Device-level AI for 5G and beyond
3 likes11,645 views
The document discusses device-level artificial intelligence (AI) for 5G networks and beyond. It describes how on-device AI can process and analyze data closer to its source, minimizing data transmission and protecting privacy while reducing latency. Examples of on-device AI applications include facial recognition and virtual assistants. The document also examines challenges of on-device AI like obtaining accurate data sets and balancing device autonomy with network impacts. It concludes that some level of device-level AI is inevitable as networks become more complex and intelligent devices are needed to help manage this complexity.
Device-level AI for 5G and beyond
- 1. Device-level AI for 5G and Beyond Yue Wang, Samsung Research UK CW TEC 2018 27th September, 2018
- 2. From smart phones to smart everything One Network 2CW TEC 2018 - The inevitable automation of Next Generation Networks
- 3. On-device AI today On-device AI: • As opposed to ‘Cloud-AI’ • Dedicated processor for AI tasks performed on the device Benefit: • Data processed and analysed closer to the data source • Minimised amount of data transmission • Consumer data privacy protected • Minimised latency • Real time analytics Applications: • Facial/voice recognition • Ad Targeting • Virtual assistant “AI and machine learning increasingly will be embedded into everyday things” - Gartner’s 2017 "On-device AI will be a big buzzword for new phones in 2018 So far, the strongest use cases are in computational photography and facial recognition” – IDC 2018 Intelligence for communications 3CW TEC 2018 - The inevitable automation of Next Generation Networks
- 4. Why? • Network: • More flexible, dynamic, and intelligent • End devices: • are connecting to an increasingly complicated network • 36.508, FDD frequency test • over 50 tables (!!!) Rel. 14 2017 vs. ~30 tables Rel. 10 2012 • The intelligence on devices: • Allow a simpler UE design • Avoid unnecessary delays and signaling overhead • Allow more flexibility of connecting Frequency bands Below and beyond 6GHz bands Carrier aggregation Access technologies Waveforms Numerology: Various subcarrier spacing Variable carrier bandwidth Variable SS block sweeping 4CW TEC 2018 - The inevitable automation of Next Generation Networks
- 5. AI in networks Core Network/Cloud NFV RCC AI AI Orchestrator AI AI AI AI AI AI Device-level AI: • RF • Power management AI Localised AI: • RAN elasticity End-to-end AI: • Slice management • Network service assurance Device-level AI Localised AI End-to-end AI Localised AI: • Flexible functional split VNFs 5CW TEC 2018 - The inevitable automation of Next Generation Networks
- 6. AI in networks Device-level AI Data is collected and stored on device – better privacy, reduce delay and no data overhead Effects to and from the network Localised AI AI applied across network domains, data needs to be passed between Data overhead Localised decision may be complimentary to end-to-end AI End-to-end AI AI applied for the end to end network, data/knowledge gathered from the different domains of the network Data challenges Deployment Green field Innovation Network architecture, policies, SLAs Protocols and signallings 6CW TEC 2018 - The inevitable automation of Next Generation Networks
- 7. An UE example – AI for cell selection Increasingly complicated procedures in cell selection and reselection in LTE and 5G • 35 parameters for system information • 10 parameters for speed dependent selection • 13 parameters for interworking • The list is getting larger: CoMP, beam sweeping • No adaptability to new technologies Increased power consumption on the UE for cell selection • Doubled power for LTE compared to 3G, RRC_IDLE -> RRC_CONNECTED • 4 times higher for LTE than 3G, RRC_CONNECTED -> RRC_IDLE Overhead Delay Power Consumption 7CW TEC 2018 - The inevitable automation of Next Generation Networks
- 8. AI for cell selection UE Actions (the selected TRP) cell selection with AI UE location, speed, measured signal strengths (RSRP/RSRQ) Feedbacks from the network UE 8CW TEC 2018 - The inevitable automation of Next Generation Networks
- 9. Benefits Current procedure Drawbacks With AI Benefits Periodically measured measurement needed even without reselection actually happening; information may be outdated Reselection is triggered Threshold based measurements Multiple factors affecting the threshold – not optimal; A massive list of parameters become unbearable with changing environment, and for different services No thresholds, less parameters and configurations Static configurations No forward compatibility – any new features developed in the radio will need either new parameters, or adding new configurations to the parameters Real-time, adapted to c hanges of the context (e.g., speed) Less overhead Faster selection Reduced power consumption 9CW TEC 2018 - The inevitable automation of Next Generation Networks
- 10. Challenges • Data • Synthesized data vs real data • Obtaining the accurate data set • Learning • The context to and from the network • Isolated AI results in sub/local optimal or even negative impacts to the network end to end • not desired by the operators • How much autonomy do you want to empower the devices? 10CW TEC 2018 - The inevitable automation of Next Generation Networks
- 11. Challenges 11 • Standard • Support both ‘legacy’ and intelligent devices • some devices may be smarter than others • Long process in standardization • leads to de-facto standard and fragmentation • Production • Device computational power, and impact on battery life – not every device needs to compete to be the most intelligent • The challenge in validation and deployment – never know what is going to happen until it is put in the real network
- 12. Future Looking and Conclusion On-device AI vs device-level AI Different levels of intelligence Network instructed device-level AI Inevitable change in the industry 1 2 3 4 12CW TEC 2018 - The inevitable automation of Next Generation Networks
- 13. “The measure of intelligence is the ability to change.” - Albert Einstein 13CW TEC 2018 - The inevitable automation of Next Generation Networks