SEA: Search Engines Amsterdam

This edition of SEA involves XIR (eXplainable Information Retrieval), with one speaker. This is a hybrid event.

Location: Lab42, Science Park Amsterdam, Room L3.36.
The Zoom link, in case you want to join online, will be available when you RSVP.

Speaker: Catherine Chen
Title: Reverse Engineering Mechanisms of Relevance in Neural IR Models
Abstract: As information retrieval (IR) systems, such as search engines and conversational agents, become integral to various domains, ensuring their transparency and explainability is crucial for accountability, fairness, and unbiased results. Explainable IR (XIR) research aims to shed light on the inner workings of these systems by developing techniques that explain model decisions in human-understandable terms. In this talk, I will present two recent studies that contribute to advancing XIR and share how we can leverage IR axioms and mechanistic interpretability methods to reverse engineer neural retrieval models, revealing how certain components (i.e., attention heads), process and prioritize relevance signals.

Bio: Catherine Chen is a 4th year PhD candidate at Brown University, supervised by Carsten Eickhoff. Her research centers on explainable information retrieval (XIR), particularly developing methods for interpretable and explainable search. She is also broadly interested in the interpretability of large language models and the applications of ML/NLP/IR to the biomedical domain. Her work aims to bridge the gap between model transparency and practical deployment, contributing to the development of more accountable and trustworthy AI systems.

Counter: SEA Talk #280.

The May edition of SEA involves Efficiency in Neural IR, with two speakers. This is a hybrid event. Note that we are meeting in a different room than usual!

Location: 904, Science Park Amsterdam, Room C0.110
The Zoom link, in case you want to join online, will be available when you RSVP.

Speaker #1: Roberto Esposito (Weaviate)
Title: How to scale similarity search for Late Interaction Models
Abstract: Late interaction models—such as ColBERT and ColPALI—have reshaped how we compute the relevance in semantic search. Unlike traditional dense retrieval models that rely on a single vector embedding per document, these models produce multi-vector representations, allowing for deeper interactions between query and document. Crucially, this approach leads to significantly improved ranking performance. However, it introduces a fundamental challenge: how do we efficiently search over multiple embeddings at scale? In this talk, we’ll explore how Weaviate adapts our vector indexing techniques—such as Hierarchical Navigable Small World (HNSW)—to support late interaction embeddings. Finally, we’ll discuss how to reduce the memory footprint of multi-vector search using vector quantization techniques like Binary Quantization (BQ), Scalar Quantization (SQ), and Product Quantization (PQ), as well as approaches to encode multi-vectors into single-vector representations for more efficient indexing
Bio: Roberto is a Research Engineer on the Applied Research team at Weaviate, where he's currently focused on multi-vector support for the Weaviate vector database. He holds a master's degree from the University of Pisa, where his research centered on Approximate Nearest Neighbor Search (ANN) and compression algorithms.

Speaker #2: Pooya Khandel (University of Amsterdam)
Title: PEIR: Modeling Performance in Neural Information Retrieval
Abstract: The efficiency of neural information retrieval methods is primarily evaluated by measuring query latency. In practice, measuring latency is highly tied to hardware configurations and requires extensive computational resources. Given the rapid introduction of retrieval models, achieving an overall comparison of their efficiency is challenging. In this paper, we introduce PEIR, a framework for hardware-independent efficiency measurements in Learned Sparse Retrieval (LSR). By employing performance modeling approaches from high-performance computing, we derive performance models for query evaluation approaches such as BlockMax-MaxScore (BMM) and propose to measure memory and/or floating-point operations while performing retrieval on input queries. We demonstrate that by using PEIR, similar conclusions on comparing the latency of retrieval models are obtained.
Bio: Pooya Khandel is a former PhD candidate at the IRLab and PCS research groups at the University of Amsterdam. His work is focused on addressing large-scale challenges in NLP and IR, particularly in the area of neural search, either through performance analysis and the application of parallel computing techniques or through data and algorithmic efficiency with sampling and distillation.

Counter: SEA Talks #281 and #282.

Details to be announced.

Details to be announced.