Course 03

Design and Train Neural Networks

This course enters the 'engine room' of deep learning. Move from abstract concepts to the concrete mechanics of training neural networks, including supervised learning, classification, sentiment analysis, and Named Entity Recognition.

Duration: 4 hours 12 Datasets

Key Topics

  • Multilayer Perceptrons (MLPs): The building blocks of deep learning
  • The Training Loop: Forward pass, loss calculation, backpropagation
  • Optimization: Gradient descent algorithms
  • Diagnostics: Identifying overfitting and underfitting

Learning Outcomes

  • Understand the mechanics of training a neural network
  • Learn to spot and mitigate common training issues
  • Evaluate model performance on classification tasks
  • Implement MLPs from scratch or using frameworks
  • Apply backpropagation theory to code

Practical Projects

  • Implement and evaluate an MLP for classification tasks
  • Social Impact Analysis: Anticipate potential risks of trained models
  • Train a Bi-LSTM-CRF model for NER in African languages

Recommended Datasets

12 datasets are aligned with this course for hands-on exercises and projects.

Malaria Cell Images

Parasitized vs Uninfected Blood Cell Images (NIH)

image-classification computer-vision 27,558 images (~339MB)
English (Metadata)

MasakhaNEWS

News Topic Classification for 16 African Languages

topic-classification language-modeling ~20,000 articles
Amharic Hausa Igbo Luganda +5 more

KaraAgro AI Maize

Annotated Maize Disease Images from Ghana

object-detection image-classification 16,552 images (~1.5GB)
English (Metadata)

MasakhaNER 2.0

Named Entity Recognition for 20 African Languages

named-entity-recognition sequence-labeling ~9,000-10,000 sentences per language
Bambara Ewe Fon Hausa +12 more

AfriSenti

Sentiment Analysis for 14 African Languages

sentiment-analysis text-classification 110,000+ tweets
Amharic Hausa Igbo Kinyarwanda +5 more

AfriHate

Hate Speech Detection for 15 African Languages

hate-speech-detection text-classification ~45,000 examples
Algerian Arabic Amharic Hausa Igbo +8 more

Tunizi

Tunisian Arabizi Sentiment Analysis

sentiment-analysis text-classification ~3,000 sentences
Tunisian Arabizi

South Africa Crop Type

Fused Sentinel-1/2 Imagery for Crop Mapping

semantic-segmentation data-fusion ~10GB
English (Metadata)

Ghana Croplands Agronomy

Biophysical & Yield Measurements for Northern Ghana Maize

regression yield-prediction <100MB
English (Metadata)

African Medical Records (OCR)

Nigerian Handwritten Clinical Records for OCR

ocr computer-vision <1K records
English (Nigeria)

Nigerian Colposcopy Images

Cervical Cancer Screening Images from Nigeria

image-classification computer-aided-diagnosis 3,356 images
English (Metadata)

Snapshot Serengeti

Camera-Trap Wildlife Images from Tanzania

image-classification object-detection ~5TB full (use a subset)
English (Metadata)

About This Course

This course covers the fundamentals of neural networks for NLP. You’ll learn to design, train, and optimize neural networks for text classification, sequence labeling, and other language tasks.

Key Topics

  • Network Architecture: Feedforward, recurrent, and convolutional networks
  • Backpropagation: Understanding gradient flow
  • Optimization: SGD, Adam, and learning rate scheduling
  • Regularization: Dropout, weight decay, and early stopping
  • Practical Training: Debugging and improving model performance

Prerequisites

  • Completion of Courses 01-02 or equivalent knowledge
  • Calculus basics (derivatives, chain rule)
  • Python with PyTorch or TensorFlow