Re:Invent 2020 is just one week away. What makes this year’s Re:Invent special is that it will be free and you can watch them online from the comfort of your home. Hence you do not need to walk for miles within hotels to jump between sessions. Still, there will be tons of AI/ML sessions for you ML Practitioner/Data Scientist which makes it hard to build your ideal schedule.
Without further ado, here are some guides which I hope will be useful for you to make your decision.
We all know that any AI/ML project starts with data preparation, which most often then not takes more time than building the AI model itself. So, do not miss this session.
How the NFL builds computer vision training datasets at scale
In this session, learn how the National Football League (NFL) uses Amazon SageMaker Ground Truth to build training datasets at a fraction of the cost so it can track all 22 players as they move on the field.
With the ever-growing number of model architecture/libraries to choose from, sometimes it is hard for us to stay on top of them all. From the installation of these libraries, downloading the pre-trained weights, fighting through the framework dependencies/drivers, getting them to run and knowing how to fine tune them. These sessions will show you how to bootstrap your model development and to remove the heavy lifting associated in building an AI model.
Get started with Amazon SageMaker in minutes
An introduction to Amazon SageMaker to accelerates your ML journey by offering a dataset and model for common ML use cases, such as predictive maintenance, churn prediction, and credit risk, so you can move quickly from concept to production.
Choose the right machine learning algorithm in Amazon SageMaker
AWS offers many choices for solving business problems through machine learning (ML), ranging from built-in algorithms, frameworks, and more in ML services down to models and algorithms available via AWS Marketplace. Amazon SageMaker supports 17 built-in ML algorithms, such as classification, regression, and recommendation. Built-in algorithms are easy to use, and they are optimized for speed, scale, and accuracy. In this session, learn how to choose the right built-in algorithm for your business problem. This session categorizes these algorithms by problem types and dives deep into popular ones.
Train and tune ML models to the highest accuracy using Amazon SageMaker
Amazon SageMaker Studio is the first fully integrated development environment (IDE) for machine learning (ML). In this session, see how easy it is to train and tune models in SageMaker Studio. Also, see how to isolate and measure the impact of changing datasets, algorithm versions, and model parameters as you train, so ML models can be trained more quickly, at low costs, and with higher accuracy. Finally, learn how to browse active experiments, search for previous experiments, and compare experiment results visually in SageMaker.
How to use fully managed Jupyter notebooks in Amazon SageMaker
Managing compute instances to view, run, or share a notebook is tedious. Amazon SageMaker offers several choices to use Jupyter notebooks, including Amazon SageMaker Studio. SageMaker Studio notebooks are one-click Jupyter notebooks that you can spin up quickly. The underlying compute resources are fully elastic, so you can easily dial up or down the available resources, and the changes take place automatically in the background without interrupting your work. You can also easily share notebooks with others, making collaboration easy and scalable. In this session, see a demo of SageMaker Studio and other ways to use Jupyter notebooks for building machine learning models.
Accelerate machine learning projects with pretrained models
Insufficient enterprise AI adoption often happens due to lack of time, data, and skills required to develop ML models for solving your business problems. In this session, learn how pretrained ML models — available in AWS Marketplace and deployed with Amazon SageMaker — can help you quickly add new ML features in your applications and enable you to prove the value of powering your applications with AI to your leadership. Learn how to explore, test, deploy, and integrate ML models securely in your existing production application.
Interpretability and AutoML
As many of you all aware, building a model with high accuracy sometimes is not enough when you cannot explain how the model come up with such prediction. Model interpretability is a popular topics in AI community. Model architectures are getting more complex and less interpretable by day. This problem is becoming more prominent with the rise of AutoML (using an AI to build the AI model for you).
At the Re:Invent 2019, AWS has launched several innovative technologies to address this problem such as Amazon SageMaker Debugger and Amazon SageMaker Autopilot, which will be covered in more details by the following sessions.
Get deep insights about your ML models during training
Amazon SageMaker Debugger provides complete insight into the training process by automating data capture and analysing training runs so you can interpret and explain how models make predictions and also identify problems. In this session, walk through how to use the real-time training metrics and set up alerts so you can reduce troubleshooting time and improve model quality.
Interpretability and explainability in machine learning
This session explores the science behind interpretability and shows how to use Amazon SageMaker to understand the workings of ML models and their predictions.
Build quality ML models easily & quickly with Amazon SageMaker Autopilot
This session covers how to use Amazon SageMaker Autopilot to automatically build the best machine learning (ML) model for your data and interpret your model’s predictions with just a few clicks. Amazon SageMaker Autopilot automatically trains and tunes hundreds of ML models to optimize your success metric and creates the model that best suits your data. In this session, data analysts can learn to quickly extract insights from data using Amazon SageMaker Autopilot to create and deploy ML models without any ML expertise. Data scientists can learn to use Amazon SageMaker Notebooks to understand how the model was generated and recreate it at any point in the future.
One step which is very often overlooked in AIDLC (AI Development Life Cycle) is the productionisation. It is often the most time consuming part which consist of optimising your model for scalability/cost, building the continuous training pipeline, monitoring model performance in production and last but not the least, building and securing the infrastructure. These talks below will provide you some guidance on the above.
How to choose the right instance type for ML inference
AWS offers a breadth and depth of machine learning (ML) infrastructure you can use through either a do-it-yourself approach or a fully managed approach with Amazon SageMaker. In this session, explore how to choose the proper instance for ML inference based on latency and throughput requirements, model size and complexity, framework choice, and portability. Join this session to compare and contrast compute-optimized CPU-only instances, such as Amazon EC2 C4 and C5; high-performance GPU instances, such as Amazon EC2 G4 and P3; cost-effective variable-size GPU acceleration with Amazon Elastic Inference; and highest performance/cost with Amazon EC2 Inf1 instances powered by custom-designed AWS Inferentia chips.
Detect machine learning (ML) model drift in production
In this session, explore how state-of-the-art algorithms built into Amazon SageMaker are used to detect declines in machine learning (ML) model quality. One of the big factors that can affect the accuracy of models is the difference in the data used to generate predictions and what was used for training. For example, changing economic conditions could drive new interest rates affecting home purchasing predictions. Amazon SageMaker Model Monitor automatically detects drift in deployed models and provides detailed alerts that help you identify the source of the problem so you can be more confident in your ML applications.
Secure and compliant machine learning for regulated industries
As organizations bring their machine learning (ML) workloads to the cloud, having access to an environment that is secure is the top requirement. In this session, learn the steps involved in provisioning a secure ML environment on Amazon SageMaker, dive deep into common customer patterns and architectures, and see how to leverage other AWS services to build these environments in a consistent and reproducible manner.
Scaling MLOps on Kubernetes with Amazon SageMaker Operators
Scaling machine learning (ML) for training and inference on Kubernetes can be time-consuming to set up and requires overhead to maintain production-level uptime. In this session, learn how to run training at scale on AWS with minimal setup using Amazon SageMaker Operators for Kubernetes. Also, learn how set up secure, high-availability endpoints with autoscaling policies. Additionally, Intuit shares how it built a Kubernetes controller that wraps around operators using SageMaker, Apache Spark, and Argo to run world-class, high-availability systems.
Implementing MLOps practices with Amazon SageMaker
MLOps practices help data scientists and IT operations professionals collaborate and manage the production machine learning (ML) workflow, including data preparation and building, training, deploying, and monitoring models. During this session, explore the breadth of features in Amazon SageMaker that help you increase automation and improve the quality of your end-to-end workflows.
Productionizing R workloads using Amazon SageMaker, featuring Siemens
R language and its 16,000+ packages dedicated to statistics and ML are used by statisticians and data scientists in industries such as energy, healthcare, life science, and financial services. Using R, you can run simulations and ML securely and at scale with Amazon SageMaker, while reducing the cost of development by using the fully elastic resources in the cloud. Learn how Siemens Energy, a technology provider for more than one-sixth of the global electricity generation, with more than 91,000 employees and presence in more than 90 countries, is enabling new digital products with Amazon SageMaker to build, train, and deploy statistical and ML models in R at scale.
These talks are not in any of the above categories, however they are equally interesting for some of you who are into deep learning and are building complex models and are constantly hitting the GPU RAM limit.
Train large models with billions of parameters in TensorFlow 2.0
When training with large models or large inputs, a single GPU’s memory is often a limiting factor during training, producing out-of-memory (OOM) errors. The solution to this problem is model parallelism, which places layers or operations of a model on multiple GPU devices to utilize the aggregate memory of all GPUs in the cluster. But implementing model parallelism efficiently is hard, time-consuming, and requires expertise in graph partitioning and setting up execution pipelines. In this session, learn how to train large models with billions of parameters using Amazon SageMaker and TensorFlow 2.0.
Fast distributed training and near-linear scaling with PyTorch on AWS
Efficiently scaling PyTorch training jobs to hundreds or thousands of GPUs is challenging. This session presents the most common issues and bottlenecks data scientists encounter with multi-node training. Learn how AWS helps solve these challenges and removes network bottlenecks with optimizations to PyTorch and AllReduce algorithms that are specific to AWS architecture. Finally, learn how to reduce training time and lower costs while maintaining high scaling efficiency.
Ok, thats it folks. Hopefully you enjoy Re:Invent 2020!