Name	Name	Last commit message	Last commit date
parent directory ..
00_CNN_RNN_ImageClassification	00_CNN_RNN_ImageClassification
01_AutoencoderTransferLearning	01_AutoencoderTransferLearning
02_VAE_FashionMNIST	02_VAE_FashionMNIST
03_GAN_FashionMNIST	03_GAN_FashionMNIST
README.md	README.md

Deep Learning Portfolio

Professional implementation of 5 advanced deep learning architectures, demonstrating mastery of CNN, RNN, Autoencoders, VAE, and GAN techniques for image classification, generation, and representation learning. All projects feature production-ready code, comprehensive documentation, and validated results.

🎯 Overview

This portfolio showcases 5 comprehensive deep learning projects covering the major neural network architectures used in modern computer vision and generative AI:

Architecture	Use Case	Key Metric	Complexity
CNN/RNN	Image Classification	88% test accuracy	Intermediate
Autoencoder	Feature Learning + Transfer	72.2% with 1,800 samples	Advanced
VAE	Generative Modeling	2D latent space visualization	Advanced
GAN	Image Synthesis	Realistic 28×28 generation	Expert
Ensemble	Multi-model Integration	Group collaboration	Advanced

Total Implementation: 5,000+ lines of Python code, 11 custom neural architectures, comprehensive evaluation frameworks

🛠️ Skills & Technologies

Core Frameworks & Libraries

tensorflow==2.10.0              # Deep learning framework
tensorflow-probability==0.18.0   # Probabilistic modeling (VAE)
keras                            # High-level neural network API
numpy==1.23.5                    # Numerical computing
matplotlib==3.7.2                # Visualization
scikit-learn                     # Machine learning utilities

Advanced Techniques Implemented

Architecture Design:

Convolutional Neural Networks (CNNs) for spatial feature extraction
Recurrent Neural Networks (RNNs/LSTM) for sequential processing
Encoder-Decoder architectures for dimensionality reduction
Generative Adversarial Networks (GANs) for image synthesis
Transfer Learning for low-data scenarios

Training Strategies:

Adversarial training with Nash equilibrium monitoring
Unsupervised pre-training for semi-supervised learning
β-weighting for loss function balancing
Custom early stopping with multi-metric heuristics
Gradient flow optimization via reparameterization trick

Model Evaluation:

Multi-class classification metrics (accuracy, confusion matrices)
Generative model quality assessment (visual inspection, latent space analysis)
Transfer learning validation (baseline vs pre-trained comparison)
Convergence analysis (loss oscillation, discriminator accuracy)

Cumulative Skills Demonstrated

Architectural Design - Implemented 11 distinct neural network architectures (CNNs, RNNs, Autoencoders, VAE, GAN)
Training Strategies - Mastered adversarial training, transfer learning, unsupervised pre-training, β-weighting
Custom Implementations - Built SampleLayer (reparameterization), GANEarlyStopping, multi-metric scoring
Critical Analysis - Identified architectural errors, validated through experiments, quantified improvements
Model Evaluation - Comprehensive metrics (accuracy, confusion matrices, latent space analysis, convergence detection)
Technical Communication - Professional documentation with theory, code, results, and real-world applications

📊 Projects

1. CNN vs RNN Image Classification

Location: 00_CNN_RNN_ImageClassification/

Objective: Systematic comparison of CNN and RNN architectures for Fashion MNIST classification, demonstrating when to use each approach.

Key Results:

✅ CNN achieved 88% test accuracy vs RNN's 86.2% - validated CNN superiority for spatial tasks
✅ Identified architectural mismatch - LSTM treats image rows as sequence (suboptimal for 2D spatial data)
✅ Category 6 (Shirts) most challenging - 65.9% accuracy due to visual similarity with T-shirts

Technical Highlights:

2 complete architectures: CNN (~140K params) and LSTM (~84K params)
Comprehensive evaluation: training curves, confusion matrices, probability distributions
Proper train-validation-test methodology with no overfitting

Real-World Applications:

Automated clothing categorization for e-commerce
Visual search systems ("find similar items")
Transfer learning foundations for custom datasets

View Full Documentation →

2. Transfer Learning with Denoising Autoencoder

Location: 01_AutoencoderTransferLearning/

Objective: Demonstrate how unsupervised pre-training on 57,000 unlabeled images improves CNN performance when only 1,800 labeled samples are available.

Key Results:

✅ Semi-supervised learning - Pre-trained model achieved 72.17% vs baseline 70.33% validation accuracy
✅ 4x improvement on hardest class - Category 6 accuracy: 32.4% (pre-trained) vs 8.8% (baseline)
✅ Denoising forces robust features - Noise factor 0.2 prevents memorization, learns true patterns

Technical Highlights:

3 distinct architectures: Baseline CNN, Denoising Autoencoder, Transfer Learning CNN
Frozen encoder layers (feature extraction) + trainable classifier head
A/B testing methodology proving value of unsupervised pre-training

Real-World Applications:

Medical imaging (few labeled scans, many unlabeled)
Rare object detection (limited annotated examples)
Domain adaptation (pre-train on similar dataset, fine-tune on target)

View Full Documentation →

3. Variational Autoencoder (VAE)

Location: 02_VAE_FashionMNIST/

Objective: Learn a 2D probabilistic latent space representation of Fashion MNIST, enabling both reconstruction and controlled generation.

Key Results:

✅ β-VAE with β=0.001 - Optimized weighting balances reconstruction quality and latent regularization
✅ Well-separated class clusters - 10 fashion categories occupy distinct 2D regions (z[0]: -3.66 to 2.76, z[1]: -3.46 to 3.25)
✅ Smooth interpolation - Walking between latent points produces realistic intermediate images

Technical Highlights:

Custom SampleLayer implementing reparameterization trick (enables backprop through sampling)
Principled β derivation via loss magnitude analysis (not trial-and-error)
2D latent space for human-interpretable visualization

Real-World Applications:

Fashion design prototyping (interpolate between existing designs)
Anomaly detection (high reconstruction error = defective product)
Data augmentation (synthesize training samples for limited datasets)

View Full Documentation →

4. Deep Convolutional GAN (DCGAN)

Location: 03_ConvolutionalGAN/

Objective: Build a DCGAN that generates realistic 28×28 grayscale pants images from random noise, demonstrating adversarial training convergence.

Key Results:

✅ Architectural error correction - Removed discriminator upsampling (50% parameter reduction: 424K → 213K)
✅ Balanced Nash equilibrium - Gen Loss: 0.6-0.9, Disc Loss: 1.2-1.6, Disc Accuracy: 55-60%
✅ Custom early stopping - GAN-specific scoring heuristic (accuracy deviation + loss ratio + range penalties)

Technical Highlights:

Critical analysis identified contradiction with DCGAN standard (Radford et al., 2015)
Generator: Progressive upsampling (7×7 → 14×14 → 28×28) via Conv2DTranspose
Discriminator: Progressive downsampling (28×28 → 14×14 → 7×7) via strided Conv2D
Training time: ~35 seconds on Colab T4 GPU vs hours/days on CPU

Real-World Applications:

Fashion design rapid ideation workflows
Data augmentation for computer vision models
Synthetic training data generation for object detection

View Full Documentation →

🔬 Key Technical Innovations

1. Custom Keras Layers for Probabilistic Modeling

Challenge: VAE requires sampling from N(μ, σ²) during forward pass, but sampling blocks gradient flow

Solution: Implemented SampleLayer with reparameterization trick:

class SampleLayer(tf.keras.layers.Layer):
    def call(self, inputs):
        z_mean, z_log_var = inputs
        epsilon = tf.random.normal(shape=tf.shape(z_mean))
        return z_mean + tf.exp(0.5 * z_log_var) * epsilon

Mathematical Foundation: z = μ + σ ⊙ ε, where ε ~ N(0,1)

Randomness separated from learned parameters (μ, σ)
Gradients flow through μ and σ, not ε
Enables end-to-end training of probabilistic models

2. Multi-Metric GAN Convergence Detection

Challenge: Standard early stopping fails for GANs (oscillating losses are healthy)

Solution: Custom scoring heuristic balancing multiple signals:

score = 1.0 / (1.0 + acc_penalty + ratio_penalty + gen_penalty + disc_penalty)

# Where:
# - acc_penalty: |disc_acc - 0.55| / 0.55
# - ratio_penalty: 2.0 if (gen_loss/disc_loss) outside [0.4, 0.8]
# - Loss range penalties for pathological values

Stopping Criteria:

Patience exhausted (15 epochs no improvement after 30 min epochs)
Convergence detected (loss variance < 0.001 over 20 epochs)

Result: Automatically detected best checkpoint (Epoch 51, score 0.8567)

3. β-Weighting Derivation for VAE Loss

Challenge: KL divergence loss magnitude often 100-1000x larger than reconstruction loss

Solution: Principled β calculation targeting 10-20% KL contribution:

# Step 1: Measure typical loss ranges
typical_mse = 0.06       # Reconstruction loss
typical_kl = 20          # KL loss (unweighted)

# Step 2: Solve for β
target_proportion = 0.15
β = (target_proportion * typical_mse) / typical_kl
  = (0.15 * 0.06) / 20
  = 0.00045
  ≈ 0.001  # Round for convenience

Validation: Post-training analysis confirmed 15% KL contribution with sharp reconstructions

🌍 Real-World Applications

Fashion & E-commerce

Automated Product Categorization (CNN) - Scale to millions of products
Visual Search (CNN + Transfer Learning) - "Find similar items" functionality
Design Prototyping (VAE + GAN) - Generate design variations, rapid ideation
Synthetic Data Generation (GAN) - Augment limited fashion datasets

Computer Vision Systems

Transfer Learning Pipelines (Autoencoder) - Pre-train on unlabeled data, fine-tune on target task
Anomaly Detection (VAE) - Identify defects via reconstruction error
Data Augmentation (GAN) - Generate training samples for rare classes
Feature Extraction (CNN) - Frozen encoders for downstream tasks

Research & Development

Architecture Comparison Studies (CNN vs RNN) - Validate architectural choices empirically
Semi-Supervised Learning (Denoising Autoencoder) - Leverage unlabeled data effectively
Latent Space Analysis (VAE) - Understand learned representations via 2D visualization
Adversarial Training (GAN) - Apply game-theoretic concepts to neural networks

⚙️ Environment Setup

Recommended: Conda Environment

All projects use Python 3.10 with TensorFlow 2.10.0 for consistency. Each project folder contains:

environment.yml - Conda environment specification
requirements.txt - pip dependencies (alternative)

# Navigate to any project folder
cd 00_CNN_RNN_ImageClassification/

# Create environment from YAML
conda env create -f environment.yml

# Activate environment (name varies by project)
conda activate dl-cnn-rnn  # or dl-autoencoder, vae-project, gan-project

# Run main script
python matheus_CNN_RNN.py  # Filename varies by project

Hardware Requirements

Project	CPU	GPU	Recommended
CNN/RNN	5-10 min	2-3 min	CPU acceptable
Autoencoder	10-15 min	3-5 min	CPU acceptable
VAE	30-35 min	3-4 min	GPU recommended
GAN	Hours/Days	35 seconds	GPU required
Group Project	Varies	Varies	Depends on models

Free GPU: Google Colab provides free T4 GPU access (sufficient for all projects)

Alternative: pip Installation

# Create virtual environment
python -m venv dl-env
source dl-env/bin/activate  # Linux/Mac
# or
dl-env\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

# Run project
python main_script.py

Common Dependencies

# Core (all projects)
tensorflow==2.10.0
numpy==1.23.5
matplotlib==3.7.2

# Additional (specific projects)
tensorflow-probability==0.18.0  # VAE only
scikit-learn                     # CNN/RNN evaluation
seaborn                          # Enhanced visualizations

Author: Matheus Ferreira Teixeira
GitHub: github.com/domvito55
LinkedIn: linkedin.com/in/mathteixeira

Note: Each project folder contains its own detailed README with complete implementation details, architecture diagrams, training procedures, and results analysis.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README.md

Deep Learning Portfolio

📚 Table of Contents

🎯 Overview

🛠️ Skills & Technologies

Core Frameworks & Libraries

Advanced Techniques Implemented

Cumulative Skills Demonstrated

📊 Projects

1. CNN vs RNN Image Classification

2. Transfer Learning with Denoising Autoencoder

3. Variational Autoencoder (VAE)

4. Deep Convolutional GAN (DCGAN)

🔬 Key Technical Innovations

1. Custom Keras Layers for Probabilistic Modeling

2. Multi-Metric GAN Convergence Detection

3. β-Weighting Derivation for VAE Loss

🌍 Real-World Applications

Fashion & E-commerce

Computer Vision Systems

Research & Development

⚙️ Environment Setup

Recommended: Conda Environment

Hardware Requirements

Alternative: pip Installation

Common Dependencies

FilesExpand file tree

01_DeepLearning

Directory actions

More options

Directory actions

More options

Latest commit

History

01_DeepLearning

Folders and files

parent directory

README.md

Deep Learning Portfolio

📚 Table of Contents

🎯 Overview

🛠️ Skills & Technologies

Core Frameworks & Libraries

Advanced Techniques Implemented

Cumulative Skills Demonstrated

📊 Projects

1. CNN vs RNN Image Classification

2. Transfer Learning with Denoising Autoencoder

3. Variational Autoencoder (VAE)

4. Deep Convolutional GAN (DCGAN)

🔬 Key Technical Innovations

1. Custom Keras Layers for Probabilistic Modeling

2. Multi-Metric GAN Convergence Detection

3. β-Weighting Derivation for VAE Loss

🌍 Real-World Applications

Fashion & E-commerce

Computer Vision Systems

Research & Development

⚙️ Environment Setup

Recommended: Conda Environment

Hardware Requirements

Alternative: pip Installation

Common Dependencies