Neural Search Talks — Zeta Alpha

ColBERT + ColBERTv2: late interaction at a reasonable inference cost

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