Scaling laws analysis for SAE hyperparameters

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Power law scalingconcept0.786
  Observation that SAE loss decreases as a power law with compute budget.
• Scaling laws can be used to guide the training of sparse autoencoders.claim0.777
  Compute-optimal hyperparameters follow predictable power-law relationships.
• A single SAE hyperparameter procedure driven by an intrinsic dictionary health audit transfers robustly across all three EEG transformer architectures.claim0.775
  Key methodological contribution claim about architecture-agnostic SAE tuning
• Inverse Scaling Lawconcept0.771
  Hypothesis cited in paper suggesting deceptive capabilities may scale with model size
• Power Law Scaling of Data and Model Performanceconcept0.760
  Empirically observed power law relationship between data scale and model performance; supports convergence hypothesis
• Single-Layer SAE Analysis Limitationconcept0.747
  Key limitation that prevents tracing inter-layer dynamics or how steering propagates through model depth
• SAE training loss (MSE + L1 penalty with decoder norm scaling)method0.740
  The objective function combining L2 reconstruction error and L1 penalty scaled by decoder norm, used to train the SAE.
• SAE features can be grounded in clinical taxonomy (abnormality, age, sex, medication) to benchmark monosemanticity and entanglement.claim0.736
  Claim that feature grounding enables interpretability metrics.