{"paper":{"title":"DiVeQ: Differentiable Vector Quantization Using the Reparameterization Trick","license":"http://creativecommons.org/licenses/by/4.0/","headline":"DiVeQ makes vector quantization differentiable by reparameterizing it as an additive error vector.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Arno Solin, Mohammad Hassan Vali, Tom B\\\"ackstr\\\"om","submitted_at":"2025-09-30T16:17:21Z","abstract_excerpt":"Vector quantization is common in deep models, yet its hard assignments block gradients and hinder end-to-end training. We propose DiVeQ, which treats quantization as adding an error vector that mimics the quantization distortion, keeping the forward pass hard while letting gradients flow. We also present a space-filling variant (SF-DiVeQ) that assigns input to a curve constructed by the lines connecting codewords, resulting in less quantization error and full codebook usage. Both methods train end-to-end without requiring auxiliary losses or temperature schedules. In VQ-VAE image compression, "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"In VQ-VAE image compression, VQGAN image generation, and DAC speech coding tasks across various data sets, our proposed methods improve reconstruction and sample quality over alternative quantization approaches.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That modeling quantization distortion as an additive error vector (and the space-filling curve assignment) produces gradients that are both stable and unbiased enough to replace auxiliary losses and temperature schedules without degrading the hard forward-pass behavior.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DiVeQ and SF-DiVeQ make vector quantization differentiable through error-vector reparameterization and space-filling curves, yielding better reconstruction in VQ-VAE, VQGAN, and DAC tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"DiVeQ makes vector quantization differentiable by reparameterizing it as an additive error vector.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"45f129a11702c01e06b9a39c42cc34b3ca90abfbb74a2d4827eacf5ee120cad0"},"source":{"id":"2509.26469","kind":"arxiv","version":4},"verdict":{"id":"bea943f8-8673-4a7f-a5b1-c4808b6df145","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-18T11:53:17.007194Z","strongest_claim":"In VQ-VAE image compression, VQGAN image generation, and DAC speech coding tasks across various data sets, our proposed methods improve reconstruction and sample quality over alternative quantization approaches.","one_line_summary":"DiVeQ and SF-DiVeQ make vector quantization differentiable through error-vector reparameterization and space-filling curves, yielding better reconstruction in VQ-VAE, VQGAN, and DAC tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That modeling quantization distortion as an additive error vector (and the space-filling curve assignment) produces gradients that are both stable and unbiased enough to replace auxiliary losses and temperature schedules without degrading the hard forward-pass behavior.","pith_extraction_headline":"DiVeQ makes vector quantization differentiable by reparameterizing it as an additive error vector."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2509.26469/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"8e883833f512c9600f1ebc0eb73d006788d75bc76a8a761eb992d24a45ad7707"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}