FROM RETROSPECTIVE TO CONTINUOUS DEEP ANALYTICS Seif Haridi KTH - PowerPoint PPT Presentation

FROM RETROSPECTIVE TO CONTINUOUS DEEP ANALYTICS Seif Haridi KTH ‐ SICS

Why most Data Analysis today is Retrospective

From OLAP Databases Data Cube LOCATION TIME

To Deep Analysis on Batches Past Data Sets Map Map Map Map Map Map Reduce Reduce Reduce Reduce Reduce Reduce Past Models

And Machine Learning Pipelines Feature Extraction Yesterday’s Dataset Model Training

Today’s Retrospective Data Processing Knowledge Data STORE LOAD PROCESS STORE

Today’s Retrospective Data Processing Knowledge Data time Takes Long to Extract Knowledge

Models and patterns on older data personalized ad winter jacket discount! are often irrelevant today

Critical Decisions demand Continuous Analysis

We propose a continuous processing architecture Knowledge Data STORE LOAD PROCESS STORE time

We propose a continuous processing architecture Knowledge Data PROCESS time

We propose a deep processing architecture PROCESS arbitrarily iterative computation

We propose a fast processing architecture PROCESS TPUs

for Dynamic Graph analysis virus outbreaks social network trends

for Online Machine Learning feature learning tensor programming

and Relational Data Streaming σ θ π σ θ π σ θ σ θ dynamic tables

Unified and Optimised on a Common Representation Dynamic Online Relational Graphs ML Streams Intermediate Representation

A Runtime Designed for Continuous Deep Analysis Intermediate Representation Distributed Runtime metrics config constraints Self‐Reconfiguration

A Full Stack for Continuous Deep Analytics Dynamic Online Relational Graphs ML Streams Intermediate Representation Distributed Runtime metrics config constraints Interpretable, Online Models of the world