![Symbols compared to Vectors
Symbols
Category - [0, 1, 0, 0, 0, 0]
Numeric - 52
Boolean - True
[0.1, 0.8, 0.34, 0.8, … 0.2]
Vectors](https://image.slidesharecdn.com/vectorsearchfordatascientists-220816183636-21da9c83/85/Vector-Search-for-Data-Scientists-pdf-17-320.jpg)
![5 Nearest Neighbors to → “Weaviate Coding Tutorial”
Content Impressions
“We have 4 Weaviate Podcast Episodes so far [ … ] how to utilize the
Weaviate Database as a Document Store in Haystack pipelines … ”
311
“We have 2 new coding tutorials on Weaviate YouTube…” 1144
“@weaviate_io Love the integration of this with the GraphQL API!” 378
“Here are some thoughts on combining Weaviate and Haystack! TLDR:
Weavaite is a great Vector Search database…”
15563
“Weaviate (@weaviate_io) is also announcing a collaboration with Jina
AI (@JinaAI_)! …”
586](https://image.slidesharecdn.com/vectorsearchfordatascientists-220816183636-21da9c83/85/Vector-Search-for-Data-Scientists-pdf-47-320.jpg)
Vector Search for Data Scientists.pdf
- 1. Powered by Vector Search for Data Scientists A Case Study with Twitter Analytics
- 2. #1 - How is my data distributed? #2 - Are there outliers in my data? #3 - Are my variables correlated with each other? Common questions in Data Science
- 3. #1 - Can we capture the semantics in vector representations? #2 - What can we learn about our data from semantic clusters? Vector Search
- 4. Social Media Clicks and Twitter Analytics
- 6. Twitter Analytics CSV Data Tweet Text Time Impressions Engagements Engagement Rate Retweets Replies Likes User Profile Clicks Url Clicks I just published “ANN Benchmarks with Etienne Dilcoker -- Weaviate Podcast #16 on Medium.. May 27th, 1:34pm 1905 50 2.6% 3 1 15 2 18 Approximate Nearest Neighbor algorithms allow us to Vector Search in massive datasets! … May 24th, 1:13pm 7182 252 3.5% 14 1 50 27 36 Feature Engineering: Contains Emoji? Character Count? Word Count? Contains “Weaviate”?
- 7. Key Takeaways: “Vector Search for Data Scientists” 1. Segmentation in Data Science 2. Vector Representations of Data 3. Vector Segmentation 4. Weaviate for Twitter Analytics 5. Research Questions and Discussion Slides, Colab Notebook, Video Presentation available on: github.com/CShorten/Vector-Search-for-Data-Scientists
- 8. Key Takeaway #1 - Segmentation in Data Science
- 9. Visualizing Distributions of Values
- 10. Segmentation in Data Science ● What Time was the Tweet sent? ● Is there a URL Link in the Tweet? ● Symbolic vs. Vector Segmentation
- 11. What Time was the Tweet sent?
- 12. Is there a URL Link in the Tweet?
- 13. Can we split Impressions based on the Semantics of the content? Weaviate Podcast Weaviate Tutorial AI Weekly Update
- 14. How can we segment analytics based on the semantics of… ● Text ● Images ● Code ● Audio ● Video ● Graph-Structure ● Biological Sequences ● … !
- 15. Summary of Takeaway #1 Segmentation in Data Science We visualize the Distribution of our data to get a sense of it. For example we see that Impressions are somewhat Normally Distributed. Is that also true for Tweets sent at 3 AM? What about Tweets related to Deep Learning for Robotics?
- 16. Key Takeaway #2 - Vector Representations of Data
- 17. Symbols compared to Vectors Symbols Category - [0, 1, 0, 0, 0, 0] Numeric - 52 Boolean - True [0.1, 0.8, 0.34, 0.8, … 0.2] Vectors
- 18. Vector Representations of Data Photo by Shayna Douglas on Unsplash 0.83 0.35 .. 0.02 Photo by Bill Stephan on Unsplash 0.74 0.01 .. 0.95
- 19. Are these puppies similar? Let’s ask Vector Distance! L2 Distance = ∑ || ai - bi ||2 L2 Distance (Puppy1, Puppy2) = (4-2)2 + (8-9)2 + (10-11)2 = 6 L2 Distance (Puppy1, Airplane) = (4-1)2 + (8-20)2 + (10-20)2 = 253 6 << 253, Puppy1 is thus much more semantically similar to Puppy2 than Airplane Vector Name Value 1 Value 2 Value 3 Puppy1 4 8 10 Puppy2 2 9 11 Airplane 1 20 20
- 20. Capturing Semantics in Vector Representations
- 21. How do Vectors represent real-world objects? 0.08 0.53 0.16 … 0.83 0.18 384 dimensional vector Does this represent how much of a “brand” this is? We aren’t sure! But there are research fields such as “Multimodal Neurons” from OpenAI, and the general field of Disentangled RepresentationLearning that are making great strides in understanding this.
- 22. Can we compress these vectors? … 384 dimensional vector Sometimes! Ideas like Binary Passage Retrieval (shown above) - fp32 to Binary values Ideas like Product Quantization - 384-d vector mapped to 32-d
- 23. Semantic Similarity with Vector Representations Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks Authored by Nils Reimers and Iryna Gurevych Published 2019
- 24. Query Point Positive and Negative Pair Sampling
- 27. Another strategy - Data2Vec, Baevski et al. 2022
- 28. Do we need to train our own models? No! There are many pre-trained models that work very well for a broad range of data!
- 29. Great place to get started: Sentence Transformers
- 30. Summary of Takeaway #2 Vector Representations of Data Data such as Images, Text, Code, … can be represented as Vectors with Deep Learning models. These models are trained to maximize semantic similarity with massive collections of data. We often do not need to train the models ourselves for particular data domains to reach reasonable performance.
- 31. Key Takeaway #3 - Vector Segmentation
- 32. ● Text ● Images ● Code ● Audio ● Video ● Graph-Structure ● Biological Sequences ● … ! We can segment analytics based on the semantics of…
- 33. Can we split Impressions based on the Semantics of the content? Weaviate Podcast Weaviate Tutorial AI Weekly Update
- 34. More Examples
- 35. House Hunting Symbols: # of bedrooms, # of bathrooms, square feet, city → With Vectors we can encode: ● Visual style ● Neighborhood structure ● Moreflexibleinterfacetodefinefeatureswithtext
- 36. e-Commerce Products Symbols: “Shoes”, “T-Shirt”, “Pants” or colors → With Vectors we can encode visual styles
- 37. Movies Symbols can differentiate between genres like “Children”, “Action”, or “Sci-Fi” → With Vectors we can encode: ● Themes ● Characters ● Storylines
- 38. Scientific Papers Symbols: “Biology”, “Machine Learning” → With Vectors we can encode ● Nuance of the ideas ● Writing style
- 39. Music Symbols can differentiate between genres like “Hip Hop”, “Dance” → With Vectors we can encode: ● Tone ● Lyrics ● Instruments
- 40. “That’s the magic of deep learning: turning meaning into vectors, then into geometric spaces, and then incrementally learning complex geometric transformations that map one space to another. All you need are spaces of sufficiently high dimensionality in order to capture the full scope of the relationships found in the original data.” - Francois Chollet, Deep Learning with Python, 2nd edition
- 41. Summary of Takeaway #3 Vector Segmentation Vector representations, also known as embeddings, enable an Interfaceto split analytics based on the Semanticsof the content. This content could be Text, Images, Code, Audio, Videos, …
- 42. Key Takeaway #4 - Weaviate for Twitter Analytics
- 43. Twitter Analytics Tweet Text Time Impressions Engagements Engagement Rate Retweets Replies Likes User Profile Clicks Url Clicks I just published “ANN Benchmarks with Etienne Dilcoker -- Weaviate Podcast #16 on Medium.. May 27th, 1:34pm 1905 50 2.6% 3 1 15 2 18 Approximate Nearest Neighbor algorithms allow us to Vector Search in massive datasets! … May 24th, 1:13pm 7182 252 3.5% 14 1 50 27 36
- 45. Cloud Data Upload There are many other ways to do this as well Google Colab Weaviate Cloud Services
- 47. 5 Nearest Neighbors to → “Weaviate Coding Tutorial” Content Impressions “We have 4 Weaviate Podcast Episodes so far [ … ] how to utilize the Weaviate Database as a Document Store in Haystack pipelines … ” 311 “We have 2 new coding tutorials on Weaviate YouTube…” 1144 “@weaviate_io Love the integration of this with the GraphQL API!” 378 “Here are some thoughts on combining Weaviate and Haystack! TLDR: Weavaite is a great Vector Search database…” 15563 “Weaviate (@weaviate_io) is also announcing a collaboration with Jina AI (@JinaAI_)! …” 586
- 49. What was the Tweet about?
- 50. Have I tweeted something like this before?
- 51. Have any Weaviate Podcast guests tweeted something like this recently?
- 56. Wikipedia Live Demo - Graph Data Model
- 58. ● Weaviate is a Vector Search Database, rather than a Library such as Facebook’s FAISS or ANNOY from Spotify ● Weaviate has a Graph-like Data Model
- 59. Expanding Twitter project with Graph Model
- 61. Summary of Takeaway #4 Weaviate for Twitter Analytics We can segment Impressions on Twitterbased on the content of the tweet without manual labeling! Weaviateis a Vector Search Databasethat can be used to store and search through semantic embeddings of data.
- 62. Key Takeaway #5 - Research Questions and Discussion
- 63. Research Questions and Discussion ● Should I fine-tune my embedding model? ● Large-Scale Vector Search with Approximate Nearest Neighbor (ANN) Algorithms ● How does Vector Search differ from Classification or Regression models?
- 64. Vector Search versus Regression on Impressions 8,530 Impressions Model Prediction
- 65. Interpretability of Vector Search Nearest Neighbors
- 66. Interpretability of Vector Search and Prediction 8,530 Impressions Model Prediction
- 67. What do we want to know about our Tweets? Should I post this? When might be a better time to post it? What might be a better phrasing of this tweet?
- 68. Expanding from individuals to teams ● Has anyone on my team tweeted something like this recently? ● Who on our team would be best fit to tell this story? ● What topics should we be tweeting about?
- 69. Summary of Takeaway #5 Research Questions and Discussion How can we improve these systems? What looks promising?
- 70. Key Takeaways: “Vector Search for Data Scientists” 1. Segmentation in Data Science 2. Vector Representations of Unstructured Data 3. Vector Segmentation 4. Weaviate Example for Twitter Analytics 5. Research Questions and Discussion Slides, Colab Notebook, Video Presentation available on: github.com/CShorten/Vector-Search-for-Data-Scientists
- 71. Connect with us! Weaviate Slack Channel YouTube: Weaviate • Vector Search Engine Weaviate Podcast Twitter @weaviate_io
- 72. Thank you for Watching! Special thanks to Sebastian Witalec in advising the development of this presentation and Svitlana Smolianova for visual styling.