Special Sessions
For SISAP 2020, we call for contributions for the following three special sessions:
- Artificial Intelligence and Similarity (organized by Giuseppe Amato, Fabrizio Falchi, Claudio Gennaro, and Fabio Carrara)
- Adversarial Machine Learning & Similarity (AMLS) (organized by Laurent Amsaleg and Michael Houle)
- Similarity Techniques in Machine Learning (SiTe-ML) (organized by Anshumali Shrivastava, Sanjiv Kumar, and Rasmus Pagh)
Special session submissions may include vision/position papers, which will be evaluated based on the quality of the arguments and ideas proposed in the papers. In order to ensure high quality of the conference papers, all papers submitted to special sessions will be peer-reviewed, including papers solicited by the special session chairs. If a special session has many high-quality submissions, some of the submissions may potentially be moved to regular sessions; likewise, relevant accepted submissions may be moved into special sessions.
Artificial Intelligence and Similarity
Special Session organized by Giuseppe Amato (ISTI-CNR, Italy), Fabrizio Falchi (ISTI-CNR, Italy), Claudio Gennaro (ISTI-CNR, Italy), and Fabio Carrara (ISTI-CNR, Italy)
Recent years have witnessed a strong renovated interest in artificial intelligence.
This has been mainly due to the outstanding performance offered by deep learning methods,
thanks to the innovative architecture, the availability of huge quantities of training data
and the high computing power provided by GPU architectures.
There are noteworthy relationships between methods of artificial
intelligence and similarity search. For example, similarity search
is more often executed on features (for instance image features)
extracted using artificial intelligence methods, rather than
hand-crafted methods. This poses new challenges, given that such
features have generally much higher (intrinsic) dimensionality,
than hand-crafted features. Artificial intelligence has also been
used as an instrument for building efficient and effective methods
for similarity search. Consider for instance methods of metric
learning, learning to index, and learning to hash. In this case,
artificial intelligence is used as an alternative to hand-crafted
structures and data coding to obtain efficient and effective
similarity search algorithms. In addition, consider that what is
generally behind many machine learning methods is the possibility
of comparing, judging relationships, and estimating similarity
among objects or entities in order to classify, recognize, and take
decisions.
In this special session, we seek contributions where artificial
intelligence and similarity evaluation/searching is either
exploited in synergy or supporting one the other. Both mature
research papers and position papers are welcome.
Topics include, but are not limited to:
- AI-based feature extraction and similarity
- Clustering methods
- Machine learning methods and similarity
- Metric learning
- Learning to index
- Learning to hash
Adversarial Machine Learning & Similarity (AMLS)
Special Session organized by Laurent Amsaleg (CNRS, France), Michael E. Houle (NII, Japan)
Most machine learning techniques are very sensitive to adversarial
perturbations, in that their outputs can be corrupted through the
addition of a small amount of carefully crafted noise. Why such
tiny content modifications succeed in producing severe errors is
still not well understood. Defensive mechanisms have been proposed
to increase the robustness of the training and test phases,
reducing the sensitivity of machine learning approaches to attacks
based on adversarial samples. Initial investigation into this issue
has concentrated on applications involving images, computer vision,
and classification.
In this SISAP special session, we solicit research that tackles the issues of
adversarial perturbation in machine learning, from the perspective of
similarity search and its applications. The goal of AMLS being to expose the
SISAP community to such issues and to foster discussions, we also welcome
vision and position papers as well as papers presenting contributions in
their early stages.
Topics include, but are not limited to:
- Similarity-based learning criteria for the detection of and defense from adversarial attack.
- Generation of adversarial samples for use in machine learning settings.
- Theoretical investigation of the phenomenon of adversarial perturbation, attack, defense and training set poisoning.
- Adversarial perturbation in contexts such as retrieval, clustering, outlier detection, feature selection, feature embeddings, search engines, hashing, similarity graphs, metric and non-metric spaces.
- Adversarial attack & defense applied to complex data, such as intrusion logs, malware, time series data, and other multimedia data.
Similarity Techniques in Machine Learning (SiTe-ML)
Special Session organized by Rasmus Pagh (IT University of Copenhagen, Denmark), Anshumali Shrivastava (Rice University, USA), and Sanjiv Kumar (Google Research, USA)
Similarity Search (SS) is a fundamental operation in data processing
applications. Research on SS has resulted in a large number of theoretical
and algorithmic developments in getting around and understanding the
fundamental hardness of SS. This has led to the development of novel
fundamental ideas such as tree-based space partitioning, metric space
similarity search, and randomized locality-sensitive hashing (LSH). Modern
machine learning algorithms are generally dealing with high-dimensional data
representations. They need to be robust to noise to be robust and to
generalize well to unseen data. It is perhaps no surprise that methods
developed for SS can be deployed to make ML algorithms more scalable. Such
transfer of knowledge and techniques has been gaining momentum in recent
years.
Ideas of getting around the curse of dimensionality from similarity search,
in particular, LSH and randomized trees, have broken computational barriers
in classical statistical estimations such as partition function estimation,
online gradient estimation, anomaly detection, and deep learning.
The natural need for fast SS, especially maximum inner product search, has
led a flurry of work in using SS for recommendation engines and deep learning
with very large output space popularly known as extreme classification.
Extreme classification is one of the prime facilitators of modern success in
natural language processing.
The special session aims to explore the above mentioned ideas and related
applications, also providing a feedback loop to researchers on similarity
techniques that may inspire new investigations. Research, application as well
as position papers are welcome.
Topics include, but are not limited to:
- Case studies of using similarity search as a building block in machine learning systems
- Machine learning algorithms that make use of techniques also used in similarity search
- Novel similarity search questions motivated by machine learning applications
