mrl-nq

AdANNS: A Framework for Adaptive Semantic Search 💃

Aniket Rege*, Aditya Kusupati*, Sharan Ranjit S, Alan Fan, Qinqqing Cao, Sham Kakade, Prateek Jain, Ali Farhadi

GitHub: https://github.com/RAIVNLab/AdANNS

Arxiv: https://arxiv.org/abs/2305.19435

<p align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/64af72d4a609b29cc7b5919b/QYOqZ25qc9pTWlFR0D0VF.png" alt="drawing" width="600"/> Adaptive representations can be utilized effectively in the decoupled components of clustering and searching for a better accuracy-compute trade-off (AdANNS-IVF). </p>

We provide four BERT-Base models finetuned on Natural Questions with Matryoshka Representation Learning (MRL).

A vanilla pretrained BERT-Base has a 768-d representation (information bottleneck). As we train with MRL, we enforce the network to learn representations at multiple granularities nested within a 768-d embedding. The granularities at which we finetune BERT-Base with Matroyshka Loss are specified in the folder name, e.g. for dpr-nq-d768_384_192_96_48, we have d=[48, 96, 192, 384, 768].

You can easily load an mrl-nq model as follows:

from transformers import BertModel
import torch

model = BertModel.from_pretrained('dpr-nq-d768_384_192_96_48')

Citation

If you find this project useful in your research, please consider citing:

@inproceedings{rege2023adanns,
      title={AdANNS: A Framework for Adaptive Semantic Search}, 
      author={Aniket Rege and Aditya Kusupati and Sharan Ranjit S and Alan Fan and Qingqing Cao and Sham Kakade and Prateek Jain and Ali Farhadi},
      year={2023},
      booktitle = {Advances in Neural Information Processing Systems},
      month     = {December},
      year      = {2023},
}