Picsart Academy

From Regression to Deep Learning

• Why Regression Isn't Enough
• Single Neuron & Bias
• ReLU & Sigmoid
• Forward Pass
• Decision Regions

How Neural Networks Learn

• Loss Functions
• Backpropagation
• SGD vs Adam
• Mini-Batch Training
• MNIST Hands-on

Convolutional Neural Networks

• Convolutions
• Receptive Fields
• Pooling
• Dropout & BatchNorm
• LeNet Architecture

Data Augmentation & Architectures

• Augmentation Pipeline
• LeNet → AlexNet → VGG
• ResNet
• Transfer Learning
• Cosine Similarity

Variational Autoencoders & Latent PCA

• Encoder / Decoder
• Reparameterization
• β-ELBO Loss
• KL Divergence
• Latent PCA Controls
• Face Generation

From Words to Embeddings

• Tokenization
• Vocabularies
• nn.Embedding
• Word2Vec
• Sentiment Classifier

Transformers & Attention

• RNNs & LSTMs
• Self-Attention
• Multi-Head Attention
• BERT & GPT
• Hugging Face

From Vision–Language Models to CLIP

• CNN + RNN Captioning
• Visual Attention
• CLIP Architecture
• Contrastive Learning
• Zero-Shot Classification
• Text-to-Image

Regression to Deep Learning Demo

Generate a spiral dataset, train a small PyTorch network with ReLU/Sigmoid, and visualize decision regions.

• Spiral Dataset
• PyTorch MLP
• Decision Boundaries

MNIST with Fully-Connected Network

Classify MNIST digits using an MLP. Achieves ~98% accuracy with live loss visualization and confusion matrix.

• MNIST
• MLP Classifier
• Confusion Matrix

MNIST with CNN

Train a CNN on MNIST for ~99% accuracy. Visualize learned convolution filters and activation maps.

• CNN Architecture
• Filter Visualization
• Activation Maps

Data Augmentation (CIFAR-10)

Compare training with and without augmentation using FastAI and xResNet18 on CIFAR-10.

• FastAI
• Augmentation Pipeline
• Transfer Learning

Word Embeddings

From tokenization to Word2Vec magic. Explore word analogies, Skip-Gram, and CBOW.

• Tokenization
• nn.Embedding
• Word2Vec Analogies

Sentence Classification (AWD-LSTM)

Classify AG News topics using a pretrained AWD-LSTM with FastAI's text module.

• AG News
• AWD-LSTM
• Fine-tuning

Bag of Embeddings Classifier

A simpler alternative: average pretrained GloVe embeddings and classify with a linear layer.

• GloVe Embeddings
• Bag of Embeddings
• Topic Classification

Transformers & Attention

Self-attention from scratch, positional encoding, multi-head attention. Use BERT & GPT-2 via Hugging Face.

• Self-Attention
• BERT & GPT-2
• Hugging Face

CLIP Basics

Zero-shot image classification, prompt engineering, image-text retrieval, and contrastive learning with CLIP.

• Zero-Shot Classification
• Prompt Engineering
• Image-Text Retrieval
• InfoNCE Loss

Vision-Language Pipelines

Image captioning and VQA with BLIP, zero-shot classification with HuggingFace pipelines.

• BLIP Captioning
• Visual Question Answering
• HuggingFace Pipelines

Caption Generator with CLIP

Train an LSTM caption decoder using frozen CLIP embeddings. Learn the "Show and Tell" architecture.

• CLIP Encoder
• LSTM Decoder
• Greedy Decoding
• Caption Training

Face Autoencoder

Train a Variational Autoencoder on LFW faces, analyze the latent space with PCA, and explore an interactive Gradio app that lets you manipulate facial features with sliders.

• VAE
• LFW Dataset
• Latent PCA
• Gradio App
• Face Generation