Claude-kit topic-landscape

Overview

This skill provides the methodology for mapping any ML/AI problem space from scratch. It's the knowledge behind how

architecture-scout

structures its search and how

combination-brainstormer

identifies gaps. Use it to turn a vague topic into a structured understanding of what exists, what's missing, and where opportunity lies.

---

Step 1: Topic Decomposition

Break any ML topic into orthogonal dimensions:

Dimension A: Problem formulation

How is the problem framed? Same underlying data can be approached differently:

• Supervised vs. unsupervised vs. self-supervised
• Classification vs. regression vs. generation vs. detection
• Online vs. offline
• Point prediction vs. sequence prediction

Dimension B: Architecture family

What model architectures are used?

• Transformer-based (attention mechanisms)
• SSM-based (Mamba, S4, state space models)
• CNN-based (convolutions, residual networks)
• GNN-based (graph neural networks)
• RNN-based (LSTM, GRU — legacy but still used)
• Hybrid (combinations of the above)
• Classical ML (random forests, SVMs, etc.)

Dimension C: Data modality

What kind of data does the method operate on?

• Time-series / sequential
• Images / video
• Text / language
• Tabular / structured
• Graph / relational
• Audio / speech
• Multimodal (combinations)

Dimension D: Training paradigm

How is the model trained?

• Supervised (labeled examples)
• Self-supervised (pretext tasks, contrastive learning, masked prediction)
• Semi-supervised (few labels + many unlabeled)
• Unsupervised (no labels — clustering, reconstruction, density estimation)
• Few-shot / zero-shot / meta-learning
• Reinforcement learning

Creating a topic map

Cross the dimensions to find your specific niche:

[Problem] × [Architecture] × [Data] × [Training] = your research position

Example: anomaly detection × Mamba × time-series × self-supervised

Each cell in this cross-product is either occupied (existing work) or empty (gap = opportunity).

---

Step 2: Method Taxonomy

For any problem domain, identify the major families of approaches:

Template

[Domain] methods:
├── Family A: [description]
│   ├── Sub-approach A1
│   └── Sub-approach A2
├── Family B: [description]
│   ├── Sub-approach B1
│   └── Sub-approach B2
└── Family C: [description]

Example: Anomaly Detection

Anomaly detection methods:
├── Reconstruction-based: learn normal → flag high reconstruction error
│   ├── Autoencoders (vanilla, variational, masked)
│   └── Diffusion-based (denoise normal, anomalies resist denoising)
├── Contrastive / self-supervised: learn representations → anomalies map far from normal
│   ├── Contrastive learning (SimCLR-style)
│   └── Knowledge distillation (student-teacher disagreement = anomaly)
├── Density-based: model the normal distribution → low-density = anomaly
│   ├── Normalizing flows
│   └── Energy-based models
├── Classification-based: learn a boundary around normal
│   ├── One-class SVM / SVDD
│   └── Deep SVDD
└── Hybrid: combine multiple signals
    ├── Multi-head (reconstruction + classification)
    └── Ensemble (multiple diverse detectors)

---

Step 3: Gap Identification

Gaps are where novelty lives. Look for:

Unexplored cells

Cross architecture × problem. Which architectures haven't been tried for this task?

• "Has anyone used [new architecture] for [this task]?" → search arXiv
• If no results: that's a gap

Underexplored combinations

Two techniques that have each been shown to work but haven't been combined:

• "Method A improves X. Method B improves Y. Has anyone combined A + B?"
• If no: potential idea

Missing modalities

A technique works for modality A. Has it been adapted for modality B?

• "Vision Mamba exists. Does Audio Mamba? Does Graph Mamba?"

Scale gaps

An approach works at small scale. Has it been made efficient / scaled up?

• "Method X works but requires 8 A100s. Can it work on a single consumer GPU?"

Evaluation gaps

A method was evaluated on dataset A only. Does it generalize?

• "All papers in this area use benchmark X. What about benchmark Y?"

---

Step 4: State of the Art Tracking

How to determine current SOTA

1. Check Papers With Code for the relevant benchmark
2. Check the most recent survey paper (last 12 months)
3. Check the most-cited recent paper — its baselines show the competitive landscape
4. Cross-reference: if a method is SOTA on PwC but no one cites it, it might be overfitting to the benchmark

What to track

• Current best: method, result, paper, year
• Trajectory: is improvement slowing down? (diminishing returns = mature area, harder to beat)
• Recent jumps: any method that significantly beat SOTA recently (new paradigm)
• Open challenges: what does the survey paper say is still unsolved?

---

Step 5: Cross-Domain Transfer Opportunities

The highest-novelty ideas often come from applying technique X from domain A to domain B.

Pattern

Technique from [Source Domain] → applied to [Target Domain]

How to find these

1. Identify what's working well in adjacent domains (e.g., language modeling techniques for time-series)
2. Ask: "Why hasn't this been tried?" — if the answer is "no one thought of it" (vs. "it fundamentally doesn't apply"), that's an opportunity
3. Look for analogies: "This problem in domain B is structurally similar to a solved problem in domain A"

Recent successful transfers

• Transformers: NLP → Vision → Audio → Time-series → Graphs
• Diffusion models: Image generation → Audio → Video → Anomaly detection
• Contrastive learning: Vision → NLP → Time-series → Tabular
• State space models (Mamba): Language → Vision → Audio → (what's next?)

Signals that transfer will work

• The data has similar structure (sequential, spatial, hierarchical)
• The task has a similar objective (reconstruction, classification, generation)
• The challenge is similar (long-range dependencies, efficiency, few labels)

---

Quick Reference: Mapping a New Topic

1. Define the topic in one sentence
2. Decompose: problem × architecture × data × training
3. Build the method taxonomy (what families of approaches exist?)
4. Check SOTA (Papers With Code + recent survey)
5. Cross the dimensions → find empty cells = gaps
6. Check adjacent domains for transferable techniques
7. Output: topic map with gaps highlighted