pyconIT.pdf - Speaker Deck

Slide 1

Slide 1 text

No content

Slide 2

Slide 2 text

How many of you use Google Gboard?

Slide 3

Slide 3 text

Do you know how Gboard models are trained?

Slide 4

Slide 4 text

👋 Luca Corbucci 👨💻 Ph.D. Student @ University of Pisa 🎤 Podcaster @ PointerPodcast 🦸 Co-Organizer @ Superhero Valley and Pisa.dev 🌏 lucacorbucci.me

Slide 5

Slide 5 text

👋 Luca Corbucci 👨💻 Ph.D. Student @ University of Pisa 🎤 Podcaster @ PointerPodcast 🦸 Co-Organizer @ Superhero Valley and Pisa.dev 🌏 lucacorbucci.me

Slide 6

Slide 6 text

How classic ML works?

Slide 7

Slide 7 text

HOW CLASSIC ML WORKS COLLECT TRAINING DATA

Slide 8

Slide 8 text

HOW CLASSIC ML WORKS TRAIN A MODEL COLLECT TRAINING DATA

Slide 9

Slide 9 text

HOW CLASSIC ML WORKS TRAIN A MODEL MAKE PREDICTIONS COLLECT TRAINING DATA WHAT’S THE COLOR OF THE CAT?

Slide 10

Slide 10 text

No content

Slide 11

Slide 11 text

TRAIN THE MODEL WITH THE RECEIVED DATA

Slide 12

Slide 12 text

PRO AND CONS

Slide 13

Slide 13 text

PRO AND CONS ✅ EASIER TO TRAIN & DEBUG

Slide 14

Slide 14 text

PRO AND CONS ✅ EASIER TO TRAIN & DEBUG ✅ MORE MATURE ECOSYSTEM

Slide 15

Slide 15 text

PRO AND CONS ✅ EASIER TO TRAIN & DEBUG ✅ MORE MATURE ECOSYSTEM ✅ GLOBAL VIEW ON THE DATASET

Slide 16

Slide 16 text

PRO AND CONS ✅ EASIER TO TRAIN & DEBUG ✅ MORE MATURE ECOSYSTEM ✅ GLOBAL VIEW ON THE DATASET ⚠ PRIVACY AND TRUST RISKS

Slide 17

Slide 17 text

PRO AND CONS ✅ EASIER TO TRAIN & DEBUG ✅ MORE MATURE ECOSYSTEM ✅ GLOBAL VIEW ON THE DATASET ⚠ GDPR AND SIMIlAR LAWS CAN LIMIT DATA SHARING ⚠ PRIVACY AND TRUST RISKS

Slide 18

Slide 18 text

PRO AND CONS ✅ EASIER TO TRAIN & DEBUG ✅ MORE MATURE ECOSYSTEM ✅ GLOBAL VIEW ON THE DATASET ⚠ SCALABILITY LIMITS ⚠ GDPR AND SIMIlAR LAWS CAN LIMIT DATA SHARING ⚠ PRIVACY AND TRUST RISKS

Slide 19

Slide 19 text

FEDERATED LEARNING

Slide 20

Slide 20 text

FEDERATED LEARNING THE IDEA WHY NOW? MORE COMPUTE AT THE EDGE RISE IN PRIVACY CONCERNS BETTER TOOLS TO IMPLEMENT FL * * * AGGREGATE THE SHARED MODELS * * MOVE THE MODELS, NOT THE DATA.

Slide 21

Slide 21 text

TRAIN A MODEL ON MY DATA TRAIN A MODEL ON MY DATA TRAIN A MODEL ON MY DATA TRAIN A MODEL ON MY DATA TRAIN A MODEL ON MY DATA TRAIN A MODEL ON MY DATA

Slide 22

Slide 22 text

TRAIN A MODEL ON MY DATA TRAIN A MODEL ON MY DATA TRAIN A MODEL ON MY DATA TRAIN A MODEL ON MY DATA TRAIN A MODEL ON MY DATA TRAIN A MODEL ON MY DATA

Slide 23

Slide 23 text

No content

Slide 24

Slide 24 text

No content

Slide 25

Slide 25 text

No content

Slide 26

Slide 26 text

No content

Slide 27

Slide 27 text

THE MODELS ARE AGGREGATED INTO A SINGLE MODEL

Slide 28

Slide 28 text

THE MODELS ARE AGGREGATED INTO A SINGLE MODEL

Slide 29

Slide 29 text

THE MODELS ARE AGGREGATED INTO A SINGLE MODEL USUALLY, THIS IS JUST AN AVERAGE OF THE WEIGHTS OF THE MODELS

Slide 30

Slide 30 text

Do you know how Gboard models are trained?

Slide 31

Slide 31 text

No content

Slide 32

Slide 32 text

No content

Slide 33

Slide 33 text

No content

Slide 34

Slide 34 text

No content

Slide 35

Slide 35 text

FL FRAMEWORKS

Slide 36

Slide 36 text

FL FRAMEWORKS AND COUNTLESS DIY IMPLEMENTATIONS ON GITHUB

Slide 37

Slide 37 text

FL FRAMEWORKS

Slide 38

Slide 38 text

WHY FLOWER?

Slide 39

Slide 39 text

WHY FLOWER?

Slide 40

Slide 40 text

WHY FLOWER?

Slide 41

Slide 41 text

WHY FLOWER?

Slide 42

Slide 42 text

WHY FLOWER? +

Slide 43

Slide 43 text

WHY FLOWER? +

Slide 44

Slide 44 text

WHY FLOWER? +

Slide 45

Slide 45 text

WHY FLOWER? +

Slide 46

Slide 46 text

WHY FLOWER? + SIMULATION DEPLOY

Slide 47

Slide 47 text

ARCHITECTURE SERVER CLIENT CLIENT CLIENT

Slide 48

Slide 48 text

ARCHITECTURE SERVER SERVER APP SUPERLINK CLIENT SUPERNODE CLIENT APP CLIENT SUPERNODE CLIENT APP CLIENT SUPERNODE CLIENT APP

Slide 49

Slide 49 text

DEMO!

Slide 50

Slide 50 text

RECAP Centralized Learning Federated Learning Data location All data is collected in one central server Data stays on the local device Privacy Higher risk Raw data never leaves device Regulations Difficult with sensitive data Ideal for sensitive or regulated environments Entry barrier Low (lots of tools, guides) Lowering

Slide 51

Slide 51 text

REAL-WORLD CONSIDERATIONS ⚡ Network Latency 🐢 Stragglers 🎲 Non-IID Data

Slide 52

Slide 52 text

RESOURCES TO GO FURTHER

Slide 53

Slide 53 text

RESOURCES TO GO FURTHER

Slide 54

Slide 54 text

CONCLUSIONS AS INTERESTS IN DATA PRIVACY GROWS, FL IS BECOMING INCREASINGLY IMPORTANT *

Slide 55

Slide 55 text

CONCLUSIONS AS INTERESTS IN DATA PRIVACY GROWS, FL IS BECOMING INCREASINGLY IMPORTANT * * EVEN LLM CAN BE TRAINED THIS WAY!

Slide 56

Slide 56 text

CONCLUSIONS

Slide 57

Slide 57 text

CONCLUSIONS AS INTERESTS IN DATA PRIVACY GROWS, FL IS BECOMING INCREASINGLY IMPORTANT * THE FUTURE OF AI IS FEDERATED! * * EVEN LLM CAN BE TRAINED THIS WAY!

Slide 58

Slide 58 text

Thank you! Questions? 👨💻 Ph.D. Student @ University of Pisa 🎤 Podcaster @ PointerPodcast 🦸 Co-Organizer @ Superhero Valley and Pisa.dev 🌏 lucacorbucci.me