

ColBERT + ColBERTv2: late interaction at a reasonable inference cost
7 snips Aug 16, 2022
Chapters
Transcript
Episode notes
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Introduction
00:00 • 3min
Colbert's Methods for Neural Retrieval
02:52 • 3min
The Similarity Matrix and Late Interaction
06:11 • 2min
Colbert's Architecture for Re-Rating Documents
07:56 • 2min
The Problem With Quantization and Dimensionality Reduction
10:13 • 2min
The Importance of Hard Negative Mining
12:14 • 2min
Interaction Base and Representation Based Winter Interaction
14:33 • 3min
The MaxSim Operation and the Difference in Performance
17:43 • 2min
The Importance of Confidence in the Average Pooling Method
20:12 • 2min
The Importance of Colbert in the IR Community
22:14 • 2min
The Differences Between Dance Retrieval and Crossing Coders
24:28 • 2min
Colbert V2: A New Way to Train Colbert
26:17 • 2min
Densfertil's Knowledge Distillation Technique
28:08 • 3min
How to Find Tokens in a Document Collection
30:46 • 2min
Hertz Performance Drop
32:34 • 2min
Benchmarking on LATA Data Sets
34:26 • 2min
MS Marko Results: A Test in Domain Performance and Beer
35:58 • 2min
The Importance of Distillation in Model Performance
37:41 • 2min
The Differences Between Played and Splayed Treble Benchmarks
39:35 • 2min
The Future of Dense Retrieval Models
41:27 • 3min
The Importance of Similarity Scores in Query Models
44:00 • 2min
The Future of Neural Information Retrieval
46:09 • 3min
The Future of Dense Retrieval
49:23 • 3min
Colbert V3: A Deep Learning Perspective
52:42 • 2min
The Painful Cost of Storage Efficiency in Academic Research
55:05 • 2min