Advanced Technology
Deep Learning (DL)
Work with neural networks for images, language, and complex pattern recognition.
Prerequisites
Basic understanding of Machine Learning is recommended before starting Deep Learning.
Certificate
Awarded by Rescue Academy on successful completion of the program's assessments and final project.
Learn Online — Live Classes
Register, receive your schedule, and join live instructor-led Deep Learning (DL) classes on Zoom. Assignments, instructor feedback, and a certificate on completion.
Learn In Person
Attend Deep Learning (DL) classes in person in Juba with hands-on labs, instructor mentorship, and a certificate on completion.
What You'll Learn
- Neural network foundations and training with PyTorch
- Convolutional networks, transfer learning and modern architectures
- Sequence models, attention and transformers
- Building image and text data sets, efficient GPU training and evaluation
- Generative models and their risks
- Detection, segmentation and deployment to phones
- Responsible deep learning: bias, privacy, cost and documentation
Program Objectives
- Build and train neural networks for images, text and sequences with PyTorch
- Use transfer learning and transformers to get strong results from small local data sets
- Evaluate deep learning models rigorously and diagnose their failures
- Compress and deploy models that run offline on modest devices
- Develop deep learning systems responsibly and document them
Who It's For
- Learners who have completed Machine Learning Beginner and want to work with images, text or audio
- Developers and researchers building AI features for field or mobile use
- Graduates preparing for AI engineering roles
Where You'll Use It
- Agriculture (crop disease and pest recognition)
- Health and humanitarian work (image and text classification)
- Language technology and mobile apps that work offline
Curriculum
Instructor-led, practical training. Each level is a block of 4, 8 or 12 weeks; every module ends with a practical activity and an assessment, and every level with a project.
Module 1: Neural Network Foundations 10 h
Understand artificial neurons, layers and activations, how a network makes a prediction and how it learns by gradient descent. 3 lessons
Module 2: Convolutional Networks for Images 11 h
Represent images as numbers, understand convolutions and pooling, and build and train a convolutional image classifier. 3 lessons
Module 3: Sequence Models for Text and Time 10 h
Represent sequences such as text and sensor readings, and understand recurrent networks and the move to attention. 3 lessons
Module 4: Transfer Learning with Pre-trained Models 11 h
Reuse models trained on large data sets to get strong results with small local data sets. 3 lessons
Module 5: Training Neural Networks Well 10 h
Choose losses, optimisers and learning rates, and prevent overfitting with regularisation, augmentation and early stopping. 3 lessons
Level project: Leaf Disease Classifier with Transfer Learning
Module 1: PyTorch in Depth 15 h
Write well-structured PyTorch code: tensors and autograd, custom datasets and loaders, and reusable training loops with checkpoints. 3 lessons
Module 2: Modern CNN Architectures 16 h
Understand the architectures behind today's image models — deeper networks, batch normalisation, residual connections and efficient mobile models. 3 lessons
Module 3: Building Image and Text Data Sets 16 h
Collect and label data well, augment it sensibly and prepare text for deep learning models. 3 lessons
Module 4: GPUs and Efficient Training 13 h
Use GPUs effectively, speed up training and manage compute time, cost and experiments. 3 lessons
Module 5: Evaluating Deep Learning Models 13 h
Evaluate deep models rigorously, analyse their errors and diagnose underfitting and overfitting. 3 lessons
Module 6: Intermediate Deep Learning Project 17 h
Deliver the medicine package recognition project from data to a recommended, mobile-ready model. 3 lessons
Level project: Medicine Package Recognition for Pharmacy Stock Checks
Module 1: Attention and Transformers 24 h
Understand attention and the transformer architecture, and fine-tune pre-trained transformers for language and vision tasks. 4 lessons
Module 2: Generative Models 20 h
Understand how autoencoders, GANs, diffusion models and large language models generate content, and assess their risks. 4 lessons
Module 3: Computer Vision Beyond Classification 22 h
Locate and outline objects with detection and segmentation models, evaluate them correctly and apply them to real tasks. 4 lessons
Module 4: Deploying Deep Learning Models 22 h
Export, compress and serve deep learning models, including offline on phones and in low-connectivity settings. 4 lessons
Module 5: Responsible Deep Learning 22 h
Address bias, privacy, consent, compute cost and misuse in deep learning, and document systems responsibly. 4 lessons
Module 6: Deep Learning Capstone 30 h
Design, train, compress, evaluate and document an offline field assistant model for a real-world task. 5 lessons
Level project: Offline Field Assistant Model
Final practical project
Offline Field Assistant Model
A deep learning model for a field task that runs offline on a mid-range phone, evaluated on field-like data and documented with a model card and risk assessment.
Deliverables
- A reproducible training repository and trained model
- A compressed, exported model with on-device measurements
- A field-like evaluation with breakdowns by condition
- A model card, risk assessment and live demonstration
Assessed on
- Accuracy, size and latency targets are met on the target device
- Evaluation reflects field conditions and reports failures honestly
- Data use respects consent, privacy and licences
- Another team could retrain and deploy the model from the documentation
Worked Examples
A selection of the visual, practical examples used in this program's lessons — each teaches a different skill.
- Worked calculation One neuron, calculated by hand Inputs, weights and a bias combine into a weighted sum, then ReLU and sigmoid turn it into an output. Beginner · Module 1 · Lesson 1
- Worked calculation How a convolution filter finds an edge The same 3 × 3 vertical-edge filter applied to three image patches gives very different responses. Beginner · Module 2 · Lesson 1
- Step-by-step Transfer learning for a leaf disease classifier The sequence that turns a model pre-trained on millions of photos into a classifier for local crop leaves. Beginner · Module 4 · Lesson 2
- Data and chart Three learning rates, three very different training runs Training loss per epoch shows learning rates that are too small, about right and far too large. Beginner · Module 5 · Lesson 2
- Troubleshooting tree My deep learning model is disappointing — what now? Use training, validation and field results to find out whether the model, the training or the data is the problem. Intermediate · Module 5 · Lesson 3
- Case study An offline pest-counting app for extension workers A composite, illustrative case of designing a vision model around field realities rather than benchmark scores. Advanced · Module 3 · Lesson 4
Free Course: Deep Learning Starter
A free conceptual tour of deep learning: neural networks, convolutional networks for images, recurrent networks for sequences, transfer learning and how training is tuned. Free, self-study, about 6 hours — no certificate. The full program above is taught live by an instructor, with practical assignments, assessments and a certificate.
- 1Free lesson 1Introduction to Artificial Neural Networks (ANNs)
- 2Free lesson 2Convolutional Neural Networks (CNNs) for Computer Vision
- 3Free lesson 3Recurrent Neural Networks (RNNs) for Sequence Data
- 4Free lesson 4Transfer Learning & Pre-trained Models
- 5Free lesson 5Model Optimization: Loss Functions, Optimizers & Learning Rate
- Final checkQuick Quiz — Deep Learning Basics
Tools & Technologies
Primary Tools
- Python with Jupyter or Google Colab — The working environment, with free GPU sessions where available Practice: Running and documenting experiments
- PyTorch and torchvision — The deep learning framework used throughout Practice: Models, datasets, training loops and pre-trained models
Supporting Tools
- Hugging Face Transformers — Pre-trained transformer models and tokenisers at Advanced level Practice: Fine-tuning language and vision transformers
- Labelling tools — Annotating images and text Practice: Building quality data sets
- Git — Versioning training code Practice: Reproducible experiments
Professional / Industry Tools
- ONNX and ONNX Runtime — Exporting and running models outside the training environment Practice: Portable, efficient deployment
- Mobile runtimes (ONNX Runtime Mobile, TensorFlow Lite or PyTorch mobile options) — Running models offline on phones Practice: On-device inference and benchmarking
- Experiment tracking (e.g. MLflow) — Recording training runs Practice: Comparing and reproducing experiments
Career Opportunities
- Data/AI analyst role at a bank, telecom, or NGO
- Research or further-study pathway in AI/ML/Data Science
- Remote/freelance AI or data work
Practical Projects
- Train a CNN-based image classifier on a real dataset
- Fine-tune a pre-trained model for a new task using transfer learning
- Given a training run with unstable or diverging loss, diagnose and fix the cause (learning rate, normalization, or batch size) and document what actually fixed it
Problem-Solving Focus
Problem: Training a deep model from scratch would need more data and compute than is realistically available
Approach: Use transfer learning — start from a pre-trained model and fine-tune it on a smaller local dataset, instead of training an entire network from zero.
Problem: A CNN trained on clean studio product photos performs poorly on real-world photos taken on a phone in different lighting
Approach: Use data augmentation during training — rotations, brightness/contrast changes, cropping — so the model generalizes to real, imperfect input instead of memorizing the training conditions.
Problem: A training run keeps failing or diverging — the loss goes to NaN or simply never improves
Approach: Diagnose it systematically: check the learning rate first, then data normalization and batch size, rather than randomly changing the model architecture and hoping something works.
Problem: A model is accurate but too slow or resource-heavy to run on a normal device or an affordable cloud plan
Approach: Treat model size and deployment cost as part of "done," not an afterthought — consider a smaller architecture or quantization instead of assuming more compute is the only fix.
Ready to register for Deep Learning (DL)?
Register — Learn Online Register — Learn In Person WhatsApp: +211926196668 Email: rescueacademy26@gmail.com