Digest #2019.09.30 – Making the Most of Machine Learning

30 September 2019

Understanding Fairness in Machine Learning – This article is a great reminder and defense for the statement, “the data speaks for itself.” Biases in training models affect the results of analyses. It is essential to understand how our models make decisions to tackle this bias by adding more balanced training data. Knowledge of biases in our training models also help in adjusting our training loss function or adjusting prediction thresholds to account for the type of fairness we want to work towards.  

Ways in which AI and ML are improving endpoint security – Take a look at the predicted billions of dollars that will be pouring in to improve cybersecurity. “Cloud platforms are enabling AI and machine learning-based endpoint security control applications to be more adaptive to the proliferating types of endpoints and corresponding threats.” The article sheds interesting light on how AI and ML can revolutionize the field of cybersecurity and improve endpoint security.

AI gears up for data analysis – The article explores how the use of Machine Learning and AI can help scientific research with the potential to improve data analyses speeds and accuracy of results dramatically. Machine learning can change scientific experiments, improve results, and significantly reduce production failures in different industries. “By looking at the long-term patterns [in the signals], you can actually spot imminent failures. One example could be a gradual increase in motor operating temperature, which may indicate that an actuation unit is on its way to overheating.” While investing in AI and ML and gathering data is resource-consuming, the returns and rewards prove to be well worth it.

Digest #2019.09.23 – Machine Learning is the word…

23 September 2019

MLOps, Rise of the Term: Most of us by now have heard this word frequenting around; MLOps, or Machine Learning Operations.. This is an interesting article on the rise of the term and the challenges actually faced by teams working cross-functionally and dealing with Machine Learning. It talks about the limitation of managed solutions for Machine Learning and hits close to home for Kubeflow. While working on Kubeflow and being in the community is exciting, we’ve only scratched the surface still have a long way to go. There’s also this funny info-graphic below about data scientists emailing Jupyter Notebooks to Developers for production – For real!

Artificial Intelligence to Improve Enterprise Storage?: This is an interesting article that explores the use of AI to improve storage operations in the enterprise. With data prevailing everywhere, petabytes and zettabytes of 0s and 1s, the lives of storage systems admins are becoming increasingly difficult. “When I’m trying to understand the IO pattern of an application or a workflow, one technique that I use is to capture the strace of the application, focusing on IO functions. For one recent application I examined, the strace output had more than 9,000,000 lines, of which a bit more than 8,000,000 lines were involved in IO. Trying to extract an IO pattern from more than 8,000,000 lines is a bit difficult.” – Jeff Layton, Nvidia Solutions Architect. Now, these tasks seem impossible for our brains but this where the mighty AI comes in. Machines champion taking large volumes of data and synthesizing it to useful chunks. Leveraging AI and ML can truly change the way we deal with, understand and use Enterprise storage. 

Apple is Building ML to Rule Them All: With every tech giant and pretty much every other business getting in on Machine Learning, it’s established that ML is the next big thing in the industry. Apple is coming in heavy on Machine Learning with something called Overton. “Apple claims to have a first-of-its kind solution with Overton – it aims to enable much of the personalization of ML models to be administered by the machine, not the human.” In general terms they are creating a Machine Learning management system; the system itself learns from the environment, responds to external factors and heals and corrects the models automatically when needed. Privacy concerns aside, our Siri responses, animojis and pet portraits are going to get snazzy – among other things…

Digest #2019.09.16 – The State of AI and ML

16 September 2019

• Machine Learning and AI in 2019: A recent survey conducted by Dresner Advisory Services shows Machine Learning and AI to rank as highest priority for enterprises. R&D, Marketing, Sales, Insurance, Fintech, Telco, Retail and Healthcare enterprise rank machine learning as their biggest bet and believe it is critical to their success. “2019 is a record year for enterprises’ interest in data science, AI, and machine learning features they perceive as the most needed to achieve their business strategies and goals.”

• Using Machine Learning in health-tech: With humans becoming increasingly health conscious and risk-averse, we’re seeing a boom in health-tech. Machine Learning is staying on top of the game here as well; researchers at MIT have invented a cardiovascular risk identifier. With heart disease being the most common cause of death in the world, the system called ‘CardioRisk’ uses a patient’s raw electrocardiogram (ECG). Using Machine Learning techniques the ECG is analysed against datasets and the system produces a risk score that places the patient in a relative risk category. “The intersection of machine learning and healthcare is replete with combinations like this — a compelling computer science problem with potential real-world impact.”

• Training on Large Images Using Spatial Partitioning on Cloud TPUs: The pain of training your Machine Learning models without enough space on a single chip can now be put to ease – You can now leverage a new spatial partitioning capability on cloud TPUs. This makes it possible to split up a single model across several TPU chips allowing processing of much larger input data sizes. Use this guide to learn how to configure spatial partitioning properly for your applications.

Digest #2019.09.09 – Machine Learning on a Budget

9 September 2019

• How To Develop Successful Machine Learning Projects On A Budget – A quick journey through some of the principles for a successful AI getting started project. The article includes an example of how to go from nothing to something – from data pipeline creation to models in production. The primary focus is on a model for hiring and tools you could use, like Kubeflow. How much money will it take to get something off the ground? Read on ..

• New homepage and improved collaboration features for AI Hub – Google’s AI Hub is powered by Kubeflow. This announcement discusses new features added to AI Hub. Although still in beta, AI Hub provides a view into how Kubeflow can be leveraged on prem. Do you have platform engineers building your corporate AI toolkit? Read this article to see what Google is doing. Advanced multi-user use cases, new machine learning taxonomy, asset favoriting, and public data sets or solutions – including a TensorRT-optimized BERT notebook from Nvidia.
• Three pitfalls to avoid in machine learning – As scientists from myriad fields rush to perform algorithmic analyses, Google’s Patrick Riley calls for clear standards in research and reporting. Machine-learning tools can turn up fool’s gold — false positives, blind alleys and mistakes. Read this article to learn about three problems in machine-learning analyses that Google faced, and solved, in the Google Accelerated Science team.
• Use case spotlight: Little Ripper deploys croc-spotting AI drones – A different take on how AI is saving lives. To help keep beachgoers safe from crocodiles 🐊 in the water and on land, the same AI drone technology that the Little Ripper Group used for its shark 🦈 detection drones is now being used to spot crocodiles in Queensland. In additino to spotting wildlife, this technology is used to help swimmers. Last summer, 51 drones were deployed around Australia to help spot rips and swimmers in distress.

Digest #2019.09.02 – The why of Kubeflow

2 September 2019

• AI Tales: Building Machine learning pipeline using Kubeflow and Minio – Understand the Kubeflow value proposition in an entertaining format. The story starts with Joe, the neighbourhood Machine learning enthusiast. Joe reads a few things, becomes an expert, and then the real fun begins. He quickly runs into problems with portability, DevOps, scaling, performance, and cost. Enter Kubeman (or Kubeperson?), who personifies Kubeflow, and saves the day!

• Basics of Data Science Product Management: The ML Workflow – Another look at the complicated space that Kubeflow helps solve. “Something I quickly learned was that managing ML products is difficult because of the complexities and uncertainties involved with the different steps in the machine learning workflow” – (1) Review of related literature; (2) Data gathering & processing; (3) Model training, experimentation, & evaluation; (4) Deployment
• Hardware Science: Researchers demonstrate all-optical neural network for deep learning – In a key step toward making large-scale optical neural networks practical, researchers have demonstrated a first-of-its-kind multilayer all-optical artificial neural network. The next generation of artificial intelligence hardware will be much faster and exhibit lower power consumption compared to today’s computer-based artificial intelligence.
• Hardware Science: Quantum computing should supercharge this machine-learning technique – Researchers from IBM and MIT show how an IBM quantum computer can accelerate a specific type of machine-learning task called feature matching. Feature matching is a technique that converts data into a mathematical representation that lends itself to machine-learning analysis. Using a quantum computer, it should be possible to perform this on a scale that was hitherto impossible.

Digest #2019.08.26 – Kubeflow basics, TensorFlow 2.0

26 August 2019

• Kubeflow for Poets – This article introduces the core concepts necessary to understand all of the moving pieces in a Kubeflow based machine learning Pipeline. It includes a brief introduction to microservices, Docker, Kubeflow, Kubernetes, virtualisation, Google cloud and more. Read this article for step-by-step low level interaction with every component in the toolchain, from dockerfiles to GKE.

• TensorFlow 2.0 RC Available – Take advantage of the next wave of innovation for TensorFlow. The 2.0 release is a major milestone and brings a focus on ease of use and simplification. It comes with many features to support this focus – eager execution, tighter Keras integration, loadable TF-Hub models, end-to-end ML pipelines with TFX, and much more.
• How AI Is Changing The Game For Recruiting – In this use case spotlight, we review how machine learning toolkits like Kubeflow and AI are changing the recruiting industry. Talent acquisition is expensive, and getting it wrong is more expensive. AI can helping improve talent acquisition efficiency and effectiveness. From finding the right candidates to creating and delivering the right message to those candidates, read this article to discover more.

How AI is changing Recruiting

20 August 2019

How AI Is Changing The Game For Recruiting – In this use case spotlight, we review how machine learning toolkits like Kubeflow and AI are changing the recruiting industry. Talent acquisition is expensive, and getting it wrong is more expensive. AI can helping improve talent acquisition efficiency and effectiveness. From finding the right candidates to creating and delivering the right message to those candidates, read this article to discover more.

Digest #2019.08.19 – Kubeflow at CERN

19 August 2019

• Replicating Particle Collisions at CERN with Kubeflow – this post is interesting for a number of reasons. First, it shows how Kubeflow delivers on the promise of portability and why that matters to CERN. Second, it reiterates that using Kubeflow adds negligible performance overhead as compared to other methods for training. Finally, the post shows another example of how images and deep learning can replace more computationally expensive methods for modelling real-word behaviour. This is the future, today.
• AI vs. Machine Learning: The Devil Is in the Details – Need a refresh on what the difference is between artificial intelligence, machine learning and deep learning? Canonical has done a webinar on this very topic, but sometimes a different set of words are useful, so read this article for a refresh. You’ll also learn about a different set of use cases for how AI is changing the world – from Netflix to Amazon to video surveillance and traffic analysis and predictions.

• Making Deep Learning User-Friendly, Possible? – The world has changed a lot in the 18 months since this article was published. One of the key takeaways from this article is a list of features to compare several standalone deep learning tools. The exciting news? The output of these tools can be used with Kubeflow to accelerate Model Training. There are several broader questions as well – How can companies leverage the advancements being made within the AI community? Are better tools the right answer? Finding a partner may be the right answer.
• Interview spotlight: One of the fathers of AI is worried about its future – Yoshua Bengio is famous for championing deep learning, one of the most powerful technologies in AI. Read this transcript to understand some of his concerns with the direction of AI, as well as the exciting developments in AI. Research that is extending deep learning into things like reasoning, learning causality, and exploring the world in order to learn and acquire information.

Digest #2019.08.12 – The Kubeflow Machine Learning Toolkit

12 August 2019

• Kubeflow — a machine learning toolkit for Kubernetes – An introduction to Kubeflow from the perspective of a data scientist. This article quickly runs through some key components – Notebooks, Model Training, Fairing, Hyperparameter Tuning (Katib), Pipelines, Experiments, and Model Serving. If you are looking for a quick overview, give this article a go. Here’s a key diagram from the article:

• Why is it So Hard to Integrate Machine Learning into Real Business Applications? – For teams just getting started, getting a trained model with sufficient accuracy is success. But that is just the starting point. There are many engineering and operational considerations that remain to be done. There are components that need to be built, tested and deployed. This post presents a real customer AI-based application, explaining some of the challenges, and suggests ways to simplify the development and deployment.
• Further afield – Techniques to improve the accuracy of your Predictive Models – a look at few techniques to improve the accuracy of your predictive models. The code base is in R, but the principles are applicable to a variety of code bases and algorithms.
• Use case spotlight – https://www.technologyreview.com/s/614043/instead-of-practicing-this-ai-mastered-chess-by-reading-about-it/. Instead of practicing, this AI mastered chess by reading about it. The chess algorithm, called SentiMATE, was developed by researchers at University College London. It evaluates the quality of chess moves by analysing the reaction of expert commentators. These learning techniques could have many other applications beyond chess – for instance, analysing sports, predicting financial activity, and making better recommendations.

Digest #2019.08.05 – Kubeflow 0.6 Release

5 August 2019

• Kubeflow v0.6: support for artifact tracking, data versioning & multi-user – version 0.6 includes several enterprise features to support multiple users and better model training pipelines. For multiple users, Kubeflow v0.6 provides a flexible architecture for user isolation and single sign-on. For data, enhancements have been added to Kubeflow Pipelines and jupyter. In total, over 250+ merged pull requests!
• Kubeflow Components and Pipelines – This is a useful article that demonstrates how to build pipelines, how to create and use components, and how to leverage them inside of notebooks. Specifically, how to run pipelines and experiments inside of a Notebook. The article also includes easy to understand, ready to use examples.
• Use Case Spotlight – The world’s most-advanced AI can’t tell what’s in these photos. Can you? A useful exploration into the the current limits of AI in visual recognition. Researchers from UC Berkeley, the University of Washington, and the University of Chicago are building the ultimate archive of photos that confuse AI. AI is judged by its answers .. we can get all sorts of unexpected bias from AI. This poses a major problem when AI systems are being used in technology like autonomous cars or fields like criminal justice.

Digest #2019.07.29 – Kubeflow Releases so far (0.5, 0.4, 0.3)

29 July 2019

• Kubeflow 0.5 simplifies model development with enhanced UI and Fairing library – The 2019 Q1 release of Kubeflow goes broader and deeper with release 0.5. Give your Jupyter notebooks a boost with the redesigned notebook app. Get nerdy with the new kfctl command line tool. Power to the people – use your favourite python IDE and send your model to a Kubeflow cluster using the Fairing python library. More training tools added as well, with an example of XGBoost and Fairing.
• Kubeflow 0.4 Release: Enhancements for Machine Learning productivity – The 2018 Q4 release will supercharge your productivity! Fairing, a python library to keep you in python code, hits alpha. Katib – the necessary hyperparameter tool for production ready models – adds support for TFjob. Jupyter notebook CRD – you can create a notebook from the command line. Kubeflow Pipelines! With pipelines, you can codify your workflows, removing boilerplate drudgery, adding hours back to your life every week.
• Kubeflow 0.3 Simplifies Setup & Improves ML Development – The 2018 Q3 release (see the pattern with releases?) sees the beginning of expanded options for training and serving. Kubeflow PyTorch and Chainer CRDs expand the model training options. Katib, hits alpha. Kubebench, a too for benchmarking, will assist you in assessing your hardware. Tensorflow Data Validation, part of TFX, added to Kubeflow Jupyter images. Model serving improvements through TF Serving and Nvidia’s TensorRT Inference Server
• Use Case Spotlight: AI in medicine – deriving intelligence from a sea of data – This article shows the power of what one person with passion can do to make an impact. Read this inspiring article about how an AI expert took on his toughest project ever: writing code to save his son’s life.
  

Digest #2019.07.22 – Kubeflow and Conferences, 2019

22 July 2019

• Kubeflow at OSCON 2019 – Over 10 sessions! Covering security, pipelines, productivity, ML ops and more. Some of the sessions are led by end-users, which means you’ll get the real deal about using Kubeflow in your production solution
• Kubeflow at KubeCon Europe 2019 in Barcelona – The top Kubeflow events from Kubecon in Barcelona, 2019. Tutorials, Pipelines, and Kubeflow 1.0 ruminations. The discussion on when Kubeflow will reach 1.0 should be of interest to those waiting for that milestone.
• Kubeflow Contributor Summit 2019 – Presentations and Slide decks, 22+ of them. Reviewing them will help you understand how the sausage is made. One of the interesting videos focuses on a panel discussion with machine learning practitioners and experts discussing the dynamics of machine learning at their workplace.
• Kubeflow events calendar – Find a past or future event. This is a great resource for reviewing content from community leaders and leveling up on the current state of Kubeflow. If you are aware of something that is missing, feel free to add the content through github – become a community member! 
• Use Case Spotlight: IBM’s photo-scraping scandal shows what a weird bubble AI researchers live in. This bubble is all about data – who owns it, who can monopolize it, who is monetizing it, and what the expectations around it. The expectations is the crux of the issue – people using the data may be at odds with the people supplying the data.