Deep Learning Specialization · Reimagined
Build what
you can explain.

Andrew Ng's five-course deep learning specialization rebuilt for community college learners. Six modules, three entry depths. You will not use a neural network until you have built one by hand.

3Tracks
6Modules
18Weeks
0Exams
Neural Network · Forward Pass
Input Hidden 1 Hidden 2 Output a⁽ˡ⁾ = σ(W⁽ˡ⁾a⁽ˡ⁻¹⁾ + b⁽ˡ⁾)
We derive this before we run it.

This vs. Andrew Ng's Original

Five courses compressed.
Depth expanded.

Andrew Ng's Deep Learning Specialization across five Coursera courses is the most thorough public DL education available. We condense it to one community-college-length course, add three entry tracks, and make implementation from scratch mandatory — not optional.

Andrew Ng's Version
5 Courses, TensorFlow-first
  • Five separate Coursera courses, 3–4 months each
  • Framework-first: TensorFlow and Keras introduced early
  • Single depth — same path for everyone
  • Graded coding assignments using pre-scaffolded notebooks
  • Limited treatment of AI bias in deep learning systems
  • No community application or deployment requirement
This Course
One Course, First Principles
  • Six focused modules, 18 weeks — matched to college schedule
  • Implement from scratch in NumPy before touching PyTorch
  • Three tracks from conceptual to from-scratch engineering
  • Portfolio: six projects you own, not auto-graded notebooks
  • Bias and harm analysis is a required module, not a footnote
  • Capstone must serve a real community problem you identify

Three Learning Tracks

One architecture.
Three depths.

All tracks attend every class, read every case study, and participate in every community discussion. The tracks differ in how far into the implementation you go — and you can move between them as your confidence builds.

Track I
I
Conceptual
DL
Prereq: High school algebra
Focus

Understand how deep learning systems work at an architectural and conceptual level. Interpret model outputs, identify failure modes, read about DL in the news without being lost. Work with pre-trained models. Prepare for roles that use or oversee DL systems without building them.

Projects
  • P1Use CLIP and DALL·E to probe what a neural network "knows"
  • P2Audit a deployed image classification system for failure modes
  • P3Write a plain-language technical report on a DL system for a public audience
Track II
II
Applied
DL
Prereq: Python · Basic calculus
Focus

Build real deep learning applications using PyTorch. Train, fine-tune, and deploy CNNs, sequence models, and transformers. Understand the key architectural decisions. By the end, you'll have built and deployed a working DL application solving a community problem you chose.

Projects
  • P1Fine-tune ResNet for a local species identification task
  • P2Build a sequence model for community text classification
  • P3Fine-tune a small transformer; document the training data's biases
  • P4Deploy a DL model with a harm analysis document
Track III
III
DL
Engineer
Prereq: Python · Calculus · Linear Algebra
Focus

Implement every architecture from scratch in NumPy: forward pass, backpropagation, Adam, batch normalization. Before you use PyTorch's autograd, you'll have written it. Before you call a Conv2d layer, you'll understand every multiplication inside it. Prepare for DL research.

Projects
  • P1Implement forward pass + backprop for a 3-layer MLP in NumPy only
  • P2Implement a convolutional layer from scratch; verify against PyTorch
  • P3Implement a single-head attention mechanism; visualize attention maps
  • P4Train a small GPT-like model on community text; analyze what it learned
  • P5Full reproducibility audit of a published deep learning paper

Six-Module Arc

Foundations to
frontier — step by step.

Each module builds on the last. You will not see attention until you deeply understand the MLP. You will not use PyTorch until you have implemented the core operations by hand. Depth before breadth.

1
Weeks 1–3
Neural Foundations

What is a neuron, really? The perceptron, sigmoid activation, and the universal approximation theorem. Implement a single neuron that learns to classify. Derive backpropagation from the chain rule — not from a textbook formula, but from first principles.

Project → Single-layer classifier in NumPy. Beat a decision tree on a community dataset.
  • Perceptron and threshold logic
  • Sigmoid, ReLU, tanh
  • Forward pass as matrix multiplication
  • Chain rule and backpropagation
  • Binary cross-entropy loss
2
Weeks 4–6
Deep Networks

Stack the layers. Understand why depth matters — and when it doesn't. Implement batch normalization and dropout from scratch. Investigate: what does each layer actually learn? Visualize activations and challenge the "black box" myth.

Project → 5-layer MLP from scratch. Visualize what each layer detects.
  • Vanishing/exploding gradients
  • Batch normalization
  • Dropout regularization
  • Adam optimizer
  • Activation visualization
3
Weeks 7–9
Convolutional Networks

Why does a filter slide? Implement convolution as a mathematical operation. Build ResNet-style skip connections. Ask: when we train an image classifier on biased data, what exactly does the convolutional filter learn? This isn't abstract — it's the coded gaze.

Project → CNN image classifier. Measure accuracy by demographic subgroup.
  • Convolution as dot product
  • Pooling layers
  • ResNet and skip connections
  • Transfer learning
  • Feature map visualization
4
Weeks 10–12
Sequence Models

Language has order. Implement RNNs and LSTMs. Understand why they struggle with long-range dependencies. When we train a language model on community text, what patterns does it amplify? Sequence models encode the statistics of whoever wrote the training data.

Project → LSTM sentiment classifier on local news. Analyze whose sentiment is captured.
  • RNN unrolling through time
  • LSTM gates
  • Word embeddings
  • Long-range dependency problem
  • Bias in language models
5
Weeks 13–15
Transformers & Attention

Attention is all you need — but what does attention mean? Implement single-head self-attention from scratch. Understand the architecture of GPT and BERT. Probe a small LLM to understand what it has and has not learned about your community.

Project → Implement attention from scratch. Fine-tune a small LLM on community text.
  • Query, key, value
  • Scaled dot-product attention
  • Positional encoding
  • GPT vs BERT architectures
  • Fine-tuning and RLHF
6
Weeks 16–18
Capstone: Build or Audit

Two paths. Build: design and deploy a deep learning system addressing a real community need you identified — with full documentation and a harm analysis. Audit: conduct a deep technical and ethical audit of an existing deployed DL system, including a reproducibility study.

Project → Community DL application or deep audit report. Public exhibition.
  • Deployment and MLOps basics
  • Model cards and datasheets
  • Reproducibility standards
  • Community consent
  • Public exhibition
"You will not call model.fit() until you have written fit() yourself. Understanding precedes convenience — always."
Course Philosophy · Ko, Ng, Buolamwini, Anderson
⚙️
Implement Before ImportingEvery architecture is implemented from scratch before any library version is used. You will know what conv2d does because you wrote it.
🔭
Interpret Every LayerYou will visualize what every layer in every network has learned. Black boxes are not acceptable. Explainability is engineering, not ethics theater.
⚖️
Audit Before DeploymentEvery model you build includes a mandatory bias and harm analysis. Model cards are not optional. Harm documentation is part of the grade.
🏘️
Real StakesEvery dataset comes from real communities. The capstone must address a real problem. You are building systems that will touch real lives.

Intellectual Lineage

Who we're learning from.

Andrew Ng
Deep Learning Spec

The five-course specialization whose conceptual depth and visual clarity we build on and condense. Module 5 directly references his attention mechanism explanations.

Andrej Karpathy
micrograd · makemore

The best from-scratch implementation tutorials in existence. Our Track III backpropagation project is inspired by his micrograd series.

Joy Buolamwini
Algorithmic Justice League

The coded gaze. Module 3's bias analysis project is built around her Gender Shades methodology applied to locally-relevant datasets.

Timnit Gebru
DAIR Institute

Datasheets for datasets. Stochastic parrots. Critical analysis of scale in language models. Core reading for Module 5.

Yoshua Bengio
Deep Learning Text

The canonical Deep Learning textbook (free online). Mathematical depth for Track III. The source for RNN and attention derivations.

Amy J. Ko
UW CS Education

Read before write. Student agency. The pedagogical framework behind the three-track structure and portfolio assessment system.

Emily Bender
Stochastic Parrots

On the dangers of stochastic parrots. What language models do and do not understand. Essential for Module 5's critical framing.

Jeff Anderson
Ungrading

Portfolio over exams. Deep learning over performance. The two-minute rule. Assessment philosophy that runs throughout.