The GenAI Revolution: Google's Gemma and Advanced Prompting for Maximum Impact

This session explores Google's GenAI models, covering key concepts like prompting, fine-tuning, and retrieval-augmented generation (RAG). We'll dive into Gemma, examining its architecture, capabilities, and why it's gaining attention in the AI community. Interactive hands-on exercises will showcase Gemma's advanced prompting techniques, followed by discussions on GenAI's future impact and a Q&A.

For Further Description, here is the Session link : https://gdg.community.dev/events/details/google-gdg-ahmedabad-presents-the-genai-revolution-googles-gemma-and-advanced-prompting-for-maximum-impact/

Transcript

1. The GenAI Revolution: Google's Gemma and Advanced Prompting for Maximum Impact Presented by: Krupa Galiya Presented for: GDG Ahmedabad
2. Outline Introduction and GenAI concepts in brief Unveiling the Hype Introduction to Gemma and Its Workings Hands-on with Gemma Future of GenAI Open discussions and Q/A
3. Introduction and GenAI concepts in brief
4. Proprietary + Confidential Evolution of AI Use Cases Predictive AI Regression & Classification Forecasting Sentiment Analysis Entity Extraction Object Detection Generative AI Text, Image & Code Generation Text & Code Rewriting & Formatting Summarization Extractive Q&A Image & Video Descriptions Multimodal Generative AI Natural Image Understanding Spatial Reasoning and Logic Mathematical Reasoning in Visual Contexts Video Question Answering Automatic Speech Recognition & Translation
5. 2015 TensorFlow democratizes AI 2016 TPUs boost AI speed and efficiency. 2017 Birth of Transformers use of Attention mechanism 2018 BERT <> Google Search 2021 LaMDA (Language Model for Dialogue Applications) 2016 AlphaGo defeats world champion Go player 2023 Bard(later it called Gemini) helps users to collaborate and Integrations with GenAI, Codey 2024-present Gemini 1.5, Gemma, PaliGemma… Reference History of Google’s Contribution in GenAI
6. Key GenAI Concepts
7. Prompting Prompting involves giving the model a specific input to guide its response or generation, allowing users to control the model’s output.
8. Fine-Tuning Fine-tuning adjusts a pre-trained model using domain-specific data to improve its performance on specialized tasks while retaining its general knowledge.
10. Retrieval-Augmented Generation RAG combines generative models with a retrieval system, allowing the model to access external knowledge for more accurate and contextually enriched responses.
11. Few-Shot Learning: Few-shot learning allows a model to generalize and perform tasks by being shown only a few examples, leveraging its pre-existing knowledge. Zero-Shot Learning: Zero-shot learning enables a model to perform a new task without having been trained on it, based on its understanding of related tasks and general knowledge. Multimodal Learning: Multimodal learning involves processing and understanding data across different formats like text, images, and audio, enabling richer interactions and insights.
12. Unveiling the Hype
13. Gemma open models A family of lightweight, state-of-the art open models built from the same research and technology used to create the Gemini models
15. Open Source vs Open Model • Primarily concerns the software's source code. • Allows complete modification and customization of the software code to suit any purpose. • Gives full access to the software's code, showing how it works. • Specifically refers to the pre-trained AI model and its related components (architecture, training details, sometimes datasets). • Enables adaptation and fine-tuning for specific tasks, but typically restricts changes to the core model architecture. • Reveals the model's design, training and weights but may still have some “Black box” parts due to AI complexity. In essence, Open Source empowers to build and change software as you wish. Open Model empowers to use and adapt powerful AI capabilities.
16. Why Open Model?
17. Control and Customization - Client khush hua! Kyunki ab hum model apni marzi se customize kar sakte hain!
18. Data Privacy and Security - Ja data, apne servers pe hi jee le apni zindagi! Open models me privacy ka full control hai!
19. Avoiding Vendor Lock-In - Kitne Vendor the? Being locked into just one vendor? That’s unfair! Open models give you more freedom.
20. Innovation and Community Support - Innovation ko rokna mushkil hi nahi, namumkin hai! Jab open-source community mil jaaye!"
21. Cost-Effectiveness- Open models can run on a company's own infrastructure, removing the need to pay for expensive API calls or cloud services from external providers.
23. Family of Gemma model
25. Gemma 2 offers best-in-class performance, runs at incredible speed across different hardware and easily integrates with other AI tools. google blog
26. Description (from kaggle) Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. They are text-to-text, decoder-only large language models, available in English, with open weights for both pre-trained variants and instruction-tuned variants. Gemma models are well-suited for a variety of text generation tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as a laptop, desktop or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone.
27. Small Models (SMs) are frequently used in practical settings, although their significance is currently underestimated What is the Role of Small Models in the LLM Era: A Survey
29. Data Preprocessing Here are the key data cleaning and filtering methods applied to the training data: 1. CSAM Filtering: 2. Sensitive Data Filtering 3. Additional methods:
30. Let’s meet gemma in studio
31. Libraries Libraries SDKs ai.google.dev REST API + Client libraries for Python, Node, Java, and Swift
32. aistudio.google.com • Generate API Keys • Create, test, and save prompts • Customize models in minutes • Generate starter code
34. Technical report
37. QKV Vectores The Scene: You're a student in a class, trying to understand a complex topic. Your teacher is explaining, but you need some clarification. You want to ask a question but also want to make sure your question is relevant to what's being discussed. The Players: You (the student): You have a Query (Q) in your mind – the question you want to ask. Your Classmates: Each of your classmates has a Key (K) – a summary of what they understand about the topic. The Teacher: The teacher has the ultimate Value (V) – the in-depth knowledge and explanation of the topic.
38. Key Findings Gemma explained: What’s new in Gemma 2 - Google Developers Blog Distillation vs. Training from Scratch (Teacher student model) Grouped Query Attention vs. Multi Head Attention Model Depth vs. Width Alternating Local and Global Attention
40. Hands-on with Gemma
41. Hands-on with Gemma https://tinyurl.com/ahm-prompt
42. reference
43. GenAI Application Demonstration https://shorturl.at/1HZqx
45. Future with GenAI
46. Next Focus on GenAI Enterprise-Ready Generative AI: Enhancing security, privacy, customization, and workflow integration to enable confident and seamless adoption by businesses. Accessibility & Impact: Expanding the user base by democratizing AI development, driving innovation across industries, and solving real-world problems.
47. Industry Transformation: Automating tasks, augmenting capabilities, unlocking new levels of creativity and productivity across sectors. Personalized Experiences: Tailored product recommendations, personalized learning, and unique experiences for everyone. Enhanced Creativity and Innovation: Generating new ideas, accelerating the creative process, creating novel content. Improved Decision-Making: Analyzing vast data, providing valuable insights, facilitating data-driven decisions. Democratization of Expertise: Making specialized knowledge and skills accessible, acting as a virtual assistant for guidance and support. Ethical and Societal Considerations: Addressing ethics, privacy, and potential biases for responsible development and deployment. Continuous Evolution: Advancements leading to even more powerful and versatile GenAI applications.
48. QA and Open Discussions