![[mldp-newsletter]-mar-2023 — machine-learning-communities:-highlights-and-achievements](https://prodsens.live/wp-content/uploads/2023/04/10643-mldp-newsletter-mar-2023-machine-learning-communities-highlights-and-achievements-550x223.png)
[MLDP Newsletter] Mar 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!
ML Campaigns
MediaPipe/TF Recommenders Sprint
ML Community Sprint is a campaign, a collaborative attempt bridging ML GDEs with Googlers to produce relevant content for the broader ML community. Throughout Feb and Mar, MediaPipe/TF Recommenders Sprint was carried out and 5 projects were completed.
- Tweet Image Maker — Learning to recommend engaging images for tweets by ML GDE Victor Dibia (USA): a slide deck and a Kaggle Notebook
- Training a recommendation model with dynamic embeddings by ML GDE Thushan Ganegedara (Australia): a slide deck and a Github repository
- Easy Image Segmentation in Android with MediaPipe, TFLite Interpreter or Task Library by ML GDE George Soloupis (Greece): a slide deck and a blog posting
- MediaPipe Intro — Image Classification & Embedding by ML GDE Margaret Maynard-Reid (USA): a slide deck and examples on GitHub
- MediaPipe Web APIs — Image segmentation by ML GDE Henry Ruiz (USA): a slide deck and a GitHub repository
ML Story
ML Story is a monthly blog article series specifically focusing on the usage and implementation of ML with Google ML products to help them be more widely known and used.
- The first ML Story is DreamBoothing Your Way into Greatness by ML GDE Sayak Paul (India).
ML Olympiad 2023
The second, ML Olympiad 2023 has wrapped up successfully with 17 competitions and 300+ participants addressing important issues of our time — diversity, environments, etc.
Featured Activities by Focus Areas
Keras
Lighting up Images in the Deep Learning Era by ML GDE Soumik Rakshit (India), ML GDE Saurav Maheshkar (India), ML GDE Aritra Roy Gosthipaty (India), and Samarendra Dash. It is an implementation of SoTA image and video restoration models for tasks such as low-light enhancement. The article also talks about Restorers, which is a library providing out-of-the-box TensorFlow and Keras implementations of SoTA image and video restoration models for low-light enhancement, denoising, deblurring, super-resolution, etc.
Keras Implementation of NeurIPS 2021 paper, Augmented Shortcuts for Vision Transformers by Usha Rengaraju (India).
Image Classification Using Vision Transformer and KerasCV by ML GDE Ayush Thakur (India) covered how one can fine-tune ViT using KerasCV on a custom dataset.
Create VAE from scratch using Keras and TensorFlow by ML GDE Derrick Mwiti (Kenya) explains how to generate images with Variational Autoencoders (VAE).
On-device ML
AI for Art and Design by ML GDE Margaret Maynard-Reid (United States) delivered a brief overview of how AI can be used to assist and inspire artists and designers in their creative space. She also discussed a few use cases of on-device ML for creating artistic Android apps.
ML Engineering (MLOps)
Textual Inversion Pipeline for Stable Diffusion by ML GDE Chansung Park (Korea) demonstrates how to manage multiple models and their prototype applications of fine-tuned Stable Diffusion on new concepts by Textual Inversion.
Various ways of serving Stable Diffusion by ML GDE Chansung Park (Korea) and ML GDE Sayak Paul (India) shares how to deploy Stable Diffusion with TF Serving, Hugging Face Endpoint, and FastAPI.
TFX T1 | E6 — Building Deep Learning Pipelines with TFX and Vertex AI (video) by TFUG São Paulo shared a session about how to put your TFX pipelines to run inside Vertex AI using Vertex AI pipelines component.
ChatGPT for Accounting: How Digits is using Generative Machine Learning to transform finance by ML GDE Hannes Hapke (United States) explores generative machine learning and what it can bring to the accounting world. He trained his model through TFX running on Vertex AI.
Two Towers Model: A Custom Pipeline in Vertex AI Using Kubeflow by ML GDE Rubens Zimbres (Brazil) shows how to create a Kubeflow pipeline in Vertex AI for the Two Towers model, in Keras, for CI/CD.
Running a Stable Diffusion Cluster on GCP with tensorflow-serving (Part 1 | Part 2) by ML GDE Thushan Ganegedara (Australia) explains how to set up a GKE cluster, how to use Terraform to set up and manage infrastructure on GCP, and how to deploy a model on GKE using TF-serving.
Scalability of ML Applications by TFUG Bangalore focused on the challenges and solutions related to building and deploying ML applications at scale. Googler Joinal Ahmed gave a talk entitled Scaling Large Language Model training and deployments.
Responsible AI
How Responsible AI plays a predominant role in the Healthcare Industry? by ML GDE Sharmistha Chatterjee (India) introduced the implications of responsible AI in the healthcare industry and demonstrated with examples of how responsible AI plays a vital role in designing robust AI/ML models especially when we are hit with uncertainties like COVID-19.
Fairness & Ethics In AI: From Journalism, Medicine and Translation by ML GDE Samuel Marks (United States) discussed responsible AI.
Women in AI/ML at Google NYC by GDG NYC discussed the hottest topics, including LLMs and generative AI. Googler Priya Chakraborty gave a talk entitled Privacy Protections for ML Models.
ML Research
Jax & Efficient Object Detection by ML GDE David Cardozo (Canada) talks about object detection using JAX. JAX Streams: Parallelism with Flax | EP4 by him and ML GDE Cristian Garcia (Columbia) explored Flax’s new APIs to support parallelism.
Sentiment after the Earthquake disaster in Turkey by ML GDE Duygu ALTINOK (Turkey) examined the online comments which were uploaded after the earthquake in Turkey and identified the emotions involved in the comments. She used FLAN-T5 for labeling the dataset with sentiment and emotion labels. After that, she trained a more compact NN for sentiment analysis with TF.
March Machine Learning Meetup by TFUG Kolkata. Two sessions were delivered: 1) You don’t know TensorFlow by ML GDE Sayak Paul (India) presented some under-appreciated and under-used features of TensorFlow. 2) A Guide to ML Workflows with JAX by ML GDE Aritra Roy Gosthipaty (India), ML GDE Soumik Rakshit (India), and ML GDE Ritwik Raha (India) delivered on how one could think of using JAX functional transformations for their ML workflows.
A Guide to Machine Learning Workflows with JAX by ML GDE Soumik Rakshit (India) shared the evolution of JAX & its power tools and a guide to writing efficient ML workflows using JAX and Flax.
A paper review of PaLM-E: An Embodied Multimodal Language Model by ML GDE Grigory Sapunov (UK) shared details of the model. He also shared his slide deck about NLP results in 2022.
An annotated paper of On the importance of noise scheduling in Diffusion Models by ML GDE Aakash Nain (India) outlined the effects of noise schedule on the performance of diffusion models and strategies to get a better schedule for optimal performance.
LLM-related
Google Colab release for people to try the Flan-UL2 20B by ML GDE Sam Witteveen (Singapore) made it accessible to people to try it. His Tweet got over 80K views on Twitter, over 500 likes, and over 100 RTs. The Colab notebook was opened over 2400 times. Sam also opened his Youtube channel exploring ML and AI. This channel got more than 6K subscribers within a month.
The Evolution of Large Language Models: from BERT, GPT to LaMDA and LLaMA by ML GDE Kuan Hoong (Malaysia) explored the key milestones in the development of LLMs, from the first neural language models to the current state-of-the-art models.
What’s the largest language model and how to use it by ML GDE Jerry Wu (Taiwan) was a session for a group of students at the University of Taipei. Jerry shared about LLMs, including BERT, T5, GPT, PaLM, and how LLM works and how to use it in business.
Alpaca-LoRA as a Chatbot service by ML GDE Chansung Park (Korea) has reached 1400+ stars on GitHub. This project boosted large interest in the deployment of ML models and helped democratizing ML serving in the large language model’s era.
Cabrita: A portuguese finetuned instruction LLaMA by ML GDE Piero Esposito (Brazil) and ML GDE Pedro Gengo (Brazil). Following the idea of Stanford Alpaca and using LoRA technique, they fine-tuned the LLaMA 7B model translating a dataset to Portuguese, and used Colab for the training step.
Vietnamese Self-Supervised Learning Wav2Vec2 model by ML GDE Binh Nguyen (Germany) was a project using 12K hours of Vietnamese speech data from YouTube. He used the data to train a pretrained self-supervised speech representation model.
Discover the power of GEN AI & LLMs at the first TF2023 Community Event (video) by TFUG Colombia was a talk about the different techniques of LLMs and generative AI.
GPT Models: The Future of Conversational AI, Image Generation, and Code Generation by TFUG Chandigarh covered the basics of language modeling, Transformer architecture that powers models like GPT, GPT architecture, and its derivatives like DALLE-2, Codex, etc.
Featured Activities by ML Frameworks
#TensorFlow.js Real-time Object Detection in the browser with YOLOv7 and TF.JS by ML GDE Hugo Zanini (Brazil) was announced as a #TFCommunitySpotlight🏅 winner in March.
#TensorFlow Advent of Code 2022 in pure TensorFlow — Day 9 & Day 11 by ML GDE Paolo Galeone (Italy) and ML GDE Tzer-jen Wei (Taiwan). Paolo also wrote Day 10 & Day 11 articles using ChatGPT.
#TensorFlow Deep Asteroid & NASA : Predicting Asteriod Impact with Tensorflow by ML GDE Gema Parreño Piqueras (Spain) was a presentation of the TensorFlow project awarded by NASA for predicting asteroid impact.
#TensorFlow Building ranking models powered by multi-task learning with Merlin and TensorFlow by ML GDE Gabriel Moreira (Brazil) describes how to build TensorFlow models with Merlin for recommender systems using multi-task learning.
#TensorFlow Combining Decision Forest and Neural Network models with TensorFlow by ML GDE Radostin Cholakov (Bulgaria) explained the basics of TF-DF for beginners and more advanced approaches for training these models using the hidden states of a neural network.
#VertexAI Stable Diffusion Finetuning by ML GDE Pedro Gengo (Brazil) and ML GDE Piero Esposito (Brazil) is a fine-tuned Stable Diffusion 1.5 with more aesthetic images. They used Vertex AI with multiple GPUs to fine-tune it. It reached Hugging Face top 3 and more than 150K people downloaded and tested it.
#VertexAI Movie review classification on Vertex AI by ML GDE Thushan Ganegedara (Australia) was a workshop sharing how to build a pipeline to classify movie reviews on Vertex AI. They used the dataset to train an ML model in two ways; AutoML and uploading a custom version of BERT trained on the data (locally) to GCS.
#VertexAI Machine Learning Cloud Platforms & Other Super Fun Stuff by TFUG Singapore talked about doing ML on cloud platforms. Cloud AI Overview by Googler Thu Ya Kyaw introduced Cloud AI and Vertex AI’s features & capabilities.
#VertexAI Prepare and manage your datasets with Vertex AI Datasets by GDG Cloud Milano shared how to prepare and manage data to train an AutoML model using Vertex AI Datasets and Vertex AI AutoML.
#VertexAI #BigQuery Scientific Data Processing: Vertex AI and TensorFlow by GDG Cloud Twin Cities was a hands-on event featuring two labs: 1) Analyzing Natality Data Using Vertex AI and BigQuery and 2) Predict Baby Weight with TensorFlow on AI Platform.
#LaMDA Google’s technologies and researches on Generative AI by ML GDE Chansung Park (Korea) was a speech about generative AI and Google Cloud, including LaMDA and Bard. The main focus was emphasizing the importance and difficulties of deploying a large language model.
#GCP #BigQuery How to start with ML on GCP — Practical Workshop by ML GDE Nathaly Alarcón (Guatemala) was a workshop for a group of university students. She talked about the options to apply ML models using Google Cloud. They also worked on a demo using ML APIs and BigQuery ML.
#GCP AI and Google Cloud at NCR by University of Michigan — Dearborn, MI shared how NCR, the global POS software provider, uses GCP and how they adopt AI in their business.
Others
Open-sourcing 100+ master’s notes & cheatsheets on machine learning, data science and more by ML GDE Merve Noyan (France)
Dev Library articles by ML GDE Rubens Zimbres (Brazil) was highlighted on Google Dev Twitter.
[MLDP Newsletter] Mar 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.
