BERT encoder pretrained from scratch với
VICReg
(Variance-Invariance-Covariance Regularization).
Hai masked text views được encode bởi một BERT encoder dùng chung, sau đó đưa qua expander MLP.
VICReg kết hợp 3 loss terms để căn chỉnh các views và ngăn feature collapse mà không cần negative samples:
Loss term
Hệ số
Mô tả
Invariance
25.0
MSE giữa z1 và z2 (căn chỉnh hai views)
Variance
25.0
Giữ std của mỗi chiều ≥ 1 (chống collapse)
Covariance
1.0
Decorrelate các chiều embedding
Kiến trúc
Text → BERT (mean-pool) → z ∈ R^768 → Expander MLP → z' ∈ R^3072
↑ VICReg loss áp dụng tại đây
Expander gồm 3 lớp Linear-BatchNorm-ReLU (dim =
3072
).
jepa_proof_vicreg huggingface.co is an AI model on huggingface.co that provides jepa_proof_vicreg's model effect (), which can be used instantly with this ducanhdinh jepa_proof_vicreg model. huggingface.co supports a free trial of the jepa_proof_vicreg model, and also provides paid use of the jepa_proof_vicreg. Support call jepa_proof_vicreg model through api, including Node.js, Python, http.
jepa_proof_vicreg huggingface.co is an online trial and call api platform, which integrates jepa_proof_vicreg's modeling effects, including api services, and provides a free online trial of jepa_proof_vicreg, you can try jepa_proof_vicreg online for free by clicking the link below.
ducanhdinh jepa_proof_vicreg online free url in huggingface.co:
jepa_proof_vicreg is an open source model from GitHub that offers a free installation service, and any user can find jepa_proof_vicreg on GitHub to install. At the same time, huggingface.co provides the effect of jepa_proof_vicreg install, users can directly use jepa_proof_vicreg installed effect in huggingface.co for debugging and trial. It also supports api for free installation.