠 Use Cases:

Use	Description
🖼️ Image Denoising	Remove noise from pictures
📄 Document Cleanup	Clean scanned papers
📢 Audio Denoising	Remove background noise
🧠 Medical	Remove sensor noise in ECG, MRI, etc.

---

🔶 2. Application 2: Dimensionality Reduction

❓ What is it?

Autoencoder compresses high-dimensional data into a low-dimensional latent representation, similar to PCA (Principal Component Analysis) — but with non-linear capabilities.

🎯 “Autoencoder = Non-linear, trainable PCA”

---

📦 Example:

| Input | 784-dim vector (28×28 image)
| Encoder | Reduces it to 2D or 3D latent code
| Decoder | Reconstructs full image
| Output | Use latent codes for clustering, visualization, etc.

---

🔧 PyTorch Sketch:

# Encoder output is just 2D
self.encoder = nn.Sequential(
    nn.Linear(784, 128),
    nn.ReLU(),
    nn.Linear(128, 2)  # 2D Latent Space
)

---

🧠 Use Cases:

Use	Description
📊 Data Visualization	Compress to 2D for t-SNE / plots
🔍 Clustering	Group similar inputs (e.g., digits, faces)
⚡ Fast Inference	Work on lower-dimensional features
📈 Feature Extraction	Use compressed codes for ML models
🎮 Game AI	Compress game states

---

📝 Practice Questions:

1. Denoising Autoencoder क्या है और कैसे काम करता है?
2. Noise हटाने के लिए Autoencoder को कैसे train किया जाता है?
3. Dimensionality reduction में Autoencoder और PCA में क्या अंतर है?
4. Latent space का क्या role है?
5. Low-dimensional representation किन real-world problems में काम आता है?

---

📌 Summary

Application	Input	Output	Benefit
Denoising	Noisy image	Clean image	Noise removal
Dimensionality Reduction	High-dim data	Low-dim features	Visualization, clustering, compression