Zero-Shot Learning Unseen Class Recognition

# Zero-Shot Learning: Unseen Class Recognition

## Introduction & Motivation

Zero-Shot: recognize unseen classes. Semantic attributes; auxiliary information. Transfer knowledge without examples. Applications: novel object recognition, rare classes.

Motivation: Generalize to never-before-seen classes.

Applications: Novel classes, scaling.

---

## Core Concepts & Theory

### Semantic Attributes

Class descriptions; transfer knowledge.

### Embedding Alignment

Map features to semantic space.

### Cross-Modal Transfer

Different modalities; shared embedding.

---

## Mathematical Formulation

Attribute prediction:
$$\hat{a} = f(x) \approx a_y$$

Semantic embedding:
$$ ext{sim}(f(x), s_y) = \max_y$$

---

## Summary & Key Takeaways

Zero-Shot Learning via semantic attributes enables recognition of unseen classes through knowledge transfer.

---

---

## Appendix: Practical Labs

### Lab 1: Attribute Prediction

import numpy as np

def predict_attributes(features, attribute_classifier):
 """Predict semantic attributes from visual features"""
 attributes = features @ attribute_classifier
 return attributes

# Test
np.random.seed(42)
features = np.random.randn(50, 256)
classifier = np.random.randn(256, 10)

attrs = predict_attributes(features, classifier)

assert attrs.shape == (50, 10), "Attribute shape"
print("✓ Attribute prediction working")

if __name__ == "__main__":
 print("Lab 1: AttributePrediction - PASSED")

### Lab 2: Semantic Similarity

import numpy as np

def semantic_similarity(predicted_attrs, class_attributes):
 """Compute similarity to class semantic vectors"""
 similarities = predicted_attrs @ class_attributes.T
 predicted_class = np.argmax(similarities, axis=1)
 return predicted_class

# Test
np.random.seed(42)
pred_attrs = np.random.randn(50, 10)
class_attrs = np.random.randn(20, 10)

classes = semantic_similarity(pred_attrs, class_attrs)

assert classes.shape == (50,), "Class predictions"
print("✓ Semantic similarity working")

if __name__ == "__main__":
 print("Lab 2: SemanticSimilarity - PASSED")

### Lab 3: Generalized Zero-Shot

import numpy as np

def generalized_zero_shot_evaluation(predictions, true_labels, seen_classes, unseen_classes):
 """Evaluate on both seen and unseen classes"""
 seen_acc = (predictions[true_labels < len(seen_classes)] == true_labels[true_labels < len(seen_classes)]).mean()
 unseen_acc = (predictions[true_labels >= len(seen_classes)] == true_labels[true_labels >= len(seen_classes)]).mean()
 
 harmonic_mean = 2 * seen_acc * unseen_acc / (seen_acc + unseen_acc + 1e-8)
 
 return seen_acc, unseen_acc, harmonic_mean

# Test
np.random.seed(42)
preds = np.random.randint(0, 20, 100)
labels = np.random.randint(0, 20, 100)

seen_acc, unseen_acc, h_mean = generalized_zero_shot_evaluation(preds, labels, list(range(10)), list(range(10, 20)))

assert 0 <= h_mean <= 1, "H-mean valid"
print("✓ Generalized zero-shot working")

if __name__ == "__main__":
 print("Lab 3: GeneralizedZeroShot - PASSED")

### Lab 4: Transfer via Attributes

import numpy as np

def transfer_via_attributes(source_features, target_classes, attribute_matrix):
 """Transfer knowledge via shared attribute space"""
 # Embed source in attribute space
 source_attrs = source_features.mean(axis=0, keepdims=True)
 
 # Find closest target classes in attribute space
 similarities = source_attrs @ attribute_matrix.T
 
 return similarities

# Test
np.random.seed(42)
src_feat = np.random.randn(50, 256)
tgt_classes = np.random.randn(10, 20)
attr_matrix = np.random.randn(20, 20)

sims = transfer_via_attributes(src_feat, tgt_classes, attr_matrix)

assert sims.shape == (1, 10), "Similarities shape"
print("✓ Attribute transfer working")

if __name__ == "__main__":
 print("Lab 4: AttributeTransfer - PASSED")