Zero-Shot Learning (ZSL)

 Zero-Shot Learning (ZSL) is when a model can correctly classify or generate outputs for unseen categories without being explicitly trained on them.

---

1️⃣ Example: Zero-Shot Image Classification (Using CLIP)

💡 Suppose we have a model trained on general image-text pairs but never trained specifically to classify a "zebra."
👉 Yet, when we give it an image of a zebra, it can classify it correctly by understanding textual descriptions.

from transformers import CLIPProcessor, CLIPModel
from PIL import Image
# Load the CLIP model
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
# Load an image (Example: A zebra image)
image = Image.open("zebra.jpg")  # Replace with your image path
# Define candidate labels (even if the model has never explicitly seen them)
labels = ["cat", "dog", "horse", "zebra", "elephant"]
# Process the image and labels
inputs = processor(text=labels, images=image, return_tensors="pt", padding=True)
# Predict using CLIP
outputs = model(**inputs)
logits_per_image = outputs.logits_per_image  # Similarity scores
probs = logits_per_image.softmax(dim=1)  # Convert to probabilities
# Print the predicted label
predicted_label = labels[probs.argmax()]
print(f"Predicted label: {predicted_label}")

✅ Even if the model was never trained on "zebra," it can recognize it using text descriptions!

---

2️⃣ Example: Zero-Shot Text Classification

💡 Suppose we have a sentiment analysis model trained only on positive/negative reviews, but we want it to classify news articles into topics without retraining.

from transformers import pipeline
# Load a zero-shot classification pipeline
classifier = pipeline("zero-shot-classification", model="facebook/bart-large-mnli")
# Input text (Example: A finance-related news article)
text = "The stock market saw a significant rise in tech shares today."
# Candidate labels (The model was never explicitly trained on these)
labels = ["sports", "politics", "finance", "technology"]
# Classify the text
result = classifier(text, candidate_labels=labels)
print(result)

✅ Even without training on these specific categories, the model can classify the news article correctly!

---

3️⃣ Applications of Zero-Shot Learning

🔹 Image recognition (e.g., CLIP for unseen object classification).
🔹 Text classification (e.g., news categorization, intent recognition).
🔹 Machine translation (translate to a language not seen before).
🔹 Medical AI (detect diseases from descriptions instead of labeled data).