Research Brainstorming

Structured methods for creative research ideation.

When to Use

•Starting a new research direction
•Generating paper ideas
•Exploring extensions of existing work
•Overcoming creative blocks
•Finding novel angles on problems

Brainstorming Principles

Diverge, Then Converge

•Divergent phase: Generate many ideas without judgment
•Convergent phase: Evaluate and select best ideas

Rules for Divergent Phase

•Quantity over quality initially
•No criticism or evaluation
•Build on others' ideas
•Wild ideas are welcome
•Combine and improve ideas

Rules for Convergent Phase

•Apply evaluation criteria
•Consider feasibility
•Rank by potential impact
•Identify quick wins vs. long-term bets

SCAMPER Method

SCAMPER is a checklist for transforming existing ideas:

S - Substitute

What can be replaced?

Prompt	Example
Different model?	BERT → GPT-4
Different data?	Text → Code
Different task?	Classification → Generation
Different metric?	Accuracy → Efficiency

C - Combine

What can be merged?

Prompt	Example
Combine methods?	RL + Language Models
Combine modalities?	Vision + Language
Combine tasks?	Multi-task learning
Combine datasets?	Domain adaptation

A - Adapt

What can be borrowed from elsewhere?

Prompt	Example
From another field?	Physics → ML theory
From another domain?	Vision → NLP
From industry?	Production systems → Research
From nature?	Biological systems → Algorithms

M - Modify/Magnify/Minimize

What can be changed in scale or intensity?

Prompt	Example
Make bigger?	Scale up model/data
Make smaller?	Efficient/compressed models
More extreme?	Harder benchmarks
More subtle?	Fine-grained evaluation

P - Put to Other Uses

What else could this be used for?

Prompt	Example
Different application?	Translation → Summarization
Different audience?	Researchers → Practitioners
Different constraint?	Accuracy → Latency

E - Eliminate

What can be removed?

Prompt	Example
Remove component?	Attention without position
Remove assumption?	Without labeled data
Remove constraint?	Without domain restriction

R - Reverse/Rearrange

What can be reordered or inverted?

Prompt	Example
Reverse process?	Generation → Understanding
Opposite approach?	Top-down → Bottom-up
Different order?	Pre-train → Fine-tune vs opposite

Morphological Analysis

Systematically explore combinations of dimensions.

Step 1: Identify Dimensions

List key aspects of your research area:

Dimension	Options
Task	Classification, Generation, Ranking
Model	Transformer, RNN, MLP
Data	Text, Code, Multi-modal
Scale	Small, Medium, Large
Supervision	Supervised, Self-supervised, RL

Step 2: Generate Combinations

Create a matrix and explore intersections:

code

Task × Model × Data × Scale × Supervision
= Many possible combinations

Step 3: Evaluate Combinations

For each interesting combination:

• Is it novel?
• Is it feasible?
• Is it interesting?
• Does it address a gap?

Template

markdown

## Morphological Analysis: [Topic]

### Dimensions
1. [Dimension 1]: [Option A, Option B, Option C]
2. [Dimension 2]: [Option A, Option B, Option C]
3. [Dimension 3]: [Option A, Option B, Option C]

### Promising Combinations
| D1 | D2 | D3 | Why Interesting |
|----|----|----|-----------------|
| | | | |

### Selected Ideas
1. [Combination]: [Why pursue this]

Cross-Domain Analogies

Find inspiration from analogous problems in other fields.

Process

•Abstract your problem: What is it fundamentally about?
•Find analogies: What other fields face similar challenges?
•Study solutions: How do they address it?
•Transfer insights: How might their solutions apply?

Analogy Sources

Your Problem	Analogous Field	Potential Insight
Scaling	Biology (growth)	Allometric scaling laws
Optimization	Physics (annealing)	Simulated annealing
Attention	Psychology (cognition)	Selective attention
Memory	Neuroscience	Working memory
Robustness	Engineering	Fault tolerance
Learning	Education	Curriculum learning

Template

markdown

## Cross-Domain Analogy

### Our Problem
[Description of the challenge]

### Analogous Problem
**Field**: [Field name]
**Problem**: [Their version of the challenge]
**Solution**: [How they address it]

### Transfer Opportunity
[How their insight might apply to ML]

### Research Idea
[Concrete research direction]

Assumption Reversal

Challenge fundamental assumptions.

Process

•List assumptions in current approaches
•For each assumption, ask "What if the opposite were true?"
•Explore implications of reversals

Template

markdown

## Assumption Reversal: [Topic]

### Current Assumptions
1. [Assumption 1]
2. [Assumption 2]
3. [Assumption 3]

### Reversals
| Assumption | Reversal | Implication |
|------------|----------|-------------|
| More data is better | Less data could be better | Data efficiency research |
| Bigger models are better | Smaller could be better | Efficient architectures |
| Pre-training helps | Training from scratch | Task-specific models |

Problem Reframing

View the problem from different angles.

Perspectives

Perspective	Question
User	What does the end user actually need?
System	What are the computational constraints?
Data	What data is actually available?
Theory	What would a theoretical analysis reveal?
Ethics	What are the societal implications?

Reframing Prompts

•"Instead of solving X, what if we solved Y?"
•"What problem are we actually trying to solve?"
•"Who else has this problem?"
•"What would make this problem go away?"
•"What would a 10x better solution look like?"

Idea Evaluation

After generating ideas, evaluate systematically.

Criteria

Criterion	Score (1-5)	Notes
Novelty		Is this new?
Impact		Would this matter?
Feasibility		Can we do this?
Clarity		Is the idea clear?
Fit		Does it match our skills/resources?

Quick Feasibility Check

• Do we have/can we get the data?
• Do we have the compute?
• Do we have the expertise?
• Can we do this in our timeframe?
• Is there a clear evaluation?

References

See references/ folder for:

•brainstorming_methods.md: Additional ideation techniques