Master the full scope of KOAI Subject 2 (Neural Networks and Deep Learning) through both theory and hands-on code. Students implement everything from MLPs to mini transformer models in PyTorch and reach a level where they can choose optimizer, regularization, and fine-tuning techniques to fit the data. Following F1, this is a required course that every KOAI applicant goes through.
Published: May 16, 2026 | Last updated: May 16, 2026 · Based on the KOAI 2026 guidelines
Track
Foundation
Required for all KOAI applicants
Target
F1 graduates
Grade 9 to Grade 11
Recommended Hours
~14 hours
1:1 basis · varies 10-20 hours
KOAI Mapping
Full scope of Subject 2
Syllabus 2-1 to 2-2
By the time they finish F2, students will master the full scope of KOAI Subject 2 (Neural Networks and Deep Learning) through both theory and practice. Starting from the mathematical derivation of perceptrons and gradient descent, the goal is to implement everything from MLPs to mini transformer models in PyTorch.
Rather than just running models, students reach a level where they can choose optimizer, regularization, and fine-tuning techniques to fit the data. In weeks 1-2 they implement perceptrons and backpropagation by hand in NumPy to understand the principles first, then move into PyTorch to cover attention, transformers, ViT, autoencoders, and LoRA fine-tuning. It's the common required course for every KOAI applicantIt is.
F2 is for F1 graduates students who have completed F1. If you can demonstrate equivalent competency from elsewhere, you may enroll by exception. If you haven't taken F1 yet, please complete F1. Foundations I first. F2 builds neural networks and deep learning on top of the NumPy, Pandas, and classical ML foundation laid in F1.
Below is the standard plan on a 1:1 basis. Depending on the student's prior knowledge and pace, some weeks may be accelerated, compressed, or covered in greater depth. Key tools: NumPy, PyTorch, torchvision, Matplotlib, Hugging Face, PEFT/LoRA.
| Week | Topic | Key Deliverable |
|---|---|---|
| 1 | Perceptrons, gradient descent (mathematical derivation) | Implement a single-layer model by hand in NumPy |
| 2 | Backpropagation, activation functions (ReLU, Sigmoid, Tanh) | Implement backprop by hand in NumPy |
| 3 | Loss functions (MSE, MAE, Cross Entropy) + intro to PyTorch | First PyTorch MLP |
| 4 | MLPs, GPU training, tensor manipulation | MNIST / Fashion-MNIST classification |
| 5 | SGD, mini-batch, momentum, Adam/AdamW | Optimizer comparison experiment |
| 6 | Learning rate scheduling, convergence diagnostics | LR finder + lr schedule comparison |
| 7 | Regularization (Dropout, Early Stopping, Weight Decay) | Overfitting control experiment |
| 8 | Weight initialization, batch normalization | BN on/off ablation |
| 9 | Data embeddings (text/image/audio) + pooling | Embedding visualization |
| 10 | Attention mechanism (theory) | Implement scaled dot-product attention by hand |
| 11 | Transformer text encoder (theory + practice) | Train a mini Transformer |
| 12 | Transformers for vision (ViT theory) | Inference with a pretrained ViT model |
| 13 | Autoencoders + model fine-tuning (full vs PEFT/LoRA) | LoRA fine-tuning of a pretrained model |
| 14 | Capstone: a DL pipeline on your own dataset | End-to-end repo + report |
※ Weeks are content units; actual time required varies by student. Recommended ~14 hours, ranging 10-20 hours.
Each week students write a PyTorch notebook and an analysis of the results, in both Korean and English. This leverages international-school students' strength in English while simultaneously preparing for the KOAI Round 2 Korean written response exam.
Students build an end-to-end deep learning pipeline for a problem they define themselves (choosing image or text). The problem definition, model design, experiments, and report all become portfolio assets.
F2 leaves behind 1 GitHub repo koai-deep-learning as a cumulative asset, and the capstone is written up as one case study in Notion. This case study can be used directly as material for the 200-300 character Prompt 2 of the KOAI personal statement.
GitHub
1 organized repo koai-deep-learning
Notion
1 case study (capstone)
Personal statement
Prompt 2 material (200-300 characters)
This track record accumulates as dated evidence in the KOAI Round 1 application's Portfolio 40% · AI competency 30% items as dated evidence. The earlier you start, the more depth you'll have by exam time.
F2 is the second course in the Foundations track of the KOAI curriculum. For the full track structure, see the KOAI Prep Curriculum Hub.
Current Course
F2. Foundations II
Neural Networks & Deep Learning
Yes. F2 is for students who have completed F1. If you can demonstrate equivalent competency from elsewhere, enrollment is possible by exception.
The full scope of Syllabus Subject 2 (Neural Networks and Deep Learning), units 2-1 through 2-2. From perceptrons and backpropagation to transformers and LoRA fine-tuning.
Yes. Week 3 covers an introduction to PyTorch, and in the preceding weeks 1-2 students implement perceptrons and backpropagation by hand in NumPy to understand the principles first.
It's a recommended figure on a 1:1 basis and varies from 10 to 20 hours per student. We estimate an individual time plan in the first diagnostic session.
High-school division applicants go deeper with A1 (CV) and A2 (NLP & Audio) and prep for the competition with C1, C2, and C3. Middle-school division applicants cover F1 and F2 in compressed form in the M1 middle-school comprehensive program. For the exact schedule, see the KOAI Competition Guide (https://citcoding.com/competitions/koai.html).
We assess your level of F1 completion and design an individualized F2 start point and time plan.