Skip to content

Commit

Permalink
move schedule down the page
Browse files Browse the repository at this point in the history
  • Loading branch information
@harsha-simhadri
harsha-simhadri committed Sep 30, 2023
1 parent 59590f6 commit 4b3208b
Showing 1 changed file with 80 additions and 80 deletions.
160 changes: 80 additions & 80 deletions neurips21.html
View file
Original file line number Diff line number Diff line change
Expand Up @@ -95,86 +95,6 @@ <h2> Code, Report, Results and Blogs</h2>
</ul>


<div id="schedule">
<h3>Summary of NeurIPS'21 event</h3>
The <a href="https://neurips.cc/virtual/2021/competition/22443#collapse21991">NeurIPS session</a> for this competition happend on Dec 8, 2021. See slides and recordings of the talks below.
<strong><a href="https://neurips.cc/virtual/2021/competition/22443#collapse-sl-21991">Overview Talk</a> and <a href="https://neurips.cc/virtual/2021/competition/22443#collapse48501">Break-out session</a> schedule (GMT)</strong>.
<ul>
<li> 11:05-11:25: Overview Talk (<a href="templates/slides/comp-overview.pptx">slides</a>, <a href="https://youtu.be/jjWxVxKSn1c">video</a>)</li>
<li> 12:00-12:45: Overview of results presented by organizers, followed by Q&A </li>
<ul>
<li>Standard hardware tracks T1 and T2 results (<a href="templates/slides/T1_T2_results.pptx">slides</a>)</li>
<li>Custom hardware track T3 results (<a href="templates/slides/T3-results.pptx">slides</a>)</li>
</ul>
<li> 12:45-13:20: Invited talk 1 by <a href="http://www.cs.columbia.edu/~andoni/">Prof. Alexandr Andoni</a>: Learning to Hash Robustly, with Guarantees (<a href="templates/slides/lsh_neurips21.pptx">slides</a>, <a href="https://youtu.be/WcOu2LF57HI">video</a>)</li>
<li> 13:20-13:55: Invited talk 2 by <a href="https://www.cs.rice.edu/~as143/">Prof. Anshumali Shrivastava</a>:Iterative Repartitioning for Learning to Hash and the Power of k-Choices (<a href="templates/slides/invited-talk-anshu.pptx">slides</a>, <a href="https://youtu.be/TOHByhrlOiw">video</a>)</li>
<li> 13:55-14:30: Talks from track winners.
<ul>
<li>Track 1: <a href="https://github.com/harsha-simhadri/big-ann-benchmarks/pull/69">kst_ann_t1</a> Li Liu, Jin Yu, Guohao Dai, Wei Wu, Yu Qiao, Yu Wang, Lingzhi Liu, <i>Kuaishou Technology and Tsinghua University</i> (<a href="https://youtu.be/dc_PGe7l5f8">video</a>)</li>
<li>Track 2: <a href="https://github.com/harsha-simhadri/big-ann-benchmarks/pull/70">BBANN</a> Xiaomeng Yi, Xiaofan Luan, Weizhi Xu, Qianya Cheng, Jigao Luo, Xiangyu Wang, Jiquan Long, Xiao Yan, Zheng Bian, Jiarui Luo, Shengjun Li, Chengming Li, <i>Zilliz and Southern University of Science and Technology</i> (<a href="templates/slides/Track2_xiaomeng_yi.pdf">slides</a>, <a href="https://youtu.be/MJcFwG5OzKM">video</a>)</li>
<li>Track 3: <a href="https://github.com/harsha-simhadri/big-ann-benchmarks/pull/63">OptaNNe</a> Sourabh Dongaonkar, Mark Hildebrand, Mariano Tepper, Cecilia Aguerrebere, Ted Willke, Jawad Khan, <i>Intel Corporation, Intel Labs and UC Davis</i> (<a href="templates/slides/optanne.pptx">slides</a>, <a href="https://youtu.be/XgtKUsGhyG4">video</a>)</li>
</ul>
</li>
<li> 14:30-15:00: Open discussion on competition and future directions (<a href="https://github.com/harsha-simhadri/big-ann-benchmarks/issues/90">github thread</a>, <a href="https://youtu.be/9I4GC1eGWCk">video</a>) </li>
</ul>
<p>
<strong>Abstract for Invited talk: "Learning to Hash Robustly, with Guarantees"</strong> <BR>
There is a gap between the high-dimensional nearest neighbor search
(NNS) algorithms achieving the best worst-case guarantees and the
top-performing ones in practice. The former are based on indexing via
the randomized Locality Sensitive Hashing (LSH), and its
derivatives. The latter "learn" the best indexing method in order to
speed-up NNS, crucially adapting to the structure of the given
dataset. Alas, the latter also almost always come at the cost of
losing the guarantees of either correctness or robust performance on
adversarial queries (or apply to datasets with an assumed extra
structure/model).

How can we bridge these two perspectives and bring the best of both
worlds? As a step in this direction, we will talk about an NNS algorithm
that has worst-case guarantees essentially matching that of
theoretical algorithms, while optimizing the hashing to the structure
of the dataset (think instance-optimal algorithms) for performance on
the minimum-performing query. We will discuss the algorithm's ability
to optimize for a given dataset from both theoretical and practical
perspective.
</p>

<p>
<strong>Abstract for Invited talk: "Iterative Repartitioning for Learning to Hash and the Power of k-Choices"</strong> <BR>
Dense embedding models are commonly deployed in commercial
search engines, wherein all the vectors are pre-computed, and
near-neighbor search (NNS) is performed with the query vector to find
relevant documents. However, the bottleneck of indexing a large number
of dense vectors and performing an NNS hurts the query time and
accuracy of these models. In this talk, we argue that high-dimensional
and ultra-sparse embedding is a significantly superior alternative to
dense low-dimensional embedding for both query efficiency and
accuracy. Extreme sparsity eliminates the need for NNS by replacing
them with simple lookups, while its high dimensionality ensures that
the embeddings are informative even when sparse. However, learning
extremely high dimensional embeddings leads to blow-up in the model
size. To make the training feasible, we propose a partitioning
algorithm that learns such high-dimensional embeddings across multiple
GPUs without any communication. We theoretically prove that our way of
one-sided learning is equivalent to learning both query and label
embeddings. We call our novel system designed on sparse embeddings as
IRLI (pronounced `early'), which iteratively partitions the items by
learning the relevant buckets directly from the query-item relevance
data. Furthermore, IRLI employs a superior power-of-k-choices based
load balancing strategy. We mathematically show that IRLI retrieves
the correct item with high probability under very natural assumptions
and provides superior load balancing. IRLI surpasses the best
baseline's precision on multi-label classification while being 5x
faster on inference. For near-neighbor search tasks, the same method
outperforms the state-of-the-art Learned Hashing approach NeuralLSH by
requiring only ~ {1/6}^th of the candidates for the same recall. IRLI
is both data and model parallel, making it ideal for distributed GPU
implementation. We demonstrate this advantage by indexing 100 million
dense vectors and surpassing the popular FAISS library by >10%.
</p>
</div>

<div id="why">
<h2>Why this competition?</h2>
In the past few years, we’ve seen a lot of new research and creative approaches for large-scale ANNS, including:
Expand Down Expand Up @@ -521,6 +441,86 @@ <h4>Timeline (subject to change)</h4>
</div>


<div id="schedule">
<h3>Summary of NeurIPS'21 event</h3>
The <a href="https://neurips.cc/virtual/2021/competition/22443#collapse21991">NeurIPS session</a> for this competition happend on Dec 8, 2021. See slides and recordings of the talks below.
<strong><a href="https://neurips.cc/virtual/2021/competition/22443#collapse-sl-21991">Overview Talk</a> and <a href="https://neurips.cc/virtual/2021/competition/22443#collapse48501">Break-out session</a> schedule (GMT)</strong>.
<ul>
<li> 11:05-11:25: Overview Talk (<a href="templates/slides/comp-overview.pptx">slides</a>, <a href="https://youtu.be/jjWxVxKSn1c">video</a>)</li>
<li> 12:00-12:45: Overview of results presented by organizers, followed by Q&A </li>
<ul>
<li>Standard hardware tracks T1 and T2 results (<a href="templates/slides/T1_T2_results.pptx">slides</a>)</li>
<li>Custom hardware track T3 results (<a href="templates/slides/T3-results.pptx">slides</a>)</li>
</ul>
<li> 12:45-13:20: Invited talk 1 by <a href="http://www.cs.columbia.edu/~andoni/">Prof. Alexandr Andoni</a>: Learning to Hash Robustly, with Guarantees (<a href="templates/slides/lsh_neurips21.pptx">slides</a>, <a href="https://youtu.be/WcOu2LF57HI">video</a>)</li>
<li> 13:20-13:55: Invited talk 2 by <a href="https://www.cs.rice.edu/~as143/">Prof. Anshumali Shrivastava</a>:Iterative Repartitioning for Learning to Hash and the Power of k-Choices (<a href="templates/slides/invited-talk-anshu.pptx">slides</a>, <a href="https://youtu.be/TOHByhrlOiw">video</a>)</li>
<li> 13:55-14:30: Talks from track winners.
<ul>
<li>Track 1: <a href="https://github.com/harsha-simhadri/big-ann-benchmarks/pull/69">kst_ann_t1</a> Li Liu, Jin Yu, Guohao Dai, Wei Wu, Yu Qiao, Yu Wang, Lingzhi Liu, <i>Kuaishou Technology and Tsinghua University</i> (<a href="https://youtu.be/dc_PGe7l5f8">video</a>)</li>
<li>Track 2: <a href="https://github.com/harsha-simhadri/big-ann-benchmarks/pull/70">BBANN</a> Xiaomeng Yi, Xiaofan Luan, Weizhi Xu, Qianya Cheng, Jigao Luo, Xiangyu Wang, Jiquan Long, Xiao Yan, Zheng Bian, Jiarui Luo, Shengjun Li, Chengming Li, <i>Zilliz and Southern University of Science and Technology</i> (<a href="templates/slides/Track2_xiaomeng_yi.pdf">slides</a>, <a href="https://youtu.be/MJcFwG5OzKM">video</a>)</li>
<li>Track 3: <a href="https://github.com/harsha-simhadri/big-ann-benchmarks/pull/63">OptaNNe</a> Sourabh Dongaonkar, Mark Hildebrand, Mariano Tepper, Cecilia Aguerrebere, Ted Willke, Jawad Khan, <i>Intel Corporation, Intel Labs and UC Davis</i> (<a href="templates/slides/optanne.pptx">slides</a>, <a href="https://youtu.be/XgtKUsGhyG4">video</a>)</li>
</ul>
</li>
<li> 14:30-15:00: Open discussion on competition and future directions (<a href="https://github.com/harsha-simhadri/big-ann-benchmarks/issues/90">github thread</a>, <a href="https://youtu.be/9I4GC1eGWCk">video</a>) </li>
</ul>
<p>
<strong>Abstract for Invited talk: "Learning to Hash Robustly, with Guarantees"</strong> <BR>
There is a gap between the high-dimensional nearest neighbor search
(NNS) algorithms achieving the best worst-case guarantees and the
top-performing ones in practice. The former are based on indexing via
the randomized Locality Sensitive Hashing (LSH), and its
derivatives. The latter "learn" the best indexing method in order to
speed-up NNS, crucially adapting to the structure of the given
dataset. Alas, the latter also almost always come at the cost of
losing the guarantees of either correctness or robust performance on
adversarial queries (or apply to datasets with an assumed extra
structure/model).

How can we bridge these two perspectives and bring the best of both
worlds? As a step in this direction, we will talk about an NNS algorithm
that has worst-case guarantees essentially matching that of
theoretical algorithms, while optimizing the hashing to the structure
of the dataset (think instance-optimal algorithms) for performance on
the minimum-performing query. We will discuss the algorithm's ability
to optimize for a given dataset from both theoretical and practical
perspective.
</p>

<p>
<strong>Abstract for Invited talk: "Iterative Repartitioning for Learning to Hash and the Power of k-Choices"</strong> <BR>
Dense embedding models are commonly deployed in commercial
search engines, wherein all the vectors are pre-computed, and
near-neighbor search (NNS) is performed with the query vector to find
relevant documents. However, the bottleneck of indexing a large number
of dense vectors and performing an NNS hurts the query time and
accuracy of these models. In this talk, we argue that high-dimensional
and ultra-sparse embedding is a significantly superior alternative to
dense low-dimensional embedding for both query efficiency and
accuracy. Extreme sparsity eliminates the need for NNS by replacing
them with simple lookups, while its high dimensionality ensures that
the embeddings are informative even when sparse. However, learning
extremely high dimensional embeddings leads to blow-up in the model
size. To make the training feasible, we propose a partitioning
algorithm that learns such high-dimensional embeddings across multiple
GPUs without any communication. We theoretically prove that our way of
one-sided learning is equivalent to learning both query and label
embeddings. We call our novel system designed on sparse embeddings as
IRLI (pronounced `early'), which iteratively partitions the items by
learning the relevant buckets directly from the query-item relevance
data. Furthermore, IRLI employs a superior power-of-k-choices based
load balancing strategy. We mathematically show that IRLI retrieves
the correct item with high probability under very natural assumptions
and provides superior load balancing. IRLI surpasses the best
baseline's precision on multi-label classification while being 5x
faster on inference. For near-neighbor search tasks, the same method
outperforms the state-of-the-art Learned Hashing approach NeuralLSH by
requiring only ~ {1/6}^th of the candidates for the same recall. IRLI
is both data and model parallel, making it ideal for distributed GPU
implementation. We demonstrate this advantage by indexing 100 million
dense vectors and surpassing the popular FAISS library by >10%.
</p>
</div>


<div id="organizers">
<h2>Organizers and Dataset Contributors</h2>
Expand Down

0 comments on commit 4b3208b

Please sign in to comment.