DS 6050: Deep Learning
A 12-module graduate training camp for building deep-learning systems from first principles, diagnosing why they fail, and making defensible engineering and research decisions.
Teaching assistant
Fokhrul Islam
UVA ID: sdh4py
Delivery
Weekly live lectures
Weekly lectures and readings on Canvas
Course purpose
Move beyond surface-level recipes
Grand challenges in medical imaging, language, and scientific discovery require more than off-the-shelf code. They require quantitative understanding, disciplined implementation practice, and the judgment to evaluate trade-offs honestly.
By the end of the course, you should be able to reason from mathematical principles, build core components from scratch, debug model behavior systematically, and adapt those learning strategies to new architectures and research after graduation.
Three learning phases
Build the mathematical foundation, then implement linear models, multilayer perceptrons, and backpropagation from scratch with NumPy. Optimization and ablation methodology become recurring tools rather than one-off topics.
Connect data modality to architecture: CNNs for spatial data, RNNs for sequences, and attention and Transformers for general-purpose modeling. Diagnose the training challenges specific to each family.
Study large-scale pretraining, vision transformers, efficient adaptation, and generative modeling while practicing paper reproduction, systematic ablations, and research judgment.
Course calendar
One module per week across the 12-week sequence.
- Week 1Module 1From Linear Regression to Neural Networks
- Week 2Module 2Backpropagation
- Week 3Module 3Training MLPs
- Week 4Module 4Convolutional Neural Networks
- Week 5Module 5Advanced CNN Architectures
- Week 6Module 6Encoder Decoder Architectures
- Week 7Module 7Recurrent Neural Networks
- Week 8Module 8Attention Mechanism
- Week 9Module 9Self-Attention and Transformers
- Week 10Module 10Transformer Models
- Week 11Module 11Prompting, PEFT & Quantization
- Week 12Module 12Generative Modeling
Assessment as learning
Assessment is designed first as guided practice: unit tests, iteration, reflection, feedback, and optional peer code review make evaluation a by-product of learning.
Programming assignments
Homework 1
Foundations & backpropagation
Linear models, gradients, and early autograd practice. The supporting Colab is on Module 1.
Homework 2
Optimization & CNNs
Training stability, convolutional stacks, and transfer learning. The supporting Colab is on Module 5.
Homework 3
Sequence-to-sequence learning
Implement and study a recurrent sequence-to-sequence model in the provided Colab.
Homework 4
Attention
Add cross-attention to a GRU-based sequence-to-sequence model.
Homework 5
Transformer 2.0
Integrate the course's Transformer concepts in the final programming assignment.
One automatic drop, equal weight
Every homework is graded out of 100 points, so all five carry exactly the same weight. Your homework grade is the average of your four highest scores out of five: one homework is dropped for everyone, automatically.
The drop is the skip
A skipped homework is scored 0 and becomes your drop. If you submit all five, your weakest score is dropped. Everyone is graded on four assignments; there is no separate skip plus drop.
Nothing is submitted late
Work not submitted by the deadline is scored 0 and becomes your drop. You do not need to notify us, apologize, or make a request.
Documented exceptions
A second missed homework requires instructor or TA approval before the deadline. If a sudden emergency prevents advance notice, contact us as soon as you are able, with documentation. An approved absence is removed from the calculation, and your automatic drop still applies.
Same benefit for everyone
All five count equally. No homework is worth more than another, and there is nothing to gain by choosing which one to skip.
One skipped
92, 88, 0, 95, 90 → 91.25
All submitted
92, 88, 71, 95, 90 → 91.25
In both examples, the lowest score is removed and the remaining four scores are averaged.
Short 10–15 question, open-book checks focus on understanding rather than memorization. Full credit is based on thoughtful completion, with immediate explanatory feedback.
Quizzes are delivered in Canvas; quiz questions are not published on the public module pages.
Teams apply course concepts to a real problem through three feedback-centered milestones.
- 1Proposal & literature review5%
- 2Mid-project checkpoint10%
- 3Final deliverables25%
Grading
Homework deadlines, the automatic drop, and excused absences follow the homework policy above.
40%
Programming assignments
Five assignments, each out of 100; grade is the average of your four highest scores (one automatic drop)
40%
Group project
Proposal, checkpoint, and final deliverables
20%
Participation
Canvas quizzes and course engagement
Professional and academic integrity
All work must comply with the UVA Honor System. Individual assignments must be original work; conceptual discussion is encouraged, and external code, ideas, and resources must be cited. Include the honor pledge with relevant submissions.