Loading...
Community presentation
## Topic: **Generative AI at the Edge: Efficient Intelligence for Resource-Constrained Devices** --- ## Slide 1 — Title Slide ### Title **GENERATIVE AI AT THE EDGE** ### Subtitl…
Create your own version
Use this visual direction as inspiration for a new Presenton deck.
Use this designCreator prompt
## Topic: **Generative AI at the Edge: Efficient Intelligence for Resource-Constrained Devices** --- ## Slide 1 — Title Slide ### Title **GENERATIVE AI AT THE EDGE** ### Subtitle **Efficient Deployment of Generative AI on Resource-Constrained Edge Devices** ### Content * CSQ 413 Seminar * Semester 7 * Department of Computer Science and Engineering * Nirmala College of Engineering Technology and Management ### Speaker Notes Introduce the topic as the convergence of Generative AI, edge computing, IoT, and on-device intelligence. ### Visuals/Charts * Illustration of Cloud → Edge Server → IoT Devices * AI chip / edge computing background image --- ## Slide 2 — Contents ### Title **CONTENTS** ### Content 1. Abstract 2. Introduction & Problem Statement 3. Core Architecture 4. Explanation / Workflow 5. Results & Performance 6. Applications 7. Key Approaches for GenAI at the Edge 8. Benefits & Challenges 9. Future Perspectives 10. Conclusion 11. References ### Speaker Notes Explain that the presentation moves from the motivation for edge GenAI to architectures, optimization techniques, experimental results, applications, and future directions. ### Visuals/Charts * Simple roadmap diagram of the presentation --- ## Slide 3 — Abstract ### Title **ABSTRACT** ### Content **Objective:** * To examine how Generative AI can be deployed directly on edge and resource-constrained devices. **Key Points:** * Cloud-based GenAI can introduce latency, privacy risks, and network dependency. * Edge deployment enables localized and near-real-time AI processing. * Resource limitations require efficient models, compression, hardware optimization, and intelligent deployment frameworks. * Recent studies demonstrate the feasibility of deploying lightweight generative models on devices such as Raspberry Pi and ESP32. **Source:** Navardi et al. (2025); Wang et al. (2025) ### Speaker Notes Emphasize that the goal is not simply to move large AI models from cloud servers to small devices. The central research challenge is making GenAI efficient enough to operate under limited computation, memory, storage, and energy. ### Visuals/Charts * Cloud AI versus Edge AI comparison diagram --- ## Slide 4 — Introduction & Problem Statement ### Title **INTRODUCTION & PROBLEM STATEMENT** ### Content **The Problem:** * Generative AI models require substantial computational and memory resources. * Traditional deployment relies heavily on centralized cloud infrastructure. * Cloud-based inference can cause: * Higher latency * Privacy concerns * Network dependency * Increased communication overhead **The Vision:** * Bring intelligent generative capabilities closer to where data is generated. **Solution:** * Deploy optimized Generative AI models directly on edge devices. ### Speaker Notes Explain that IoT devices continuously generate data, but sending every request to the cloud is inefficient for real-time applications. ### Visuals/Charts **Cloud Processing** IoT Device → Network → Cloud AI → Response ### Citation Navardi et al. (2025); iGenEdge (2025) --- ## Slide 5 — Introduction & Problem Statement ### Title **WHY GENERATIVE AI AT THE EDGE?** ### Content Edge deployment provides: * **Low Latency** AI inference occurs closer to the data source. * **Improved Privacy** Sensitive data can remain within the local environment. * **Reduced Bandwidth Usage** Less raw data needs to be transmitted to centralized servers. * **Improved Reliability** Local AI services can reduce dependence on continuous cloud connectivity. ### Speaker Notes Clarify that edge AI does not completely replace cloud computing. Instead, cloud and edge systems can work together depending on resource requirements. ### Visuals/Charts Two-column comparison: | Cloud AI | Edge AI | | ---------------------------- | -------------------- | | Centralized processing | Local processing | | Network dependency | Reduced dependency | | Higher communication latency | Lower latency | | Data transmitted remotely | Better data locality | ### Citation Wang et al. (2025) --- ## Slide 6 — Introduction & Problem Statement ### Title **WHAT IS GENERATIVE AI AT THE EDGE?** ### Content **Generative AI at the Edge** refers to deploying generative models on or near data-generating devices to perform local inference and intelligent processing. ### Key Characteristics * Real-time processing * Resource-constrained operation * Local data processing * Improved privacy * Reduced dependence on cloud infrastructure ### Examples of Edge Platforms * Smartphones * Raspberry Pi devices * IoT sensors * Autonomous systems * Edge servers ### Speaker Notes Define the concept simply before introducing the technical architectures. ### Visuals/Charts Device icons connected to a local edge AI node. ### Citation Wang et al. (2025); Navardi et al. (2025) --- ## Slide 7 — Core Diagrams ### Title **CORE ARCHITECTURE** ### Content **General GenAI Edge Architecture** 1. **IoT / Edge Device Layer** * Generates data and AI requests. 2. **Edge Computing Layer** * Hosts optimized AI models. * Performs local inference. 3. **Collaborative Edge Layer** * Shares workloads and resources among edge nodes. 4. **Cloud Layer** * Model training * Large-scale storage * Model distribution ### Speaker Notes Explain that heavy training can remain in the cloud, while optimized models are distributed to edge devices for inference. ### Visuals/Charts ``` CLOUD Model Training & Storage ↓ EDGE COMPUTING LAYER Model Deployment / Inference ↓ IoT & EDGE DEVICES Sensors / Users / Applications ``` ### Citation iGenEdge (2025) --- ## Slide 8 — Core Diagrams ### Title **COLLABORATIVE EDGE INTELLIGENCE** ### Content **How collaborative edge systems work:** * Multiple edge nodes provide computing resources. * AI services can be dynamically placed near high-demand users. * Requests can be distributed across available edge nodes. * Cloud connectivity supports model storage and dissemination. * Collaboration improves resource utilization under fluctuating workloads. ### Speaker Notes Use iGenEdge as the primary example of collaborative deployment. ### Visuals/Charts ``` CLOUD ↓ ┌─────────────────┐ │ Edge Node 1 │ │ ↕ │ │ Edge Node 2 │ │ ↕ │ │ Edge Node 3 │ └─────────────────┘ ↓ ↓ ↓ IoT Devices ``` ### Citation iGenEdge (2025) --- ## Slide 9 — Explanation ### Title **PHASE I: MODEL PREPARATION & OPTIMIZATION** ### Content Before deployment, GenAI models must be optimized for edge hardware. ### Main Techniques * **Quantization** * Reduces numerical precision and memory requirements. * **Pruning** * Removes unnecessary model components. * **Knowledge Distillation** * Transfers knowledge from large models to smaller models. * **Neural Architecture Search (NAS)** * Helps design efficient architectures. ### Speaker Notes Explain that model optimization is necessary because original foundation models are often too large for edge hardware. ### Visuals/Charts ``` Large GenAI Model ↓ Optimization ┌─────┼─────┐ Quantization Pruning Distillation ↓ Optimized Edge Model ``` ### Citation Navardi et al. (2025) --- ## Slide 10 — Explanation ### Title **PHASE II: EDGE DEPLOYMENT & INFERENCE** ### Content After optimization: * The AI model is deployed to edge devices or edge servers. * Incoming requests are processed locally. * The system monitors: * Latency * Memory usage * CPU utilization * Throughput * Models can be dynamically placed based on demand and available resources. ### Speaker Notes Explain that deployment is not static. Frameworks such as iGenEdge dynamically manage services according to incoming workloads. ### Visuals/Charts ``` IoT Request ↓ Edge Resource Monitor ↓ Model Selection ↓ Local AI Inference ↓ Generated Response ``` ### Citation iGenEdge (2025) --- ## Slide 11 — Explanation ### Title **PHASE III: COLLABORATION & ADAPTATION** ### Content When one edge device has insufficient resources: * Requests can be redirected to collaborative edge nodes. * Models can be placed closer to high-demand devices. * Edge nodes balance workloads dynamically. * Federated approaches enable collaborative learning without centralizing raw data. ### Example: FedGen-Sensor * Edge sensors perform local reconstruction. * Raspberry Pi servers aggregate federated model updates. * Raw sensor data remains distributed. ### Speaker Notes Explain the difference between collaborative inference and federated learning. ### Visuals/Charts ``` Sensor Nodes ↓ ↓ ↓ Local AI Processing ↓ ↓ ↓ Federated Aggregation ↓ Updated Global Model ``` ### Citation Pei et al. (2026) --- # Slide 12 — Results ### Title **RESULTS: GENERATIVE AI ON EDGE DEVICES** ### Content Benchmarking on Raspberry Pi 5 devices demonstrated major performance differences based on model size. ### Key Observation Smaller models generally provide: * Lower latency * Higher throughput * Reduced memory requirements ### Example Results * Yi 1.48B: **47 s average latency, 12 tokens/s** * Llama3 3.21B: **87 s average latency, 4.8 tokens/s** * Mistral 7.25B: **220 s average latency, 0.3 tokens/s** ### Key Finding **Model size strongly affects the feasibility of local edge inference.** ### Speaker Notes Explain that this benchmarking data demonstrates the trade-off between model capability and resource efficiency. ### Visuals/Charts Recommended: **Bar chart comparing Model Size vs Latency** ### Citation Nezami et al. (2025) --- ## Slide 13 — Results ### Title **RESULTS: FEDGEN-SENSOR PERFORMANCE** ### Content The FedGen-Sensor framework demonstrated: * Reconstruction fidelity comparable to FP32 FedAvg baseline: * **RMSE: 0.086 vs 0.082** * **5.5× reduction** in communication overhead * **2.5× reduction** in average edge inference latency ### Key Result Lightweight federated generative models can provide efficient and privacy-preserving AI directly on low-power sensor systems. ### Speaker Notes Highlight that the performance improvement comes from hardware-software co-design, quantization, and sparse update transmission. ### Visuals/Charts Recommended comparison chart: | Metric | Baseline | FedGen-Sensor | | ------------------- | -------- | -------------- | | Reconstruction RMSE | 0.082 | 0.086 | | Communication | 1× | 5.5× reduction | | Inference latency | 1× | 2.5× faster | ### Citation Pei et al. (2026) --- ## Slide 14 — Applications ### Title **APPLICATIONS: SMART IoT & SENSOR NETWORKS** ### Content GenAI at the edge can support: * Real-time sensor data reconstruction * Fault recovery * Missing data recovery * Industrial monitoring * Smart cities * Environmental sensing * Wearable devices ### Example FedGen-Sensor reconstructs missing sensor data directly at the edge without transmitting raw sensor streams to a centralized cloud. ### Speaker Notes Explain how local reconstruction is important when communication is unstable or sensor failures create gaps. ### Visuals/Charts * Smart factory * Environmental sensor network * ESP32 sensor nodes ### Citation Pei et al. (2026) --- ## Slide 15 — Applications ### Title **APPLICATIONS: INTELLIGENT IoT SERVICES** ### Content Edge GenAI can support: * Navigation assistance * Situational awareness * Predictive analytics * Weather services * Natural language interaction * Intelligent decision support ### Why Edge? * Faster responses * Reduced network latency * Localized service delivery ### Speaker Notes Explain that many future IoT applications require immediate AI responses and cannot always wait for cloud communication. ### Visuals/Charts * Smart vehicle * Smart home * IoT device ecosystem ### Citation iGenEdge (2025) --- ## Slide 16 — Applications ### Title **APPLICATIONS: PRIVACY-SENSITIVE SYSTEMS** ### Content On-device and edge AI is useful when sensitive data should remain local. ### Application Areas * Healthcare monitoring * Wearable devices * Smart homes * Medical devices * Autonomous systems ### Key Advantage Local processing reduces the need to transmit sensitive information to centralized cloud systems. ### Speaker Notes Mention that privacy is one of the strongest motivations for on-device AI, but it does not automatically solve every security issue. ### Visuals/Charts * Wearable health device * Smart home * Privacy lock symbol ### Citation Wang et al. (2025) --- ## Slide 17 — Applications ### Title **APPLICATIONS: GENERATIVE AI SERVICES ON EDGE DEVICES** ### Content GenAI models can be deployed as services across collaborative edge systems. ### Supported Services * Language-based AI services * Text generation * Image and multimodal processing * Local decision support * Generative sensor intelligence ### Deployment Concept ``` IoT Device ↓ Edge GenAI Service ↓ Generated Output ``` ### Speaker Notes Explain that future edge systems may provide AI models as localized services rather than requiring every device to host a complete model. ### Citation Navardi et al. (2025); iGenEdge (2025) --- # Slide 18 — Key Approaches for GenAI at the Edge ### Title **KEY APPROACHES FOR GENAI AT THE EDGE** ### Content ### 1. Software Optimization * Quantization * Pruning * Knowledge distillation * Neural Architecture Search ### 2. Hardware Optimization * Specialized AI accelerators * Hardware-aware model design * Efficient memory management ### 3. Deployment Frameworks * Dynamic model placement * Collaborative edge computing * Federated learning * Resource-aware orchestration ### Speaker Notes This slide replaces the original "Types of NeRF" slide with the three major categories identified in the GenAI-at-the-edge survey. ### Visuals/Charts Three-column diagram: **Software → Hardware → Frameworks** ### Citation Navardi et al. (2025) --- # Slide 19 — Benefits & Challenges ### Title **BENEFITS OF GENERATIVE AI AT THE EDGE** ### Content ### Low Latency * Processing occurs closer to the data source. ### Enhanced Privacy * Sensitive data can remain local. ### Reduced Bandwidth Usage * Less dependence on transferring raw data. ### Improved Reliability * Local services can operate with reduced cloud dependence. ### Real-Time Intelligence * Supports applications requiring immediate responses. ### Speaker Notes Keep this slide positive and concise. The next slide addresses the limitations and trade-offs. ### Citation Wang et al. (2025); Navardi et al. (2025) --- # Slide 20 — Benefits & Challenges ### Title **CHALLENGES OF GENERATIVE AI AT THE EDGE** ### Content ### Resource Constraints * Limited CPU, memory, storage, and energy. ### Large Model Size * Foundation models may be too large for edge devices. ### Performance Trade-Offs * Compression can affect model quality. ### Energy Consumption * AI inference can increase device power requirements. ### Heterogeneous Hardware * Edge devices have different capabilities and architectures. ### Speaker Notes Explain that the main research problem is balancing performance with efficiency. ### Citation Wang et al. (2025); Navardi et al. (2025) --- # Slide 21 — Future Perspectives ### Title **FUTURE PERSPECTIVES** ### Content ### More Efficient Foundation Models * Smaller and edge-native generative models. ### Advanced Compression * Improved quantization, pruning, and distillation. ### Hardware-Aware AI * Models designed specifically for edge hardware. ### Collaborative Edge Intelligence * Dynamic workload sharing across edge nodes. ### AI-Native Networks * Integration with future 6G and decentralized computing systems. ### Autonomous Edge Sensors * Sensors capable of local intelligence and self-healing data services. ### Speaker Notes The final point is strongly supported by FedGen-Sensor, while the other directions are drawn from the surveys. ### Citation Wang et al. (2025); Navardi et al. (2025); Pei et al. (2026) --- # Slide 22 — Conclusion ### Title **CONCLUSION** ### Content * Generative AI at the edge brings intelligent computation closer to data sources. * It can reduce latency, improve data privacy, and reduce cloud dependence. * Resource constraints remain the central challenge. * Efficient deployment requires: * Model optimization * Hardware optimization * Intelligent orchestration * Collaborative edge computing * Recent experiments demonstrate that lightweight GenAI systems can operate on resource-constrained platforms. ### Final Statement **The future of GenAI is not only larger cloud models—it also includes efficient, collaborative, and intelligent systems operating directly at the edge.** ### Speaker Notes End by connecting all the papers: optimization enables deployment, benchmarking measures feasibility, orchestration manages resources, and federated learning enables collaborative intelligence. --- # Slide 23 — References ### Title **REFERENCES** ### Content **[1]** M. Navardi, R. Aalishah, Y. Fu, Y. Lin, H. Li, Y. Chen, and T. Mohsenin, “GenAI at the Edge: Comprehensive Survey on Empowering Edge Devices,” *Qeios*, 2025. **[2]** X. Wang, Z. Tang, J. Guo, T. Meng, C. Wang, T. Wang, and W. Jia, “Empowering Edge Intelligence: A Comprehensive Survey on On-Device AI Models,” *ACM Computing Surveys*, 2025. **[3]** F. Akram, A. W. Malik, and S. U. Khan, “iGenEdge: Intelligent Generative AI Service Deployment for Edge-Connected IoT Devices,” *IEEE Internet Computing*, 2025. **[4]** Z. Nezami, M. Hafeez, K. Djemame, S. A. R. Zaidi, and J. Xu, “Descriptor: Benchmark Dataset for Generative AI on Edge Devices (BeDGED),” *IEEE Data Descriptions*, 2025. **[5]** J. Pei, M. Feng, P. Muthukumar, P. V. Krishna, and M. Dai, “Generative AI-Native Edge Sensors: Collaborative Data Reconstruction via Lightweight Federated Learning,” *IEEE Journal of Selected Areas in Sensors*, 2026. ### Speaker Notes Use the full citation style required by your department. Some bibliographic details may need final formatting directly from the PDFs. --- # Slide 24 — Thank You ### Title **THANK YOU** ### Subtitle **Questions & Discussion** ### Suggested Closing Question **Can the future of Generative AI move from centralized cloud intelligence to collaborative intelligence at the edge?** --- Follow Design: {"palette":["Deep silicon graphite #14171A — primary dark background","Phosphor lime #39FF14 — active edge inference indicators and top highlights","Cool terminal mist #E2E8F0 — readable high-contrast body text","Industrial steel #334155 — card backgrounds and boundary dividers","Flash amber #F59E0B — resource constraint markers and alerts","Subdued slate #94A3B8 — metadata, timeline milestones, and footnotes"],"fonts":{"Archivo":"https://fonts.googleapis.com/css2?family=Archivo:ital,wght@0,100..900;1,100..900&display=swap","Inter":"https://fonts.googleapis.com/css2?family=Inter:ital,wght@0,100..900;1,100..900&display=swap","Fira Code":"https://fonts.googleapis.com/css2?family=Fira+Code:wght@300..700&display=swap"},"type":"Archivo in heavy industrial weights for impactful section headers; Inter for dense multi-bullet technical explanations; Fira Code for architecture layer labels and latency benchmark values.","layout":"Horizontal pipeline flow grid emphasizing tiered system layers (Sensors to Edge Nodes to Aggregators to Cloud), framed with modular HUD-style corner brackets and uniform content blocks.","framework_treatment":"Dark industrial cards with tactical cut-corner chamfers, glowing neon connection nodes for sensor networks, high-contrast latency delta badges, and modular step-by-step optimization cards.","feels_like":"A cutting-edge DARPA IoT edge-computing prototype demo and silicon hardware briefing"}
Presentation overview
This community presentation was shared by Varghese CJ as a complete 23-slide example. ## Topic: **Generative AI at the Edge: Efficient Intelligence for Resource-Constrained Devices** --- ## Slide 1 — Title Slide ### Title **GENERATIVE AI AT THE EDGE** ### Subtitl… The preview lets you review the full sequence in order, rather than judging the design from a single cover image.
Use it as a reference for planning your own deck: notice how the amount of information changes from slide to slide, where visual emphasis appears, and how repeated design choices help the presentation feel connected. Community examples are inspiration, not locked templates, so you can keep the ideas that fit your audience and replace anything that does not.
Learn from the example
Look at the job each slide performs. Identify where the deck introduces its subject, develops the main points, adds evidence or examples, and moves toward a conclusion.
Notice which element attracts attention first and how headings, supporting text, images, and data are separated. Strong hierarchy makes the intended reading order obvious.
Compare spacing, typography, color, and repeated components across the slides. Consistent rules help varied content feel like one presentation instead of unrelated screens.
Make it your own
Begin with your audience and the decision, lesson, or action the presentation should support. Replace the topic, examples, evidence, and visuals with material you can verify, then edit every slide for one clear takeaway. You can borrow the visual direction without copying the creator's wording or message. Presenton will use this community deck as a design reference while you develop content for your own purpose.
