# CS1470 Deep Learning: Workshop & SRC Makeup Assignment

**Due Date:** Sunday, December 14, 2025 via Gradescope

## Overview

This makeup assignment is designed for students who were unable to attend all required Workshop and SRC discussion sessions. The assignment captures the essence of both components: **technical exploration** of emerging deep learning topics and **critical reflection** on the societal, ethical, and practical implications of these technologies.

The assignment has three tiers based on the number of sessions missed. Each tier builds incrementally on the previous one, requiring progressively deeper engagement with both technical concepts and their broader implications.

## Assignment Tiers

### Tier 1: Missed ONE session (1 Workshop OR 1 SRC)
**Page Limit:** 2 pages, double-spaced

### Tier 2: Missed TWO sessions
**Page Limit:** 3-4 pages, double-spaced

### Tier 3: Missed THREE sessions
**Page Limit:** 4-5 pages, double-spaced

## Topic Selection Requirements

**Critical Requirement:** You MUST choose a deep learning topic, technique, or application that was **NOT directly covered** in CS1470 lectures or assignments.

### Prohibited Topics (Too Broad/Already Covered):
- Basic CNNs, ResNets, or standard architectures
- Standard RNNs, LSTMs, or basic Transformers
- Vanilla autoencoders, VAEs, GANs, or diffusion models (as covered in class)
- Basic RL algorithms covered in lecture (DQN, Policy Gradients, PPO)

### Encouraged Topics (Niche and Exploratory):
- **Architecture Innovations:** Vision Transformers (ViT), Swin Transformers, Sparse Transformers, State Space Models (Mamba), Mixture of Experts (MoE), Joint Embedding Predictive Architectures (JEPAs)
- **Specialized Applications:** Neural Radiance Fields (NeRF), protein structure prediction (AlphaFold), neural drug discovery, climate modeling with DL
- **Training Techniques:** Federated learning, continual/lifelong learning, meta-learning (MAML), neural architecture search (NAS), knowledge distillation, Distribtued Training techniques (FSDP, Model Parallelism) (you can reach out to me for more information on this if you want)
- **Emerging Paradigms:** Neuromorphic computing, quantum machine learning, physics-informed neural networks (PINNs)
- **Specialized Domains:** Graph neural networks for molecules, deep learning for genomics, audio synthesis (WaveNet, Jukebox), 3D generation
- **Safety & Alignment:** Constitutional AI, RLHF implementation details, mechanistic interpretability, adversarial robustness

**If you're unsure whether your topic qualifies, ask on EdStem or email Armaan.**

---

## Tier 1 Requirements (2 pages)

Complete ALL of the following components:

### Part A: Technical Exploration (1-1.5 pages)

Select a niche deep learning topic and provide a **detailed technical explanation** that demonstrates understanding beyond surface-level description.

**Your explanation must include:**

1. **Problem Formulation**
- What specific problem does this technique address?
- Why do standard approaches (covered in class) fall short?

2. **Technical Mechanism**
- How does the technique work? Include mathematical formulations where relevant
- What makes this approach unique or innovative?
- Explain at least one key algorithmic component or architectural choice

3. **Empirical Results**
- What are the reported performance gains or capabilities?
- On what benchmarks or applications has this been tested?
- Include at least one concrete example or case study

**Requirements:**
- Cite at least **2 primary sources** (research papers, not blog posts)
- Use diagrams, equations, or pseudocode where helpful (doesn't count toward page limit if clearly labeled)
- Write for an audience that is the TAs (we are familiar with basic DL concepts)

### Part B: Critical Reflection (0.5-1 page)

Analyze **one significant ethical, societal, or practical concern** related to your chosen topic.

**Address the following:**
- What is a specific risk, limitation, or ethical consideration?
- Who could be impacted (positively or negatively)?
- What makes this concern particularly relevant to THIS technique?

**Examples of strong reflections:**
- Environmental impact of training massive models
- Bias amplification in specialized domain applications
- Accessibility barriers to deploying cutting-edge techniques
- Privacy concerns in federated learning
- Dual-use concerns (beneficial vs. harmful applications)
- Interpretability challenges in safety-critical domains

---

## Tier 2 Requirements (3-4 pages)

Complete **ALL Tier 1 requirements** PLUS the following:

### Part C: Comparative Analysis (0.5-1 page)

Compare your chosen technique with a **related approach covered in CS1470**.

**Your analysis must:**
- Identify a relevant baseline or alternative from the course (e.g., standard Transformers, CNNs, VAEs, etc.)
- Explain the key technical differences
- Discuss trade-offs: When would you use one vs. the other?
- Consider computational cost, data requirements, performance, interpretability

### Part D: Limitations & Failure Modes (0.5-1 page)

Conduct a **deep dive into where this technique struggles**.

**Explicitly address:**
1. **Technical Limitations**
- What types of problems or data does it handle poorly?
- Are there known failure modes or edge cases?
- What are the computational or scaling challenges?

2. **Real-World Case Study**
- Find ONE example where this technique was deployed or tested in practice
- What challenges emerged during deployment?
- How did performance differ from controlled benchmarks?

### Part E: Extended Ethical Analysis (0.5-1 page)

Expand your Tier 1 reflection by analyzing **multiple stakeholder perspectives**.

**Consider:**
- Developers/researchers
- End users
- Affected communities
- Regulatory bodies

**Questions to explore:**
- Who benefits most from this technology? Who might be harmed?
- What safeguards or governance mechanisms might be needed?
- Are there existing frameworks (fairness metrics, audit procedures) that apply?

---

## Tier 3 Requirements (4-5 pages)

Complete **ALL Tier 1 & Tier 2 requirements** PLUS the following:

### Part F: Cross-Domain Synthesis (0.5-1 page)

Identify and analyze **connections between your technical topic and a broader deep learning challenge**.

**Examples of synthesis:**
- How does your technique relate to the interpretability vs. performance trade-off?
- How might this approach intersect with data efficiency or few-shot learning?
- Could this technique help address fairness or bias concerns in other applications?
- How does this fit into the broader landscape of scaling laws, emergent capabilities, or alignment research?

**Your synthesis should:**
- Draw explicit connections to concepts from multiple course units
- Identify unexpected relationships or implications
- Demonstrate systems-level thinking about deep learning

### Part G: Critical Position & Proposal (1-1.5 pages)

Develop a **well-reasoned position** on the responsible development and deployment of your chosen technique.

**Your position should:**

1. **Take a Clear Stance**
- Should this technology be developed further? Under what conditions?
- What guardrails, if any, should be in place?
- Are there applications that should be prioritized or avoided?

2. **Provide Justification**
- Support your position with technical reasoning AND ethical considerations
- Acknowledge counterarguments and address them
- Reference specific examples or precedents

3. **Propose Concrete Actions**
- What should researchers, companies, or policymakers do?
- Suggest at least 2-3 specific, actionable recommendations
- Consider short-term and long-term implications

### Part H: Future Directions & Open Problems (0.5 page)

Identify and discuss **1-2 unresolved challenges or promising research directions**.

**For each direction, you may try to answer:**
- What is the open problem?
- Why is it important?
- What makes it technically difficult?
- What progress has been made (if any)?

---

## Formatting & Submission Guidelines

### Document Format
- **Font:** 12pt Times New Roman or similar serif font
- **Spacing:** Double-spaced
- **Margins:** 1-inch margins on all sides
- **Headers:** Include your name, Banner ID, and tier number
- **Citations:** Use any consistent citation style (APA, IEEE, etc.)
- **File Format:** Submit as PDF via Gradescope

### Page Limits
- **Tier 1:** Exactly 2 pages (±0.25 pages)
- **Tier 2:** 3-4 pages
- **Tier 3:** 4-5 pages

**Note:** Citations, figures, and diagrams do NOT count toward the page limit if they are placed in a clearly labeled "References" or "Figures" section. However, in-text citations and small inline equations DO count.

### Academic Integrity
- All sources must be properly cited
- Paraphrase in your own words; do not copy-paste from sources
- Use of AI tools (ChatGPT, Claude, etc.) is permitted for brainstorming or editing, but final work must demonstrate your understanding

---

## Grading Rubric (Pass/Fail)

The rubric for this assignment is given to you. In order to **PASS**, your submission must meet ALL of the following criteria:

### Content Requirements
- Chosen topic is sufficiently niche and not directly covered in CS1470
- All required components for your tier are present and substantive
- Technical explanations demonstrate genuine understanding (not just summarizing abstracts)
- Critical reflections show thoughtful analysis, not generic statements
- Minimum citation requirements met with proper attribution

### Quality Standards
- Writing is clear, organized, and free of major errors
- Claims are supported with evidence (citations, examples, data)
- Analysis goes beyond surface-level description
- Appropriate technical depth for someone who has completed CS1470
- Meets page length requirements

### Standards for Not Passing
- Topic is too broad or was covered extensively in lecture
- Superficial treatment (e.g., just summarizing Wikipedia or a blog post)
- Missing required components
- Ethical reflection is generic or disconnected from the specific technique
- Poor writing quality that obscures understanding
- Insufficient citations or evidence of plagiarism
- Significantly under/over page limit

---

## Tips for Success

### Finding Good Sources
- Start with recent conference papers (NeurIPS, ICML, ICLR, CVPR)
- Check Google Scholar, arXiv, Papers with Code
- Look for survey papers for broader context
- Don't rely solely on blog posts or medium articles (use primary sources!)

### Choosing Your Topic
- Pick something that genuinely interests you
- Make sure there's enough literature to understand it deeply
- Avoid topics that are too new (might lack critical analysis)
- Avoid topics that are too obscure (might lack context)

### Writing Efficiently
- Start with an outline mapping to the required components
- For Tier 2/3: Write Tier 1 content first, then expand
- Use clear section headers corresponding to assignment parts
- Budget your pages: don't spend 3 pages on technical explanation leaving no room for analysis

### Strong vs. Weak Approaches

**Strong Example (Tier 1):**
> Topic: Neural Radiance Fields (NeRF)
> - Explains implicit 3D scene representation and volume rendering
> - Compares to explicit 3D representations (meshes, point clouds)
> - Discusses computational cost for real-time applications
> - Reflects on accessibility barriers due to compute requirements

**Weak Example (Tier 1):**
> Topic: "Advanced CNNs"
> - Too broad, covered in class
> - No specific innovation identified
> - Generic ethical discussion about facial recognition

---

## Getting Help

- **Topic Selection:** Post on EdStem or email me (Armaan)
- **Extension Requests:** No extensions on this assignment will be granted
---

## Submission

**Submit your PDF to Gradescope.**

**In the title page of your PDF, include:** `[Tier X] - [Your Topic]`
Example: `[Tier 2] - State Space Models for Long-Range Dependencies`

**Deadline:** Sunday, December 14, 2025 at 10:00 AM EST

Late submissions will NOT be accepted. (**NO EXCEPTIONS**)

---