[MLDP Newsletter] Oct 2023 — Machine Learning Communities: 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!
Generative AI/LLM
MakerSuite / PaLM
PaLM2 API with Gradio Chat by ML GDE Chansung Park (Korea) is a project demonstrating how to build a feature-rich chatbot UI in Gradio. He implemented features including multi chat channels, save/restore history, pause text response generation, regenerate last conversation, clean chat history, adjust text generation parameters, inspect actual prompt that the model sees and etc. His other project, Zero2Story is a framework built on top of PaLM API, Stable Diffusion, MusicGen for ordinary people to create their own stories. This has been accepted for the NeurIPS workshop that will be held in person in December.
Content creation using MakerSuite and PaLM API by ML GDE Dimitre Oliveira (Brazil) is about using MakerSuite for fast prototyping and experimenting with ideas, and then transitioning to PaLM API to create a real app to deploy in production.
Exploring Generative AI using MakerSuite and PaLM API by Mikaeri Ohana (Brazil) explains solutions to prototype and how to use tools such as MakerSuite, PaLM, Bison, and etc. She covered the best scenarios to apply and some use cases.
PaLM API for Health Pre-Screenings for Women (Github) by ML GDE Sara EL-ATEIF (Morocco) is to help women identify health issues by answering questions and queries about their health symptoms. The goal is to encourage women to use this conversational model regularly to stay informed and seek care when needed.
Converting a Presentation Transcript to Speaker Notes by ML GDE Martin Andrews (Singapore) explains how to automatically convert an audio recording of a talk & the content of the slides into speaker notes.
How to extract knowledge from documents with Google PaLM 2 LLM by ML GDE Al Kari (United States) offers a quick and straightforward method leveraging PaLM 2 API to extract knowledge and ask questions from text, specifically PDF documents.
Code Generation with PaLM 2 by ML GDE Jéssica Costa (Brazil) is about experiments on the PaLM 2 code-bison model. She shared two code generation experiments, Python and SQL.
How can we compose music using AI? by ML GDE Khongorzul Munkhbat (Mongolia) discussed three methods of using AI for music composition: through LSTM, Magenta, and the combination of PaLM API and MakerSuite.
At Artificial Intelligence Information Security Day hosted by GDG Taipei, ML GDE Jerry Wu (Taiwan) shared about text embeddings with encryption in PaLM. ML GDE Tzer-jen Wei (Taiwan) demonstrated how to use Generative AI Studio and PaLM API to find security issues in smart contract codes.
AI In the “PaLM” of Your Hand by ML GDE Stephen Wylie (United States) helped people make their own models on top of PaLM using MakerSuite/Vertex AI. He kicked off with an introduction to the foundation of LLMs, and some specifics around PaLM 2 such as its strengths, training considerations, and potential costs.
Implementing Generative AI FAQ Bot on own documents by ML GDE Yogesh Kulkarni (India) explained how to use PaLM APIs to build custom chatbots.
Building Generative AI Applications on Google Cloud by ML GDE Yüksel Tolun (Turkey) covered from basic concepts to its practical applications. This talk tailored for those who want to leverage the power of Vertex AI and its new features — Search and Conversation, and how to synthesize VertexAI with Langchain.
Building Web Solutions Using ML by TFUG Surat helped community members learn how to use MakerSuite. As a speaker, Firebase and Web GDE Vrijraj Singh (India), walked people through the ML in Web.
AI/ML & Data Talks Podcast Episode 19 by ML GDE Kuan Hoong (Malaysia) & ML GDE Sachin Kumar (Qatar) shared a board-ranging discussion covering their career and working with PaLM 2.
DevFest 2023 by GDG Montreal covered various ML topics such as Crafting Intelligent Flutter Apps: A Journey with FlutterFlow, MakerSuite, and PaLM API by Flutter GDE Boris-Wilfried (Canada); Using LLMs to bridge the Fuzzy Human/Digital Computer Boundary by ML GDE Allen Firstenberg (United States); and Leveraging BigQuery and Python Notebooks to Power Your AI/ML Projects by JL Marechaux (Googler).
DevFest Mauritius 2023 by GDG Mauritius introduced Adding generative AI to Flutter with PaLM API (by Dart GDE Sylvia Dieckmann) and Building Smart Apps with TensorFlow Lite (by Pranav Gupta Chummun).
More activities
LLaMA-2 hands-on workshop by ML GDE Sam Witteveen (Singapore) is on how to use LLaMA2 for creating LLM apps.
Open Source LLMs: Owning Your Own LLMs in a Minute (slides) by ML GDE Chansung Park (Korea) summarized his open source projects, including LLM As Chatbot and some other side open source projects. The main focus was to share the best practice of how to choose an LLM from millions of open sources, and why it is important to actually try it out yourself for specific use cases.
AI Workshop: Build ChatGPT-like Chatbot from Scratch by ML GDE Tarun R Jain (India) covered LLMs and Embedchain library workflow on how to build a chatbot like ChatGPT, VectorDB, and RAG. His article Empower Multiple Websites with Langchain’s Chatbot Solution also explains the details.
LLMX — An API for Chat Fine-Tuned Large Language Models by ML GDE Victor Dibia (United States) is a Python package providing a unified interface to several LLM providers of chat fine-tuned models — PaLM Chat-Bison, CodeChatBison, HF, OpenAI, Cohere and etc.
alpaca-lora: Experimenting with home-cooked Large Language Model by ML GDE Wei Zheng (China) introduced the concept of large foundation models (LFMs) and several fine-tuning methods making LFMs behave as desired. He focused on LoRA and explained the fine-tuning code as well as performance improvement techniques.
Is it impossible to fine tune LLMs to get new knowledge? by ML GDE Mitsuhisa Ohta (Japan) shared his experiment to clarify whether it is possible to fine-tune LLMs to get new knowledge. He also conducted a hands-on workshop on utilizing multiple-GPU and multiple nodes on ABCI.
Singapore Spotlight by TFUG Singapore invited four local Singaporean speakers and explored LLM topics. The speakers covered reliability of LLMs in production, how to use LLMs as a system of multiple expert agents, how to use open-endedness techniques, and introduced Token-Crisis on LLM scaling and OpenMoE project separately.
How Generative AI improves the productivity of Software developers by Kotlin GDE Monika Kumar Jethani (India) looked at various ways Gen AI can improve productivity of software developers.
Keras
Segment Anything Model with 🤗 Transformers by ML GDE Merve Noyan (France) and Sayak Paul (India) is a Keras example on how to use and fine-tune Segment Anything Model from Meta. And their post, 📌Controlling Stable Diffusion with JAX, diffusers, and Cloud TPUs on Google Opens Source Blog introduced the community sprint hosted by HuggingFace & Google Cloud and showcased a few projects from the sprint.
My Keras Chronicles (ML) (video) by ML GDE Aritra Roy Gosthipaty (India) introduced Keras 3 and the use of JAX with it. It was a part of Keras Knights event hosted by GDG Nairobi.
How to survive to Naked and Afraid using an Autoencoder by ML GDE Arnaldo Gualberto (Brazil) was a DevFest talk explaining different types of autoencoders with Keras and TensorFlow for a large variety of applications like: dimensionality reduction, embeddings search, image generation, segmentation, image denoising, and recommendation systems.
Introduction to Keras Core: unlocking the power of JAX with Keras Core by ML GDE Kuan Hoong (Malaysia) was about the power of Keras Core’s multi backend support and JAX.
KerasCV for the Young and Restless by ML GDE Suvaditya Mukherjee (India) was an introduction to Keras, KerasCV, and the newest features from the development of Keras into Keras 3.
KerasNLP: From Words to Wisdom (slides) by Abheesht Sharma (India) & Anshuman Mishra (India) at DevFest New Delhi introduced LLMs and how to use KerasNLP to train models.
Keras Core for the Curious and Creative by Ngesa Marvin (Kenya) introduced Keras Core with JAX and showed object detection demo using Colab & multi-framework modeling with KerasCV. He also led Keras hands-on workshops at DevFest Kigali and DevFest Nairobi.
#KCD In Monthly AI/ML Meetup — Keras Community Day hosted by GDG Nuremberg, ML GDE Imran us Salam Mian (Germany) led a workshop to help people understand what Keras is and how to get started with it.
#KCD Keras Core with JAX: Streamlining Deep Learning for Robust Acceleration (slides) by ML GDE Jeongkyu Shin (Korea) introduced JAX, and explored how to leverage the performance of JAX through Keras Core.
#KCD Improving digital interaction for Women with Machine Learning by ML GDE Lesly Zerna (Bolivia) gave a talk about how to get started with Keras, MakerSuite and Bard to grow a professional career and improve digital interactions with tech communities. She also hosted Keras Community Day Bolivia as an organizer and main presenter.
#KCD Quantization aware training with Keras by ML GDE Rajesh Shreedhar (United States) shared a technique, quantization of model, which is to improve the performance of deep learning models by reducing the precision of the model’s weights. He went through details of quantization-aware training with Keras examples.
#KCD Unleashing the Power of Keras: Embracing the Multi-backend Future with TensorFlow, PyTorch, and JAX by ML GDE Sam Witteveen (Singapore) delivered what new Keras Core multi backend and how you can take advantage of it to create deep learning models.
#KCD Free-market ML: competing frameworks, abstractions, and types by ML GDE Samuel Marks (United States) was a talk & workshop walking you through the library level implementation, open-source contribution, and the likely future.
#KCD Build Your Own GenAI Model by ML GDE Vikram Tiwari (United States) discussed the basics of generative AI and how to build your own model.
Kaggle
Fine-tune Llama 2 for sentiment analysis by ML GDE Luca Massaron (Italy) is a hands-on tutorial on fine-tuning Llama 2. He dealt with a sentiment analysis on financial and economic information. Plus he uploaded articles explaining theory and practice of his project: Fine-tuning a large language model on Kaggle Notebooks for solving real-world tasks — part 1 & part 2.
Two Towers Recommender by ML GDE Rubens Zimbres (Brazil) is a dataset and a Kaggle notebook about the Two Towers Recommender in Tensorflow/Keras.
KerasNLP starter notebook writing quality by ML GDE Alexia Audevart (France) is shared for the competition. KerasNLP starter notebook Contradictory DearWatson (competition) & KerasNLP starter notebook Disaster Tweets (competition) are also shared for competitions hosted by Kaggle. The last notebook was copied and edited more than 360+ times.
Responsible AI in Predictive Underwriting by ML GDE Guan Wang (Singapore) is a Kaggle notebook delving into the ethical application of AI in insurance underwriting with some Kaggle datasets and dummy modeling. He explored transparency and fairness with the open source ‘What-If-Tool’ from Google.
XGBoost for tabular data (video) by ML GDE Luca Massaron (Italy) is a hands-on tutorial on using XGBoost to solve a regression problem with open data. This session is a part of Kaggle X Mentorship Program.
MLOps
#Keras #TFX LLMOps: from fine-tuning to deployment: with KerasNLP and TensorFlow Extended (slides) by ML GDE Chansung Park (Korea) covers ML pipeline from fine-tuning to deployment of a LLM with Keras. His article, Building a Machine Learning Pipeline with TensorFlow Extended and W&B shared how W&B’s experiment tracking and model registry can be integrated into TFX.
#Kubeflow ML GDE David Cardozo participated in the recent Kubeflow update — 📌Kubeflow 1.8: Kubernetes MLOps delivered via Python workflows. Thanks to the update, ML pipelines are now constructed as modular components, enabling easily chainable and reusable ML workflows.
#VertexAI #Kubeflow Cloud-Scale AI: Mastering MLOps with Vertex AI and Kubeflow by ML GDE David Cardozo (Canada) explored how the tools are revolutionizing container orchestration on Kubernetes, making ML project deployment, monitoring, and maintenance easy. His other talk, From Bits to Jokes: Our Comic Journey from sending ML to Production (video), shared his implementation experiences in MLOps-based products.
#VertexAI #Kubeflow Automating your ML pipelines using Kubeflow and Vertex AI by ML GDE Henry Ruiz (United States) introduced how to automate ML pipelines on Vertex AI using Kubeflow at DevFest Central Florida.
#TFX BERT as a service by ML GDE Dimitre Oliveira (Brazil) is designed to demonstrate a simple yet complete ML solution that uses a BERT model for text sentiment analysis using a TFX end-to-end pipeline.
#Dialogflow Deploying your Dialogflow CX Webhook by ML GDE Xavier Portilla Edo (Spain) is a repository of webhook for Dialogflow, using Typescript types and local debugging.
In Exploring AI use cases and building scalable ML pipelines, ML GDE Hannes Hapke (United States) shared his MLOps journey such as how to manage GPU resource allocation and tackle challenges of implementing TF server. He also gave a talk, Fine-Tune LLMs or Integrate 3rd party APIs? A Financial Case-Study comparing fine-tuning open source models vs. consuming model APIs. He shared lessons learned from running LLMs in production as well.
On-device ML
#MediaPipe Empowering Android Apps with On-Device Machine Learning by ML GDE Pankaj Rai (India) focused on introducing ML Kit and its features and various capabilities. He also shared how to use MediaPipe and how to run a custom ML model using MediaPipe Studio.
#MediaPipe AI/ML Symposium 2023: From Theory to Impact by GDG Hong Kong was a big event covering various ML topics such as MediaPipe, Kaggle, VertexAI, and ML research. One of the sessions was “ML for Everyone: Building On-Device ML Applications Without the Hassle” (by GDG Hong Kong organizer Emma Wong.
[ML Story] Machine Learning on the browser: TF Lite meets TF.js by ML GDE Nitin Tiwari (India) demonstrates object detection in web browsers, showing the deployment of TFLite models through TF.js. In this particular example, he showed the detection of different categories of waste found in our surroundings aiming to solve some environmental problems around us.
#Dialogflow Build conversational AI experiences powered by LLMs with Vertex AI Conversation and Dialogflow CX by ML GDE Xavier Portilla Edo (Spain) shared the latest generative AI features in Vertex AI Conversation and Dialogflow CX and explained how to combine traditional agent design techniques and best practices with Google’s latest LLMs to create complex conversational applications.
#Cloud Community Connect with TFUG Prayagraj by TFUG Prayagraj at IIIT Allahabad aimed to explore AI and Cloud domains with passionate students. The agenda included an overview of Google Cloud’s AI offerings, including VertexAI, Dialogflow, Duet and AI infrastructure like TPUs.
#Cloud [ML Story] How I cleared the Google Cloud Machine Learning Certification in 2023 by ML GDE Bhavesh Bhatt (India) shared his preparation strategy and tips to get the certification.
#VertexAI Accelerating Model deployment using Transfer Learning and Vertex AI by ML GDE Robert John (Nigeria) was a talk at DevFest Kigali 2023 sharing about ML pipelines, transfer learning, and model deployment with Vertex AI.
#VertexAI Developing solutions using LangChain and VertexAI (video) by ML GDE Pedro Gengo (Brazil) shared how to use LangChain with Vertex AI with two demos.
#VertexAI A CAMEL ride: A Story of AI Role-Playing using CAMEL, Langchain and VertexAI by ML GDE Yogesh Kulkarni (India) showed how to use the AI agent, CAMEL and how to make AI entities collaborate and interact with each other.
#TensorFlow 3D Vision models for MRI image analysis by ML GDE Mohamed Berrimi (Algeria) shared the challenges that researchers face when dealing with 3D data such as CT and MRI. He gave a talk at the IPTA23 conference in Paris to show how TensorFlow can be used especially for that.
Others & regional highlights
In 📌Getting started with community contributions, diffusion models, and more, ML GDE Sayal Paul explored what it means to be a GDE and how to leverage the power of community through community contributions. It is a podcast session in People of AI: Session 2.
In honor of Hispanic Heritage Month, ML GDE Henry Ruiz (United States) was introduced in 📌How Machine Learning GDE Henry Ruiz is inspired by resilience in his community.
GalsenAI hosted Indaba X Senegal (official site: INDAVA Senegal 2023) as a coorganizer along with ESP and WiMLDS. The theme of the year was challenges and prospects for the effective implementation of national data and artificial intelligence strategies in Senegal.
Python For Data Science For Dummies 3rd Edition by ML GDE Luca Massaron (Italy) offers new datasets, examples, Python commands, and ML techniques, all in Colab environment.
Afterwork on research in Machine Learning by TFUG Abidjan & Bassam is the first session of ML Paper Reading Clubs cohosted by the two TFUGs (Côte d’Ivoire). They learned how to read ML papers and shared tips.
Cloud Googlers & ML GDEs presented Cloud Architecture center, TensorFlow capabilities to generate music, Palm API’s, LLM capabilities at Google developers panel at Central Asia’s largest forum Digital Bridge in Astana, Kazakhstan.
Google dev community in Discord is becoming a hub for ML developers covering various Google ML products as well as hosting events online.
The 6th CASSINI Hackathon by the European Commission hosted 4 GDEs and they helped the use of ML and GCP for the participating teams.
