A Future Look of Data Stream Processing as an Architecture for AI
0 likes182 views
The document discusses the components and capabilities of a proposed system for achieving human-level artificial intelligence, focusing on the neurobiological foundations and features of the human brain that could inform its design. It explores how compilers can optimize instructions for diverse workloads and emphasizes the integration of data stream processing and advanced analytics for complex decision-making tasks. The architecture includes a neural network runtime that supports parallel processing, low latency, and incremental updates, aiming to leverage these principles for solving problems in fields like climate science.
![!11
▪ It will be able to solve complex Climate Science problems, fast
val rawStreams = streams("models/*/ts*.nc").
withType[LabelledTensor[Inf x Int x Int -> Double,
Float x (Float, Float) x (Float, Float)]].
dimensionLabels('time x 'lat x 'lon);
val averageStreams = rawStreams.map { raw =>
val timeSliced = raw.sliceBy('time);
val aligned = timeSlices.tile(360 x 720).
map(grid => average(grid));
val gridSlices = aligned.sliceBy('lat, 'lon);
val agg12h = gridSlices.window('time, t => t.between(TimeOfDay(6.h), TimeOfDay(18.h))).
average;
val agg1d = gridSlices.window('time, t => Day(t)).average;
val agg1month = gridSlices.window('time, t => Month(t)).average;
val agg1Season = gridSlices.window('time, t => Month(t).in(
Set(Dec, Jan, Feb),
Set(Mar, Apr, May),
Set(Jun, Jul, Aug),
Set(Sep, Oct, Nov)).average;
(agg12h, agg1d, agg1month, agg1season)
}.unzip4;
val diffs = averageStreams.map { inv =>
val merged = inv.mergeOn('time, 'lat, 'lon);
val averageModels = merged.map(models => (models, average(models)));
averageModels.map {
case (models, avg) => models.map(t => t-avg)
};
}](https://image.slidesharecdn.com/csl18-paris-180619153309/85/A-Future-Look-of-Data-Stream-Processing-as-an-Architecture-for-AI-11-320.jpg)
A Future Look of Data Stream Processing as an Architecture for AI
- 1. !1 Full Human-Level Artificial Intelligence and … Data Stream Processing Heh, I got this! Paris Carbone
- 2. ▪ There is one known runtime for human-level intelligence
- 3. ▪ What is so special about the human brain structure? !3 Neurobiological Foundations of Action Planning and Execution - Human Action Control — B.Hommel et al. ▪ Diverse functionality/workloads ▪ Common runtime (neurons)
- 4. ▪ The Brain Neural Network Runtime !4 ▪ Distributed ▪ Organised in Logical Units ▪ Embedded State with Computation ▪ Shared Network ▪ Configured Data Dependencies ▪ Messages (signals) ▪ Supports Low latency Serving ▪ Supports Incremental Updates ▪ Supports Iterative Tasks ▪ Asynchronous Processing ▪ 100% Organic
- 5. ▪ Distributed ▪ Organised in Logical Units ▪ Embedded State with Computation ▪ Shared Network ▪ Configured Data Dependencies ▪ Messages (signals) ▪ Supports Low latency Serving ▪ Supports Incremental Updates ▪ Supports Iterative Tasks ▪ Asynchronous Processing ▪ 100% Organic ▪ The Data Stream Processing Runtime !5
- 6. !6 ▪ Compilers - Our first and best “super-human” invention ▪ Instead, compilers can understand instructions… ▪ explained by humans in a high-level declarative language ▪ and then optimise them ▪ and translate to stupid machines to execute them reliably “A revolutionary technology that does NOT require you to throw tons of data to your problem to be able to solve it”
- 7. !7 ▪ Our ‘Continuous Deep Analytics’ Project Compilers + Data Streams
- 8. ▪ Modern Data Pipelines need to combine diverse workloads! (ML Training & Serving, Relational Algebra, Streams, Tensors, Graphs) !8 ⋈ ⋈ ⋈ σθ σθ σθ σθ π π Relational Data Streams Feature Learning Tensor Programming Dynamic Graphs
- 9. !9 Arc Compiler ▪ diverse workloads ▪ common runtime
- 10. !10 Intelligence: Smart Choice / Responce Time Pipeline (CPU) - Optimised Pipeline (GPU/TPU) - Optimised time until decision Pipeline (CPU) Pipeline (GPU/TPU) critical decision making
- 11. !11 ▪ It will be able to solve complex Climate Science problems, fast val rawStreams = streams("models/*/ts*.nc"). withType[LabelledTensor[Inf x Int x Int -> Double, Float x (Float, Float) x (Float, Float)]]. dimensionLabels('time x 'lat x 'lon); val averageStreams = rawStreams.map { raw => val timeSliced = raw.sliceBy('time); val aligned = timeSlices.tile(360 x 720). map(grid => average(grid)); val gridSlices = aligned.sliceBy('lat, 'lon); val agg12h = gridSlices.window('time, t => t.between(TimeOfDay(6.h), TimeOfDay(18.h))). average; val agg1d = gridSlices.window('time, t => Day(t)).average; val agg1month = gridSlices.window('time, t => Month(t)).average; val agg1Season = gridSlices.window('time, t => Month(t).in( Set(Dec, Jan, Feb), Set(Mar, Apr, May), Set(Jun, Jul, Aug), Set(Sep, Oct, Nov)).average; (agg12h, agg1d, agg1month, agg1season) }.unzip4; val diffs = averageStreams.map { inv => val merged = inv.mergeOn('time, 'lat, 'lon); val averageModels = merged.map(models => (models, average(models))); averageModels.map { case (models, avg) => models.map(t => t-avg) }; }
- 12. !12 equi-join time slices then map: average then diff sink: 12h sink: 1d sink: month sink: season src20 window: 12h aggregate with shared tree of partials: average window: 1d window: month window: season src1 tile map: average window: 12h aggregate with shared tree of partials: average window: 1d window: month window: season equi-join time slices then map: average then diff equi-join time slices then map: average then diff equi-join time slices then map: average then diff ▪ And generate an optimised stream process graph (program)
- 13. !13 Using an Intermediate Representation (IR) f f’…. ….Data knowledge f+f’ IR IR IR f f’
- 14. !14 Weld IR (Stanford DAWN Project) + supports large number of existing libraries - currently limited to short-lived local task execution Matei Zaharia (Spark architect) et. al. !14
- 15. The Arc Compilation Stack Available Resources Stream Metadata Intermediate Representation (IR) Frontends Logically Optimised IR Physically Optimised IR Binaries Arc: Weld for Streams
- 16. !16 JIT - Live Rewiring of Continuous Programs Physically Optimised IR Binaries Change in Resources Change in Load Distribution Monitoring Discovered better Plan
- 17. !17 The Current CDA Team (RISE SICS + KTH) Computer Systems Machine Learning Lars Kroll Paris Carbone Christian Schulte Seif Haridi Theodore Vasiloudis Daniel Gillblad MSc Students • Klas Segeljakt • Oscar Bjuhr • Johan Mickos
- 18. ▪ The Brain Neural Network Runtime !18 ▪ Distributed ▪ Organised in Logical Units ▪ Embedded State with Computation ▪ Shared Network ▪ Configured Data Dependencies ▪ Messages (signals) ▪ Supports Low latency Serving ▪ Supports Incremental Updates ▪ Supports Iterative Tasks ▪ Asynchronous Processing ▪ 100% Organic ▪ Just in Time Reconfiguration ▪ Executes Declarative Instructions Reliably