Slip-box

Each note makes one claim, stated in its title. Notes link to the claims they depend on, and each link says why.

• A coarser aggregate can't add information the finer one lacked, but a denser one can expose signal the sparser one buried
• A function's robustness is determined by the component with the worst asymptotic growth rate
• 2 A loss function is a preference over error distributions
• 1 A permutation-invariant representation guarantees that the feature vector depends only on content and never on order
• A quantity that compounds multiplicatively across steps can be made additive by applying the log function to it, which lets the sum-based statistical toolbox apply to it
• A randomized control group isolates a causal effect because regression to the mean affects both arms equally and cancels in the comparison
• 2 A replication knob is a source of variation under a fixed experimental design
• 1 A replication knob must be both an arbitrary choice and unrelated to the target variable
• 1 A rule of thumb for performing propositionalization is starting bottom-up by under-aggregating
• A single observation of extreme value can't reveal its own noise-to-signal ratio, but repeated observations can - by how far values pull back toward the mean
• An experiment can be a comparison, or a gate, depending on the intent, not its mechanics
• 2 Averaging an ensemble's predictions preserves shared signal and cancels uncorrelated errors
• 4 Avoiding outlier magnitudes makes the measure robust
• Because the true value of a parameter is unreachable, the Bayesian expresses it as a distribution over possible values
• 1 Choosing a loss means deciding the price of errors
• 1 Decision trees select the feature and its split threshold of a node based on the sum of errors of their child leaves
• Distribution shape determines which measure of location is trustworthy
• 4 Every data point is made of signal and noise
• 3 Extreme values regress toward the mean because they carry noise that doesn't repeat
• 1 Frequentists put the probability on the data, believing the parameter is fixed, while Bayesians put the probability on the parameter and believe the data is fixed
• 1 Git worktrees allow working on multiple branches in parallel
• 2 Group leakage happens when evaluating generalization but splitting at the observation level
• Harmonic mean is sensitive to small values
• 1 Log-space target can be used to encode proportional error price across the target distribution
• MAD rescaled by 1.4826 is comparable to σ under a Gaussian distribution
• 1 Mean and median can be misleading when the distribution is multimodal
• Mean calculated on a weighted sample ≈ a weighted mean of the population
• 1 Measures of location and dispersion are complementary
• Measures of location represent position and dispersion represents distance
• Modeling and monitoring are inverses
• 1 Noise can be absolute (randomness) or relative (to some target)
• On standardized axes, the regression slope of a linear function tells us the effect of regression to the mean on the data
• 3 Outliers mask themselves under ±2σ detection because they inflate the σ
• 2 Over-aggregation is irreversible, under-aggregation is recoverable
• 1 Parallel rates add; sequential rates don't
• 1 Plain git makes it difficult to work on multiple branches in parallel
• 1 Random splitting temporal data makes the target interpolation not forecasting
• 1 Reciprocals of sequential rates are additive when the segment lengths are equal
• Reproducibility is the ceteris-paribus condition that makes replication-variance a clean measurement
• 1 Robust outlier detection requires robustness on both location and dispersion layers
• 1 Run variation sources the axes along which ceteris paribus can be violated
• Seed granularity (shared vs per-source) is a design choice between measuring total noise and attributing it to sources
• Target leakage happens when training set contains a feature carrying target information unavailable at inference time
• 2 Test set's split structure determines the claim about the model
• 1 The ±2σ outlier rule assumes a Gaussian distribution
• 1 The model will sacrifice errors the loss considers cheap in order to minimize the errors the loss considers expensive
• 5 The split type has to preserve the structure you're generalizing along
• 2 The split unit follows what you generalize to
• The stage-in-workflow components form a handoff chain
• 2 The stage-in-workflow dimension classifies experiments by which artifact is the subject of iteration at this stage
• Trimmed mean is the dial between a mean and the median
• 1 Unlike inter-hypothesis, running intra-hypothesis experiments allows comparing models ceteris paribus
• Using Mutual Information first allows identifying which features carry any signal, regardless of their relationship type
• 3 Validation and Test sets must have identical split structure to support the same claim
• 2 Validation set's split structure selects the model
• 1 Weighted mean allows introducing encoded preference to the measure
• 2 What's noise and what's signal in a data point depends on how the target variable is defined
• 1 Zettelkasten notes should be claim-based
• Zettelkasten's value comes from complex conclusions emerging from atomic claims

The number shows how many other claims link to it.