Effective Communication Quiz

Deep Learning DSAI 308

• Course instructors are Dr. Khaled Mostafa ElSayed and Dr. Mohamed Maher Ata
• Course is offered at the University of Science and Technology, Zewail City

Siamese Network

• Siamese networks are used in distance-based learning methods, such as face recognition and signature verification
• Useful for feature learning, as a feature extractor
• Also used for dimensionality reduction, enabling data visualization to show class separability
• These networks use a convolutional architecture
• The network maps input patterns into a target space through non-linear mappings, transforming inputs to a lower-dimensional space
• Distances in this space approximate the semantic distance in the input space
• The training method is discriminative, extracting information from available data, without requiring category-specific details
• Training is conducted on pairs, triples, or quads of samples

Traditional Classification Approach

• Traditional approaches to classification, exemplified by neural networks or support vector machines, typically require predefined categories and training examples for each category
• These methods are inherently limited to a fairly small number of categories (around 1,000)

Why Another Approach (Siamese Networks)

• Siamese networks address situations where the number of categories is high
• Cases where samples per category are limited
• When only a subset of categories is known at the beginning of training

Example Applications (Face Recognition)

• Many categories (e.g., thousands of individuals)
• Few training examples per category (e.g., several images per individual)
• A common approach is using distance-based methods
• This involves calculating a similarity metric between the input pattern (e.g. face) and stored prototypes

Siamese Network Architecture (Training Phase)

• Consists of a convolutional part and a similarity measurement part
• Images are mapped to latent representations (embeddings)
• Networks often involve shared weights for efficient computation

Siamese Network Architecture (Prediction Phase)

• Uses pre-trained CNNs
• Embeddings of available images are stored in a database.
• During testing, a query image is input to the pre-trained neural network to get its embedding.
• The embedding is then compared to the embeddings in the database
• The classifier uses these comparisons to predict the class

Siamese Network Architecture (Question Answering)

• Non-siamese networks use separate question and answer encoders to calculate cosine similarity
• Siamese networks use a single encoder for both question and answer input
• The single encoder is used to calculate similarity metrics between questions and answers

Siamese Network: Two Inputs Architecture

• The network accepts two inputs (images, patches, etc.)
• This architecture is often used to assess the similarity between the two inputs

Using Cross Entropy Loss (With Image Pairs)

• Loss functions are used to measure the difference between predicted and actual outputs
• Cross-entropy loss is a common function in classification problems

Input Pairs to Siamese Networks: Positive & Negative Pairs

• Positive pairs refer to similar input images (e.g., two images of the same person)
• Negative pairs represent dissimilar input images (e.g., images of different individuals)

Training Siamese Networks

• Create a dataset with labeled image pairs indicating similarity
• Similarity information is crucial, not categories
• Appropriate metrics of similarity within the input space are required (e.g., distance metrics)
• The training involves forward pass (computing loss) and backward pass (updating weights).
• Weights of identical networks are frequently averaged to create consistency

Review: Cosine Similarity

• Vectors must be normalized before measurement.
• This measure is based on the dot product of two normalized input vectors, which falls between -1 and 1 in values

Few Shots Prediction

• Instead of processing all available samples, the approach uses the average embedding vector of each class for prediction.
• All embeddings are normalized
• Cosine similarities are calculated between the unknown input and class averages.
• The class with the highest similarity probability is predicted
• The "Top 5" most likely classes can be also provided

Improving Similarity Learning (by using ensemble learning, choosing good loss and augmenting training pairs)

• Using loss functions (e.g., triplet, contrastive loss)
• Generating diverse training samples
• Execution time is improved by dividing training data into clusters and using multiple networks in parallel

Transfer Learning for Siamese Networks

• Leveraging trained models (like ResNet-50) for feature extraction
• Trainable modules are used to adapt the feature extraction task

Comparison between Different Pretrained Models

• Different pre-trained models are evaluated for use as the backbone in Siamese networks, comparing their top-1 accuracies on datasets like ImageNet

Types of Siamese Networks Based on Provided Inputs

• Two-input networks use cross-entropy or contrastive loss
• Three-input networks utilize triplet loss
• Four-input models employ quadruplet loss

Applications of Siamese Networks

• Face verification and classification
• MNIST Clustering
• Emotion Prediction (using the NimStim data set)

Traditional Techniques for Establishing Image Correspondence

• These techniques use edge features from consecutive images, calculating transformations to match images.

Establishing Image Correspondence Using Siamese Networks

• This approach uses localized sections (patches) to find similar image parts without relying on classical methods like SIFT/SURF

Signature Verification with Siamese Networks

• Siamese architecture is suitable for evaluating the similarity between signatures

Code Snippets for Siamese Network Implementation

• Code examples are provided for image pairs creation, building Siamese models, and calculating Euclidean loss using MNIST datasets used.

Test your knowledge on the key principles of effective communication and conflict resolution. This quiz covers essential aspects such as body language, teamwork, and common barriers to communication. Perfect for improving your interpersonal skills.