1. A Berkeley View of Systems Challenges for AI
강한결 (정보디스플레이학과)
김수근 (정보디스플레이학과)
조연수 (정보디스플레이학과)
진건기 (관광학부)

2. Today’s AI (Artificial Intelligence)

3. Trends and Challenges, Research Opportunities
Mission-critical AI
Acting in dynamic environments
R1. Continual learning
R2. Robust decisions
R3. Explainable decisions
Personalized AI
Secure AI
R4. Secure Enclaves
R5. Adversarial learning
R6. Shared learning on confidential data
AI across organizations
AI specific architectures
AI demands outpacing Moore’s law
R7. Domain specific Hardware
R8. Composable AI systems
R9. Cloud-edge systems

4. Trends and Challenges, Research Opportunities
Mission-critical AI
Acting in dynamic environments
R1. Continual learning
R2. Robust decisions
R3. Explainable decisions
Personalized AI
Secure AI
R4. Secure Enclaves
R5. Adversarial learning
R6. Shared learning on confidential data
AI across organizations
AI specific architectures
AI demands outpacing Moore’s law
R7. Domain specific Hardware
R8. Composable AI systems
R9. Cloud-edge systems

5. Trends and Challenges, Research Opportunities
Mission-critical AI
Acting in dynamic environments
R1. Continual learning
R2. Robust decisions
R3. Explainable decisions
Personalized AI
Secure AI
R4. Secure Enclaves
R5. Adversarial learning
R6. Shared learning on confidential data
AI across organizations
AI specific architectures
AI demands outpacing Moore’s law
R7. Domain specific Hardware
R8. Composable AI systems
R9. Cloud-edge systems

6. Trends and Challenges, Research Opportunities
Mission-critical AI
Acting in dynamic environments
R1. Continual learning
R2. Robust decisions
R3. Explainable decisions
Personalized AI
Secure AI
R4. Secure Enclaves
R5. Adversarial learning
R6. Shared learning on confidential data
AI across organizations
AI specific architectures
AI demands outpacing Moore’s law
R7. Domain specific Hardware
R8. Composable AI systems
R9. Cloud-edge systems

7. Three topics of research opportunities
Acting in dynamic environments
Secure AI
AI-specific architectures

8. Topic 1: Acting in dynamic environments
Many future AI applications will operate in dynamic environments that change often rapidly and unexpectedly and often in non-reproducible ways
1) Continual learning
2) Robust decisions
3) Explainable decisions

9. Systems that continually learn and adapt to asynchronous changes
R1: Continual learning
Most of today’s AI systems perform training offline and then make predictions online
This makes them a poor fit for environments that change continually and expectedly
Systems that continually learn and adapt to asynchronous changes

10. Systems that track data provenance
R2: Robust decisions
As AI applications are increasingly making decisions on behalf of humans they need to be robust to uncertainty and errors in inputs and feedback
Increasingly, learning systems leverage data collected from unreliable sources, possibly with inaccurate labels
Systems that track data provenance

11. R3: Explainable decisions
AI systems will often need to provide explanations for their decisions
Systems which is able to record and faithfully reply the computations that led to a particular decision

12. Attacker compromising the integrity of the decision process
Topic 2: Secure AI
Attacker compromising the integrity of the decision process
Attacker learning the confidential data on which an AI system was trained on, or learning the secret model
1) Secure enclaves
2) Adversarial learning
3) Shared learning on confidential data

13. R4: Secure enclaves
Enclave
Code base
Other codes

14. R5: Adversarial learning
Evasion attack
Data poisoning attack

15. R6: Shared learning on confidential data
Fraud
detection
. . .
Bank 1
Bank 2
Bank 3
Share data
Predict flu
outbreaks
. . .
Hospital 1
Hospital 2
Hospital 3
Share data

16. Topic 3: AI-specific architectures
INNOVATION OF
System & Architecture
AI demands
improve the performance
simplify the development of the next generation of AI applications

17. R7: Domain specific hardware.
data continues to grow exponentially
LIMITATION!!
NEW WAY
developing domain-specific processors
Innovations in computer architecture instead of semiconductor process improvements.

18. R8: Composable AI systems
Today’s monolithic AI system
MODULARITY
COMPOSITION
Key to increasing development speed and adoption of AI

19. edge devices are extremely heterogeneous
R9: Cloud-edge systems
expect a rapid increase in AI systems that span edge devices and the cloud
developing cloud and the cloud-edge systems is challenging for several reasons.
a large discrepancy btw the capabilities of edge devices and datacenter servers
edge devices are extremely heterogeneous
the hardware and software update cycles of edge devices are significantly slower
increase in the storage capacity <<< the growth in the data being generated

20. R9: Cloud-edge systems
Two general approaches to addressing the mix of cloud and edge devices.
to repurpose code to multiple heterogeneous platforms via retargetable software design and compiler technology.
to design AI systems that are well suited to partitioned execution across the cloud and the edge.

21. EOD