Closed-form affine-regression ablation of a continuous conditioning variable

methods / Causal Validation / Closed-form affine-regression ablation of a continuous conditioning variable

Closed-form affine-regression ablation of a continuous conditioning variable

Techniqueintermediate

Fits a closed-form (Moore-Penrose pseudo-inverse) affine regression predicting each activation frame from a continuous, separately-derived conditioning embedding (e.g. a speaker-identity vector), then subtracts the predicted component at inference time — a continuous-variable analogue of LEACE/mean-projection's discrete-class subspace removal.

Used in (2 observations)

An encoding probe decomposes feature-group variance in wav2vec2 and BERTmeasured

structure: Linear Subspace · models: Wav2Vec 2.0 Base (LibriSpeech-960h), BERT-base-uncased · paper: Beyond Decodability: Reconstructing Language Model Representations with an Encoding Probe

Subtracting a regression-fit speaker component removes speaker identitymeasured

structure: Linear Subspace · models: WavLM Large · paper: Eta-WavLM: Efficient Speaker Identity Removal in Self-Supervised Speech Representations Using a Simple Linear Equation