[Dec 2024] ML Community — Highlights and Achievements

[Dec] ML Community — Highlights and Achievements

Let’s explore highlights and accomplishments of the vast Google Machine Learning communities over the month. We appreciate all the activities and commitment by the community members. Without further ado, here are the key highlights!

AI/ML Training Campaign highlights in 2H 2024

Throughout 2024, the AI/ML communities have actively engaged in the campaigns. 304 AI/ML Study Jams, 65 AI/ML Math Clubs, 46 AI/ML Paper Reading Clubs, and 16 AI/ML Paper Writing Clubs were held across the globe. Here are the highlighted activities in the 2H 2024.

AI/ML Paper Reading Clubs | 23 events by 9 communities

• TFUG Taipei (playlist: 6 videos) hosted a year round series of AI/ML Paper Reading Club sessions covering topics from “Understanding diffusion models” to “Introduction to JAX”. TFUG Islamabad (playlist), AI Tamilnadu (playlist), Lucknow AI Labs (playlist) are also rolling out AI/ML Paper Reading Clubs.

AI/ML Math Clubs | 22 events by 10 communities

• TFUG Ghaziabad (playlist 9 videos with 2000+ views) hosted a series of AI/ML Math Club sessions covering the mathematical foundations to linear regression. Sayma Shahid (Bangladesh, playlist) hosted a series of Calculus with ML/AI for Machine Learning, AI, Deep Learning & NLP Community — Bangladesh. GDG On Campus ISAMM (video) and TFUG Islamabad (video) also have started AI/ML Math Clubs.

AI/ML Paper Writing Clubs | 8 projects led by 6 communities

• AI/ML GDE Rabimba Karanjai’s paper, TPU as Cryptographic Accelerator (arXiv) has been accepted at HASP’24.
• AI/ML GDE Anshuman Mishra and Abheesht Sharma’s co-authored paper with the Keras team at Google, KerasCV and KerasNLP: Vision and Language Power-Ups (arXiv), has been accepted at JMLR.
• AI/ML GDE Sayak Paul’s paper, Getting it Right: Improving Spatial Consistency in Text-to-Image Models (Project page, arXiv) has been accepted at ECCV 24.

AI/ML Study Jams | 113 events organized by 61 communities

• Machine Learning Uyo (playlist) and TFUG Hyderabad (playlist) hosted a series of AI/ML Study Jams leveraging Kaggle Learn courses.

Gemini

https://medium.com/media/3825a0998c29b0dea02373891d168ee8/href

Gemini running in real time, from my phone. With spatial awareness. Interacting with me in realtime! (Bengali version) by AI/ML GDE Rabimba Karanjai (US) is a short video showing how Gemini can interact with you.

Bringing Images to Life: Object Detection with Google’s Gemini 2.0 Spatial Understanding

Bringing Images to Life: Object Detection with Google’s Gemini 2.0 Spatial Understanding by AI/ML GDE Burak PASTIRMACI (Türkiye) shows how the Gemini API performs object detection and its spatial understanding. He also summarized the major update to Gemini 2.0 and what it can do.

Gemini 2.0 — Google announces its entry into the world of AI Agents by AI/ML GDE Liu Yu-Wei (Taiwan) summarizes the new updates including main functions and applications. He also shared his Colab Notebook of an AI Agent example using Gemini 2.0.

LINE Bot with Gemini-2.0-flash-exp (repository) by AI/ML GDE Jimmy Liao (Taiwan) shows how to integrate LINE bot with Gemini 2.0 Flash from scratch.

Gemini 2.0 Cookbook (Traditional Chinese translation) by AI/ML GDE Will Huang (Will Huang) is a Chinese translation version of the Gemini 2.0 Cookbook.

https://medium.com/media/683f82f94b80a3200546c57e81b07791/href

Real-World Multimodal AI Agents use case | Phidata and Gemini 2.0 (repository | demo) by AI/ML GDE Tarun R Jain (India) walks you through how to build a multimodal agents app that gets you a detailed analysis of ingredients from images of consumer products.

ai-student-screen-reviewer (video) by AI/ML GDE Cyrus Wong (Hong Kong) is the code for an actual test using Vertex AI and Gemini 2.0 Flash to analyze a large number of student computer screen videos.

Developing an AI-powered English Tutor (repository) by AI/ML GDE Carlos Alarcon (Colombia) explored the process of creating an English tutor app interacting with users. Along with the fundamentals of Automatic Speech Recognition and Text-to-Speach technologies, he covered how to design and implement the AI core and the system architecture of the app using Gemini and Gemma.

Exploring Google Gemini: Applications in Software Engineering (slides | demo video) by AI/ML GDE Fernanda Wanderley (Brazil) showed how developers can harness the power of Gemini to accelerate their tasks, by, for example, searching for a code at a GitHub repository.

AI Workshop on Google AI Studio & Gemini at REC Sonbhadra (slides) by TFUG Prayagraj focused on empowering developers who are new to AI and ML by introducing AI Studio and Gemini. This event was held 170km away from the group’s home city, Prayagraj, as part of their commitment to support developers in remote areas.

Gemma

Welcome PaliGemma 2 – New vision language models by Google

#PaliGemma Welcome PaliGemma 2 — New vision language models by Google co-authored by AI/ML GDE Aritra Roy Gosthipaty (India) is a post introducing PaliGemma 2 with its capabilities, demo, fine-tuning methods and other resources.

#PaliGemma Semi-Automated Data Labeling for Computer Vision tasks with PaliGemma (repository) by AI/ML GDE Rajesh Shreedhar Bhat (US) was a talk about how to speed up the labeling process and enhance accuracy with PaliGemma.

#ShieldGemma Creating Endpoint Development for Chatbot and Personal Assistant based on LLM by AI/ML GDE Joan Santoso (Indonesia) was a workshop focusing on creating functional chatbots and personal assistants. He also discussed how to secure the app using ShieldGemma to prevent creating dangerous content.

#DataGemma How to avoid hallucinations with DataGemma (Colab Notebooks | slides) by AI/ML GDE Mikaeri Ohana (Brazil) delved into various types of hallucination, exploring their implications and the critical need for frameworks to mitigate them effectively. She provided an in-depth analysis of concepts developed by DataGemma, focusing on RIG and RAG.

Building Advanced Assistants (Agentic Workflows) with Gemma 2 and Llama Index by AI/ML GDE Marvin Naftali Ngesa (Kenya) dove deep into RAG with Gemma 2 and Llama Index by covering agentic RAG, router agents, and research agents.

Responsible AI

Panel Discussion: AI for Good at KL DevFest 2024 by AI/ML GDEs, Esther Irawati Setiawan (Indonesia), Poo Kuan Hoong (Malaysia), Leong Lai Fong (Malaysia), and Googler Thu Ya Khaw was a conversation on about Gen AI and the potential risks of the cutting-edge technologies.

Designing & Implementing Responsible Agents with Gemma by AI/ML GDE Tahreem Rasul (Pakistan) focused on building responsible AI agents using Gemma. The session highlighted how Gemma was designed and trained in the light of responsible AI and offered practical suggestions for incorporating Responsible AI into workflows using Responsible AI Toolkit.

Responsible AI and Gemma Scope by AI/ML GDE Xiaohu Zhu (China) introduced Google’s Responsible AI approaches and Gemma Scope, a mechanistic interpretability tool for LLMs. He also showcased Gemma Scope examples.

https://medium.com/media/dcf047b43f83560599da3f5cecaedea9/href

AI Ethics and Responsible AI Session (video) by Machine Learning Uyo focused on the critical role of ethical considerations in AI development and the importance of implementing responsible AI practices across industries. Armel Yara (Canada) provided practical insights into how ethical guidelines can be integrated into AI workflows to prevent harm and enable better decision-making. They also covered how to build responsible AI systems with an introductory session by AI/ML GDE Sara El-Ateif (Morocco).

Kaggle

Gemini Powered Multilingual Finance Assistant by AI/ML GDE Bhavesh Bhatt (India) makes interactions faster and efficient, especially for users seeking specific information on subsidy eligibility, loan options, tax benefits, and support programs by leveraging Gemini 1.5’s context window, context caching, and multilingual capabilities.

https://medium.com/media/2489134fc379a3e98d5ad54f2919974c/href

AI Trailer — Using Gemini to create video trailers (Colab version) by AI/ML GDE Dimitre Oliveira (Brazil) shows how to generate customized trailers for longer video files. He leveraged Gemini’s multimodal power and its long-context capability to analyze a long video file and create a trailer for it. In this video, he walks you through the code.

Keras

AI & Deep Fakes presentation — X-Ray Generation with Keras & Tensorflow by AI/ML GDE Mikaeri Ohana (Brazil) explored the technical aspects of Deep Fakes and related technologies such as Face Swap and facial manipulation, focusing on deep learning and neural networks. The curriculum emphasized mathematical models and CNNs, and participants used TensorFlow to run the demonstration of fake X-Ray image generation. This talk was part of an initiative that integrated applied knowledge with awareness efforts regarding the risks and ethical considerations of AI misuse.

Keras 3.0 : The Sequential model for images classification (Colab Notebook) by Amel Yara (Canada) was a talk at Devfest Yamoussoukro covering the fundamentals of sequential models with a practical live coding.

From Text to Insight: Exploring LLMs with Keras (Colab Notebook KerasNLP, Keras CV | slides) by AI/ML GDE Ahirton Lopes (Brazil) covered practical application of LLMs using KerasNLP and KerasCV. He featured demonstrations of pre-trained models, such as OPT for text completion tasks and Stable Diffusion for image generation.

ODML

Zolup browser, the mobile app that leverages AI (Google Play) by AI/ML GDE George Soloupis (Greece) is a Gemini-powered Android app designed to enhance mobile browsing accessibility and usability for all users. It addresses challenges such as reading small texts, navigating ad-heavy articles, and creating audio content for web pages.

JAX

[Lab-MLX] part 1: Inference with Gemma2 using JAX and Flax on Apple Silicon

[Lab-MLX] part 1: Inference with Gemma2 using JAX and Flax on Apple Silicon (repository) by AI/ML GDE Jimmy Liao (Taiwan) shares how to perform inference with Gemma2 using JAX/Flax on Apple Silicon, M2. He stressed it’s essential to grasp how to perform inference using native GPUs to truly understand performance tuning.

Firebase Genkit

Go Tiny Agents by AI/ML GDE Anubhav Singh (India) is a Go-based multi-agent system that simulates characters from the Sitcom, “The Office” using Gemini. Each agent has a unique personality and responds to messages based on their character traits.

VertexAI

From Data to Decisions: Simple AI model using Vertex AI (slides) by AI/ML GDE Adkham Zokhirov (Uzbekistan) explored how to transform raw data into actionable insights by building a simple AI model using Vertex AI. He covered from data preparation to model deployment, and demonstrated how Vertex AI simplifies the development and management of ML models.

Vertex AI lesson one: from studio to code & Vertex AI lesson two: setup Gemini Code Assist on VSCode and using Gemini 2.0 by AI/ML GDE Jimmy Liao (Taiwan) are a series of writings introducing how to leverage Vertex AI Studio and Gemini Code Assist to generate code.

Building a Multimodal RAG Application with Gemini & Vertex AI (Colab Notebook) by AI/ML GDE Tahreem Rasul (Pakistan) guided participants through the creation of a multimodal RAG application using Gemini & Vertex AI and how to deploy it.

Others

Unleash Your Creativity with Imagen 3 and Multimodal Gemini (slides) by AI/ML GDE Margaret Maynard-Reid (US) introduced Imagen 3 and its ability to generate photorealistic visuals. She also discussed how developers can rapidly prototype and bring AI-powered visions to life with either AI Studio or Vertex AI.

Virtual Try-On with Imagen 3 (Colab Notebook | repository) by AI/ML GDE Nitin Tiwari (India) demonstrates a project that allows virtual outfit try-ons using Imagen 3. It uses a pipeline combining the Gemini and SAM-2 models for text prompting to generate segmentation masks, which are then used to inpaint a new outfit image created by Imagen 3.

A translation application using Chrome’s AI and Angular by Angular GDE Connie Leung (Hong Kong) covers using Chrome’s Translation API and Angular to build a translation application. She also shared how to build a simple language detection application locally using Chrome’s Built-In Language Detection API and Angular.

Road to AI Community Day by Machine Learning Kolkata highlighted and showcased various topics in Gen AI including hallucination mitigation, PaliGemma and VLMs, and AI agents using Golang and Genkit. AI/ML GDEs Ritwik Raha (India), Suvaditya Mukherjee (US), Aritra Roy Gosthipaty (India), and Anubhav Singh (India) participated in the event as a speaker.

https://medium.com/media/f52aa56d3bb56c2ebbeed25988bffdba/href

NotebookLM: From research notes to personal podcasts (slides) by AI/ML GDE Allen Firstenberg (US) was a conversation with students studying Communications and Journalism. He introduced NotebookLM sharing how they can use it as part of their education and work. He also talked about NotebookLM’s features and potential use cases in Episode 214 — NotebookLM: The Future of Personalized AI Learning for Developers?

DCGAN implementation (Colab Notebook) by AI/ML GDE Arnaldo Gualberto (Brazil) is an implementation of a DCGAN using TensorFlow.

[Dec 2024] ML Community — Highlights and Achievements was originally published in Google Developer Experts on Medium, where people are continuing the conversation by highlighting and responding to this story.