[MLDP Newsletter] July 2023 — Machine Learning Communities: highlights and achievements

[MLDP Newsletter] July 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!

Keras Community Day

Keras Community Sprint 2023 focusing on KerasCV, KerasNLP has successfully ended, producing model implementations, tutorials and slide decks. The content is being distributed through the Keras Community Day campaign. GDG Cloud Almaty, TFUG Bassam hosted their events already and many KCD events are lined up.

https://medium.com/media/3de04d429329f07ea092f5020025bb13/href

A new library, Keras Core, was released as well! ML GDE Aakash Nain (India) and ML GDE Aritra Roy Gosthipaty (India) are listed as contributors. Introduction to Keras Core with Francois Chollet and Introduction to KerasCV with Google Software Engineer by ML GDE Aritra Roy Gosthipaty (India) and ML GDE Ritwik Raha (India) went live in line with the Keras Core release.

https://medium.com/media/92caa973dc2360520dfc396c9bd8c191/href

Keras

ML GDE Aritra Roy Gosthipaty (India) and Suvaditya Mukherjee (India) awarded 🏅TFCommunitySpotlight. Their When Recurrence meets Transformers tutorial features Keras patterns like creating RNNs from custom RNN cell layers.

What Is Keras Core? by ML GDE Aritra Roy Gosthipaty (India) and ML GDE Ritwik Raha (India) includes an introduction to Keras and it dives deep into Keras Core. They shared a custom training loop in JAX with Keras Core and harnessing the model.fit API to train the model.

Keras XLA Benchmarks by ML GDE Sayak Paul (India) presents an extensive benchmark around the vision models shipped in the Keras ecosystem (tf.keras.applications, TF Hub, and KerasCV).

Implementation of DeepLabV3+ for semantic segmentation in KerasCV by ML GDE Soumik Rakshit India (India) is a generalized implementation of DeepLabV3+, the latest model in the DeepLab family of semantic segmentation models.

Gen AI: the potential of BARD and the new features of Keras by ML GDE Nathaly Alarcon (Bolivia) was a mentoring for participants at I/O Extended La Paz. She covered various products, KerasCV, KerasNLP, Keras Core. TensorFlow, and etc.

How to Perform Image Augmentation With KerasCV & Text Classification With BERT and KerasNLP by ML GDE Derrick Mwiti (Kenya) explained how to perform image augmentation for an image classification project and how to fine-tune NLP models respectively.

At I/O Extended Silicon Valley 2023 hosted by GDG Silicon Valley, ML GDE Rajesh Bhat (United States) gave a talk titled Harnessing Transformers for various Computer Vision Tasks. He explained capabilities of the Transformer in revolutionizing computer vision tasks, different components of Transformer, and examples for image classification using Keras.

July ML meetup with TFUG Kolkata hosted by TFUG Kolkata was to make ML enthusiasts aware of the latest developments in Keras. One of the sessions was “Taking KerasNLP on a GenAI ride” and it explored deep into KerasNLP.

TensorFlow Kigali Kickoff hosted by TFUG Kigali was to discuss ethical considerations and societal impacts of AI technology. Through the event, the group raised awareness about TFUG and ML product updates including Keras and TensorFlow.

Kaggle

Introduction to Machine Learning with Google Colab with Pokemon Classification by ML GDE Kinni Mew (Hong Kong) was a hands-on workshop training a group of ICT teachers to build a Pokemon classifier using Kaggle dataset and Keras API.

ML GDE Luca Massaron (Italy) has participated in KaggleX BIPOC Mentorship Program (formerly known as the BIPOC Grant Program). The goal of this program is to increase representation, create career opportunities, and develop individual growth for BIPOC (Black, Indigenous, People of Color) people in the data science industry.

On-device ML

Bake a cake with TensorFlow — ML part by ML GDE George Soloupis (Greece) demonstrates the usage of an electronic nose (e-Nose) during the cake baking process. The custom device collected necessary data for training the final ML model. By establishing a connection between the e-Nose and an Android device, he built a real-time ML inference system and facilitated a comprehensive inspection of the cake baking procedure.

MediaPipe On-Device Machine Learning hosted by TFUG Durg introduced new aspects of ML with MediaPipe. They shared the capabilities of MediaPipe and how MediaPipe is making devices smarter.

TensorFlow Lite: Tflite Model Optimization for On-Device ML (Codelab) hosted by TFUG Ibadan explored the powerful capabilities of TF Lite to deploy ML models on edge devices. Attendees learned how to optimize models for efficient inference and resource usage, enabling them to run seamlessly on devices with limited computational resources.

Google I/O Extended Pwani 2023 hosted by GDG Pwani covered several ML topics including MediaPipe: Making On-Device Machine Learning a Piece of Pie.

At I/O Extended Surabaya 2023, TFUG Surabaya organizer Joan Santoso gave a talk, AI — Keras / Tensorflow Recommendations. At Google IO Extended Kisii 2023, Nairobi AI organizer Marvin Ngesa gave a talk on Diffusion Model With KerasCV. They shared practical experiences on how to enable diffusion models to generate novel yet coherent images and how to use KearsCV.

LLM / Generative AI

https://medium.com/media/8074c5186b1e4a75aa7cbeeedee1fb20/href

In Hands-on with the PaLM2 API to create smart apps, ML GDE Ruqiya Bin Safi (Saudi Arabia) shared PaLM API, Generative AI Support on Vertex AI, and prompt engineering best practices.

In Getting Started With Generative AI in Google Cloud, ML GDE Yannick Serge (Cameroon) covered generative AI, how it works, and Generative AI Studio UI & demos.

At I/O Extended Kuala Lumpur 2023 hosted by GDG Kuala Lumpur, ML GDE Kuan Hoong (Malaysia) explained how to fine tune LLM using Generative AI Studio and MakerSuite to build a prototype Bard app.

Google’s generative AI and LLMs trend and Comparison of Chinese large-scale natural language model PaLM2 and GPT by ML GDE Jerry Wu (Taiwan) shared Chinese versions of recent LLMs and Vertex AI features. He also shared his talk at Google I/O Extended’23@Taichung. This event also included ML GDE Tzer-jen Wei’s 7 cool things you can do with Generative AI session.

The Evolution and Impact of Large Language Models by ML GDE Mikaeri Ohana (Brazil) discussed the evolution and impact of LLMs, highlighting their crucial role in AI and NLP.

Community and Artificial Intelligence at Google I/O Connect by ML GDE Gema Parreño Piqueras (Spain) shared her experience at I/O Connect Amsterdam and PaLM API workshop.

Generative AI — Getting Started with PaLM 2 by ML GDE Sascha Heyer (Germany) explains how to get started with Generative AI and PaLM 2 on Vertex AI. He is writing a series of posts about generative AI on his Medium channel.

At Build LLM with Palms API hosted by GDG Waterloo, ML GDE David Cardozo (Canada) talked about how to access the advanced capabilities of LLM models like PaLM 2 with the PaLM API.

GenAI meetup at the Google Berlin office (event page of Practical GenAI Meetup) hosted by DevEcosystem Team Europe in collaboration with GDG Berlin & GenAI Berlin August Meetup. This event covered the latest tools including Generative AI Studio, Vertex AI, PaLM 2, and more. 450+ RSVPs and a long queue lined up to get into the event on the day.

GenAI Empower Education by ML GDE Sara EL-ATEIF (Morocco) was about the use of generative AI in education. The goal was to help participants in NEXT-GEN HACKATHON EMSI RABAT. She spoke about Bard and her recent hackathon project NeuroGuru built using PaLM 2 and Codey with Vertex AI.

I/O insights: the future of Google’s AI by 4 ML GDEs in Brazil, Arnaldo Gualberto, Bianca Ximenes, Fernanda Wanderley, and Pedro Gengo was a 2-hour panel discussion about I/O highlights, new releases, and community tips. Topics included Bard, PaLM2, KerasNLP, MST Scale, Vertex AI, speech-to-text models, accessibility, data laws, open source, and TFUG & developer communities.

Understanding Size Tradeoffs with Generative Models — How to Select the Right Model? by ML GDE Victor Dibia (United States) explains how to select the right LLM model for you. One of his approaches is to consider functional and non-functional requirements of business problems.

Async for LangChain and LLMs by ML GDE Gabriel Cassimiro (Brazil) shared how to make LangChain chains work with Async calls to LLMs, speeding up the time it takes to run a sequential long chain.

Generation Art: Exploring Generative Artificial Intelligence and Google Tools by ML GDE Lesly Zerna (Bolivia) was a talk about what is generative AI and its relation with deep learning. She covered I/O updates related to Gen AI, how to use Generative AI studio, and recommendations about responsible AI.

NLP in LLM era by ML GDE Junpeng Ye (China) was a talk about how LLMs have changed the NLP industry and how we should embrace the change.

Autonomous Agents with LLMs hosted by TFUG Singapore covered a number of LLM agent topics and what models, systems, and techniques have come out recently related to code and reasoning tasks. They also had a lightning talk about the mathematics underlying diffusion models.

ML Research

All Things ViTs: Understanding and Interpreting Attention in Vision by ML GDE Sayak Paul (India) and Hila Chefer from Google Research was presented at CVPR as a full-fledged tutorial.

BioBigBird by ML GDE Vasudev Gupta (India) showcased how to pre-train JAX/Flax versions of Bigbird model from HuggingFace transformers. He pre-trained using MLM objectives on TPU-v3–8.

Introduction of NNX library to ML GDEs by ML GDE David Cardozo (Canada) and ML GDE Cristian Garcia (Colombia) was a special session introducing Neural Networks for JAX (NNX) library, which they developed, to ML GDEs in a monthly sync.

How to improve RLHF performance by ML GDE Rumei LI (China) explains several methods to improve the performance of RLHF, including DeepMind’s Sparrow model.

Monitoring your Stable Diffusion fine-tuning with Neptune by ML GDE Pedro Gengo (Brazil) is about how using a tool like Neptune to log and organize your experiments can help in decision-making and allow the visualization of intermediate results as the training unfolds.

Paper Reading: Cramming- Training a Language Model on a Single GPU in One Day (paper) by ML GDE Dan lee (China) discussed how much performance a transformer-based language model can achieve in environments with limited computational complexity.

TensorFlow v FLAX: A Comparison of Frameworks by Wesley Kambale (Uganda) compared TensorFlow and Flax focusing on their features, functionality, advantages, and use cases.

Weekly ML Paper Reading Club — July — 01 hosted by TFUG Ibadan was a paper reading club session studying A multi-label learning model for psychotic diseases in Nigeria together.

15 lines of Python can beat the most advanced AI! by ML GDE Mathis Hammel (France) is a Twitter thread about the new paper from University of Waterloo using simple gzip kNN heuristics for text classification. His tweet got 3700+ Likes & 700K+ impressions.

Activities by ML Frameworks

#TensorFlow TensorFlow and CV (slides) by ML GDE Yu Chen (China) was a talk about the latest updates in TensorFlow, what machine vision is, how to implement machine vision with TF and optimization direction of TF model.

#TensorFlow At Leveling your skills with TensorFlow hosted by GDG Cúcuta, ML GDE Lesly Zerna (Bolivia) gave a lecture about foundations on TensorFlow, resources to learn it: codelab, TF guides, and ML crash course. Also adding information about TF recommendations for responsible AI.

#TensorFlow ML In Depth sessions from classification to deploying ML models by Taha Bouhsine (Morocco) were online hands-on workshops covering the basics of using deep learning for CV with TensorFlow. Participants learned basic knowledge and how to train their own models using TensorFlow.

#TensorFlow Machine Learning Bootcamp hosted by TFUG Hyderabad was a beginner-friendly event introducing the world of machine learning. They covered ML basics as well as how to use TensorFlow to build your own models.

#TensorFlow AI Fundamentals Bootcamp: Introduction to Machine Learning hosted by GDG Glasgow covered ML introduction, data preprocessing, model training, neural network, and TensorFlow.js.

#TensorFlow AI for Good: Building AI Systems for Fun and Profit, and Advancing Accessibility hosted by GDG Cape Town included transformative capabilities of TensorFlow.js and explore how it can revolutionize eyesight testing in remote areas where access to healthcare is limited.

#TensorFlow Diver Meal by Vasu Arora (India) is an app providing personal, nutritious, and diverse meal recommendations. He accomplished an image classification accuracy rate of 99.2% by utilizing transfer learning with Keras and TensorFlow.

#Cloud Google Cloud for Education: Strategy Segmentation using Generative AI by ML GDE Rubens Zimbres (Brazil) presented a generative and semantic approach to organize groups of Colombian schools in clusters. He used OCR to scan documents related to pedagogical projects, translated them to English, summarized and generated embeddings (LLM) and used unsupervised learning to generate groups of similar content.

#Cloud Recommendation Systems on Cloud Workshop — DevSummit 2023 by ML GDE Khongorzul Munkhbat (Mongolia) was a talk introducing the usage of learning cloud computing on the Cloud Skills Boost.

https://medium.com/media/3a804d1eec6aa4f5634b467aabab4efe/href

#Cloud Google Cloud in the Age of Generative AI by ML GDE Jéssica Costa (Brazil) introduced concepts to LLMs and showed how to get started in Generative AI Studio.

#Cloud #BigQuery Multivariate Time-Series Prediction with BQML (Korean version) by ML GDE JeongMin Kwon (Korea) is a post sharing the test results of new features and some of the ARIMA feature in BQML that are still valid briefly.

#Cloud #BigQuery Building ML models using BigQuery ML by ML GDE Nitin Tiwari (India) was a mentorship for second year engineering students. He guided the students on harnessing the power of BigQuery ML for building and deploying machine learning models for image classification tasks.

#Cloud #VertexAI Deploying ML Models in Google Cloud with Vertex AI by ML GDE Olayinka Peter (Nigeria) covered deploying ML models in Vertex AI as well as the basic TensorFlow model building process. The process involved creating and deploying a model and a model version, and then performing prediction with the model via a rest API.

#Cloud #DocumentAI Streamlining Business Operation with Document AI by ML GDE Ralph Regalado (Philippines) was a discussion about the value of Document AI to franchise owners on how they can use it to be more efficient in their business operation.

#Cloud #VertexAI Week 4 study session by ML GDE David Cardozo (Canada) in Certification Study Group — Professional ML Engineer was an office hour session hosted by GDG NYC to show Vertex AI, Vertex AI Pipelines and Flax.

#Cloud Hyperparameter Tuning with GCP AI Platform by ML GDE Imran Salam (Germany) presented 3 techniques which can be used to find the best set of the hyper parameters.

#SimpleML In I/O Extended Davao 2023: SimpleML for Sheets, ML GDE Ralph Vincent Regalado (Philippines) discussed and demonstrated SimpleML for Sheets. He gave a talk at I/O Extended Bogor 2023 as well.

Others

https://medium.com/media/030e81507c4eb942ad09e4f7de737dd7/href

Chat with ML GDE Hannes by ML GDE Margaret Maynard-Reid (United States) is a video chatting with ML GDE Hannes Hapke: how he became an ML GDE, what he likes the most about being an ML GDE, and his work and community projects.

TFUGCbe Reboot 2023 hosted by TFUG Coimbatore aimed to revitalize the TFUG Coimbatore community by bringing together ML professionals, researchers, developers, and students. The event focuses on knowledge expansion, networking, and collaboration within the TensorFlow ecosystem. Topics include AI & remote sensing for sustainability, generative AI (Stable Diffusion), and etc.

I/O Extended Kumasi hosted by GDG Accra & Zindi Ghana discussed the intricacies of ML explainability and interpretability using Vertex AI, exploring how transparency can demystify those AI models and make them more accessible to all.

Power Women in AI/ML at Google NYC hosted by GDG NYC talked about design security infrastructure for generative AI tooling and how to build a closed-domain information retrieval system using Flan-T5.

IO Extended: Intro to Large Language Models with the PaLM API and MakerSuite hosted by GDG Detroit shared how to use MakerSuite, how to develop with the PaLM API. Googler Josh Gordod participated in this workshop as a speaker and trainer.

[MLDP Newsletter] July 2023 — Machine Learning Communities: 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.