vec2vec-two-level-adversarial-training-latent-and-output

IN premise — summaries/2026/08/24/jha-2025-vec2vec-s2-problem-formulation-unsupervised-embedding-translation.md

Created 2026-08-24T17:10:57+00:00

vec2vec trains with discriminators at two levels—latent space (D¹ₗ, D²ₗ) and output embedding space (D₁, D₂)—yielding four GAN loss terms: L_GAN(D₁,F₁) + L_GAN(D₂,F₂) + L_GAN(D¹ₗ,T∘A₁) + L_GAN(D²ₗ,T∘A₂).

Summary

vec2vec doesn't just check that its final translated word vectors look right; it also forces the intermediate representations to be plausible, using separate adversarial checks at both stages. This means the model is penalized for producing bad translations at any level of its pipeline, which should make the learned cross-lingual mapping more robust than an approach that only judges the end result.