Edge AI in Action: Accelerating Development across the Global South

Abstract

The session explored how moving artificial intelligence from centralized clouds to the edge can accelerate innovation and address real‑world challenges in the Global South. After an opening by ITU’s Fred Werner, two research‑focused keynotes highlighted (i) edge‑centric haptic and tactile applications at IIT Delhi and emerging standards, and (ii) federated learning as a privacy‑preserving, low‑latency strategy for telecom networks. A moderated panel then examined concrete XR deployments, practical edge‑AI roll‑outs across diverse geographies, and metrics for evaluating edge solutions. The meeting concluded with UN‑level reflections on AI governance, the importance of inclusive standards, and a call to action for continued collaboration.

Detailed Summary

• Context & Vision – Werner opened by asking whether invention itself might become the last human invention, citing a conversation with AI‑safety expert Roman Jampolski. He framed AI for Good as a long‑running UN initiative (since 2017) that has moved from hype‑driven PowerPoint concepts to concrete, standards‑driven actions.
• AI‑for‑Good Milestones – Highlighted the 2023 generative‑AI boom, the emergence of autonomous AI agents, embodied AI (robotics, brain‑computer interfaces), and “space AI computing.”
• Strategic Pillars – Emphasised three pillars of AI for Good: Solutions (machine‑learning challenges, startup pitching, edge‑AI “build‑athon” in India, TinyML competitions), Skills (AI Skills Coalition, sandbox environments for government training), and Standards (over 400 AI standards in development, especially for future networks – 5G/6G and AI‑native architectures).
• Edge‑AI Focus – Noted a recent build‑athon on edge AI in India and announced that the current session would centre on “edge AI in action in the Global South.”

2. Keynote 1 – Edge‑AI Research at IIT Delhi (Prof. Brejesh Lall & Dr. Rathinamala Vijay)

2.1 Why Edge Is Critical

• Convergence – Edge becomes indispensable due to the convergence of communication, compute, and control.
• Latency & Safety – For applications like haptics, sub‑millisecond latency is essential because delayed tactile feedback can be “catastrophic.”
• Contextual Relevance – Large foundational models lack the fine‑grained context required for many local tasks; edge devices can infuse locality into AI decisions.

2.2 Haptics as a Showcase

• Fundamentals – Haptics combines kinesthetic (force/pressure) and tactile (texture) sensing, which permeates everyday interactions (temperature, hardness, greeting gestures).
• QoE Challenges – Unlike speech or video, haptic QoE suffers dramatically from unsynchronized or delayed feedback.
• Edge Solutions – IIT Delhi pursues two strategies:
  1. Split‑Control Architecture – Moving substantial processing from the cloud to the edge, reducing round‑trip latency.
  2. Intent‑Based Signalling – Translating raw haptic data into higher‑level intent at the edge, enabling heterogeneous devices to interoperate (e.g., a camera‑based system can understand a haptic command without sharing low‑level sensor data).

2.3 Standards & Technical Reports (TSDSI & ITU)

• Dynamic AI/ML Models for V2X – Reports on self‑sustaining vehicular‑to‑everything applications.
• Security‑Enhanced Passive Digital Twins – Early work on AI‑augmented digital twin security.
• Architectural Support for Tactile Applications – Guidelines for integrating haptic streams into 6G RAN and AI‑native network slices.
• Quality‑of‑Experience Metrics – Ongoing studies to define multi‑modal QoE at the edge.

2.4 Takeaway

• Edge research at IIT Delhi is tightly linked to standardisation efforts (IMT‑2030, ITU‑T, CGPP, M2M) to ensure global interoperability, especially for emerging tactile and V2X use cases.

3. Keynote 2 – Federated Learning for Telecom Networks (Dr. Ranjitha Prasad)

3.1 Motivation

• Data Explosion – Mobile traffic is growing exponentially, especially with 6G use cases (eMBB, URLLC).
• Network Bottlenecks – Legacy architectures cannot meet sub‑10 ms latency requirements for mission‑critical closed‑loop control.

3.2 Edge‑Native AI & ORAN

• AI‑Native Network – AI is embedded directly into the Radio Access Network (RAN) via the RAN Intelligent Controller (RIC), part of the O‑RAN Alliance.
• MEC (Multi‑Access Edge Computing) – Provides a hierarchical stack: Core ↔ RAN ↔ User Equipment (UE), enabling inference and limited training at the edge.

3.3 Federated Learning (FL) as Enabler

• Privacy‑Preserving – “Bring code to the data”; raw user logs never leave the device. Only model updates (metadata) are shared.
• Latency & Bandwidth Reduction – Local training removes the need for massive upstream bandwidth.
• Personalisation – Edge servers can tailor lightweight models for local conditions while the global model continues to train centrally.

3.4 Illustrative Use Cases

Use Case	Description	Edge Role
Traffic‑Prediction for Large Events (France)	Base stations locally forecast traffic spikes during a stadium football match; edge nodes aggregate locally‑derived forecasts for the core to re‑allocate resources.	Real‑time data collection & preliminary inference at edge.
V2X Safety Information	Vehicles exchange road‑condition and incident data; each car runs a local FL client that contributes to a global safety model.	Edge servers on‑board each vehicle; global model updated via the cloud.
Security‑Incident Detection (IIIT‑Delhi)	A federated client detects anomalous security events; the edge aggregates alerts to inform the whole network.	Edge‑level anomaly detection, minimal data exposure.

3.5 Programme & Partnerships

• IntelliCom Lab (IIIT‑Delhi) + IIT Delhi – Joint research funded by METI and the “IndieI” initiative.
• Publications & Demos – Several papers on FL for telecom security, to be presented at the conference exhibit (Hall 2, METI Pavilion).

3.6 Takeaway

• FL provides a practical pathway to embed AI at the edge for telecom, balancing privacy, latency, and scalability while supporting the AI‑for‑Good agenda.

4. Panel Discussion – “Demystifying Edge AI”

Moderator: Frederic Werner (ITU)

4.1 Panelists

Panelist	Affiliation	Expertise
Mala	ARTPARK (Center of Excellence – Wired & Wireless Technologies)	XR applications, private 5G, AI‑enabled medical emergency care
Alagan Mahalingam	Rootcode (Founder, CEO)	Edge‑AI deployments across 27 countries, hardware‑software co‑design, agricultural advisory
Shakshi Gupta	Qualcomm (Global Government Affairs)	Edge AI hardware, Tech‑for‑Good programme, on‑device AI for health & automotive

4.2 Discussion Topics

4.2.1 XR Applications (Mala)

• XR‑Assisted Facility Tours – Private 5G network inside ARTPARK’s “Art Garage” allows visitors to explore startups via XR headsets, with multilingual immersive overlays.
• XR‑Assisted Emergency Care – Public 5G connects first‑responders wearing XR glasses, IoT wearables, and AED kits; live vitals are over‑laid on video streams for remote medical experts to guide CPR/AED use.
• Scalability & Trade‑offs
  • Private 5G enables low‑latency, on‑premise edge AI for industrial/educational settings.
  • Public 5G offers broader coverage for life‑saving emergency scenarios but may face higher contention.
  • Cost – Private networks require upfront CAPEX; public networks rely on telecom‑operator tariffs.
  • Security – Private networks are easier to isolate; public deployments need robust encryption and access‑control.

4.2.2 Edge‑AI Deployments Across Geographies (Alagan)

• Global Footprint – Rootcode’s solutions serve 92 million users in 27 countries (e.g., Estonia, Portugal, Sri Lanka).
• Case Study: Precision Agriculture in Portugal – Sensors in soil + mobile‑app image analysis → AI model predicts crop health.
• Edge Adaptation in Sri Lanka – In villages lacking reliable connectivity, a Raspberry Pi edge node runs quantised models (Gemma‑V, custom 2‑D CNNs) to deliver the same advisory service offline.
• Edge in Rural Healthcare (USA) – Remote patient monitoring device streams vitals to an edge server; alerts generated locally without cloud dependence.
• Design Principle – Task‑first, model‑second: Define the concrete problem, then distil or prune the AI model to fit the target edge hardware (quantisation, pruning, knowledge‑distillation).

4.2.3 Metrics & Evaluation of Edge AI (Shakshi)

• Availability – Edge AI now runs on smartphones (10‑billion‑parameter on‑device models), cars, and IoT devices; Qualcomm’s “Tech for Good” programme showcases startups using Qualcomm AI‑on‑Device chips.
• Key Evaluation Dimensions
  1. Latency (sub‑10 ms for mission‑critical loops)
  2. Power Consumption (critical for low‑resource regions)
  3. Privacy (data stays on‑device; only model updates leave)
  4. Bandwidth Savings (reduced upstream traffic)
  5. Scalability & Interoperability (aligned with emerging standards – IMT‑2030, O‑RAN)
• Showcase: Raksa Health (India) – On‑device AI health assistant that works offline, stores patient records locally, and offers prescription lookup without internet.

4.2.4 Audience Interaction

• Panelists reiterated the need for regional standards to avoid fragmented solutions and emphasized that open‑source AI models (hosted via ITU’s AI‑for‑Good sandbox) can accelerate replication across borders.

4.3 Panel Conclusions

• Edge AI is not a luxury for the Global South; it is a necessity for connectivity‑constrained, power‑limited contexts.
• Local‑first model design (task‑driven, model‑size optimisation) yields solutions that are portable, affordable, and privacy‑preserving.
• Standardisation, open data, and cross‑regional collaboration are essential to scale successes.

5. Closing Remarks – UN AI Governance & Global Dialogue

5.1 Reflections (UN Representative, New York)

• Edge‑AI Definition – “AI closer to where things happen: people, services, communities.”
• Development Impact – Improves speed, reduces cost, and safeguards privacy where connectivity is limited.

5.2 Three Strategic Messages

1. Human‑Centred AI – Technology must protect and empower people; all examples reaffirm this principle.
2. Closing the Global AI Gap – Replicating successful models across countries requires decisive support and knowledge‑sharing.
3. Avoid Fragmentation – Coordinated, inclusive approaches across national and regional initiatives are needed to prevent a splintered AI ecosystem.

5.3 Global Dialogue on AI Governance (Amb. Rain Tamzar)

• Mandate – First UN‑level Global Dialogue on AI Governance (July 2024, Geneva) will gather governments and multi‑stakeholder groups to produce actionable outcomes.
• Core Themes – Trust, transparency, interoperability, equal participation, human‑rights grounding, and practical deliverables (not endless theory).
• Process – Dialogue will run parallel to the ITU AI‑for‑Good summit, building on member‑state inputs, stakeholder wisdom, and the International Scientific Panel.

5.4 Closing Ceremony

• Fred Werner presented “momentos” to panelists and organisers. The session formally closed after a brief thank‑you round.

Key Takeaways

• Edge AI is essential for the Global South – it overcomes connectivity gaps, lowers latency, saves bandwidth, and preserves privacy.
• AI‑for‑Good’s three pillars (Solutions, Skills, Standards) provide a concrete framework to move from ideas to deployed, interoperable technologies.
• Haptic and tactile edge applications demand sub‑millisecond latency; split‑control and intent‑based signalling are viable architectural patterns.
• Federated Learning enables privacy‑preserving AI at the edge of telecom networks, delivering sub‑10 ms closed‑loop control for 6G use cases.
• Real‑world deployments (precision agriculture in Portugal, offline advisory in Sri Lanka, XR‑assisted emergency care) demonstrate that task‑first, model‑second design yields portable, low‑cost solutions.
• Standardisation work (TSDSI, ITU, O‑RAN, IMT‑2030) is already addressing edge‑specific requirements such as V2X AI models, tactile QoE, and AI‑native network slices.
• Metrics for evaluating edge AI should include latency, power consumption, privacy guarantees, bandwidth savings, and compliance with emerging standards.
• Open‑source AI models and sandbox environments (ITU AI‑for‑Good sandbox) are critical for rapid replication and testing across regions.
• The upcoming UN Global Dialogue on AI Governance will aim to embed these technical advances within a human‑rights‑based, inclusive policy framework, avoiding fragmented national approaches.
• Collaboration between academia, industry, and international bodies (e.g., ITU, Qualcomm, Rootcode, ARTPARK) is already producing scalable edge‑AI solutions; continued multi‑stakeholder engagement is the path forward.

---

See Also:

• aligning-ai-governance-across-the-technology-stack
• scaling-ai-solutions-through-southsouth-collaboration
• building-a-trusted-and-resilient-ai-infrastructure-ecosystem-balancing-innovation-security-and-rights
• ai-innovators-exchange-accelerating-innovation-through-startup-and-industry-synergy
• towards-a-safer-south-launch-of-the-global-south-network-on-ai-safety-and-evaluation
• ai-for-inclusive-economic-progress-the-public-services-ai-stack
• responsible-ai-at-scale-governance-integrity-and-cyber-readiness-for-a-changing-world
• democratizing-ai-resources-and-building-inclusive-ai-solutions-for-india
• inclusion-for-social-empowerment