Scoping Advanced AI Research Projects
Given a research topic (even a one-line title), produce a structured research scoping document with: problem statement, research questions, methodology outline, datasets/tools, evaluation metrics, risks/limitations, and a phased plan.
Example invocation: "Scope a project on deepfake detection via micro-expression analysis" → produces the full structure below filled in for that topic.
Progress:
- Step 1: Clarify domain and narrow the topic to a falsifiable research question
- Step 2: Survey state-of-the-art (identify 3-5 key baseline methods/papers per topic area)
- Step 3: Define methodology (model architecture, data pipeline, experimental design)
- Step 4: Identify datasets, simulators, or environments
- Step 5: Define evaluation metrics and success criteria
- Step 6: Identify ethical/societal risks and mitigation
- Step 7: Produce phased timeline (exploration → prototype → evaluation → writeup)
Step 1: Narrow the topic
Convert a broad theme into a specific, testable question.
- Broad: "Optimizing data center energy with RL"
- Narrow: "Can a multi-agent deep RL policy reduce PUE by ≥10% vs. rule-based cooling control, under variable workload traces, without violating thermal safety constraints?"
Step 2: Survey baselines
For each domain, anchor to known reference points rather than reinventing:
- DC energy + RL: DeepMind's DC cooling RL, Soft Actor-Critic for HVAC control, digital-twin simulators (EnergyPlus, CFD-based)
- AV ethics: Trolley-problem variants, Moral Machine dataset, SAE levels, value-alignment frameworks (not just "be ethical" — operationalize via constraint sets, e.g., ISO 26262 safety cases + explicit harm-weighting functions)
- Deepfake detection: Face-forensics++ , Celeb-DF, micro-expression datasets (CASME II, SAMM), action-unit (AU) based temporal features, frequency-domain artifacts as complementary signal
- GNN epidemic modeling: Mobility graphs + SEIR compartmental dynamics fused via spatio-temporal GNN (e.g., ST-GCN variants), mobile mobility datasets, synthetic urban contact networks
- Quantum-resistant IoT crypto: NIST PQC finalists (Kyber, Dilithium, Falcon), constrained-device benchmarks (ROM/RAM/energy), lattice-based vs. hash-based tradeoffs for low-power MCUs
- Algorithmic polarization: Echo chamber / filter bubble literature, engagement-optimization vs. diversity-optimization recommenders, causal inference on observational social graph data (confounding is the central challenge here)
Step 3: Methodology design
Always specify:
- Independent variable(s) — what you manipulate (algorithm, architecture, protocol parameter)
- Dependent variable(s) — measurable outcome (energy, detection AUC, R0 prediction error, polarization index)
- Controls/baselines — what you compare against (must include at least one non-trivial baseline)
- Confounders — especially critical for ethics/social-impact topics (selection bias, feedback loops)
Step 4: Data/tools
State explicitly whether using real, synthetic, or simulated data, and justify. For sensitive domains (deepfakes, social polarization), flag data access/privacy constraints upfront.
Step 5: Metrics
Pick metrics that match the claim being made. Don't use accuracy alone for imbalanced detection tasks (use AUC, precision-recall, EER for deepfakes). Use domain-appropriate epidemiological metrics (RMSE on case counts, calibration of R0) rather than generic regression metrics alone.
Step 6: Ethics/risk
For every topic — even "technical" ones like crypto or RL — name at least one dual-use or deployment risk (e.g., RL control systems failing under distribution shift; deepfake detectors enabling an arms race; crypto migration breaking legacy IoT fleets; GNN epidemic models misused for surveillance).
Step 7: Timeline
Default to 4 phases: (1) literature + baseline replication, (2) method development, (3) evaluation + ablation, (4) writeup + reproducibility package.
Example 1: Input: "Квантово-устойчивая криптография для IoT" Output:
- RQ: "Can Kyber-512 be implemented on a Cortex-M0 IoT device with <20KB RAM and <50ms handshake latency while maintaining NIST Level 1 security?"
- Baselines: ECC (Curve25519), RSA-2048, Kyber reference implementation
- Methodology: port + optimize PQC library, benchmark on real constrained hardware (not just simulation), measure RAM/ROM/energy/latency
- Metrics: handshake latency, energy per handshake (mJ), code size, security level
- Risk: hybrid classical+PQC transition period compatibility; side-channel resistance often overlooked in constrained ports
Example 2: Input: "Поляризация мнений из-за рекомендательных алгоритмов" Output:
- RQ: "Does engagement-maximizing ranking increase opinion polarization more than a diversity-regularized ranking, holding user population constant?"
- Methodology: agent-based simulation + natural experiment design (A/B if accessible, else observational with instrumental variable to handle self-selection confound)
- Metrics: polarization index (e.g., variance of opinion distribution over time), echo-chamber modularity on interaction graph
- Risk: causal claims from observational social data are fragile — explicitly state identification strategy and its limits
- Always convert a broad theme into one falsifiable, measurable research question before writing anything else
- Name a concrete baseline — "no baseline" is the most common reviewer rejection reason
- For ethics-adjacent topics (AV, polarization), operationalize abstract values into measurable proxies; don't leave "responsible AI" undefined
- Treat simulation-only results as preliminary; flag the sim-to-real gap explicitly for RL/control topics
- Reuse established datasets/benchmarks where they exist rather than building bespoke ones, unless justified
- Proposing deepfake/polarization work without addressing dataset bias or demographic skew in training data
- Using accuracy as the sole metric for imbalanced classification (deepfake detection, anomaly detection)
- Treating RL-for-DC-cooling as solved by simulation alone — real deployments must address safety constraints and distribution shift
- Vague ethics sections ("we will consider fairness") instead of a concrete constraint, metric, or audit procedure
- Ignoring energy/compute cost of the proposed method itself when the topic is energy optimization