Machine Learning for developers: 5 Essential Podcasts

Probabilistic Programming has been discussed as a programming paradigm that uses statistical approaches to dealing with uncertainty in data as a first class construct. On today’s podcast, Wes talks with Mike Lee Williams of Cloudera’s Fast Forward Labs about Probabilistic Programming. The two discusses how Bayesian Inference works, how it’s used in Probabilistic Programming, production-level languages in the space, and some of the implementations/libraries that we’re seeing.

Key Takeaways
* Federated machine learning is an approach of developing models at an edge device and returning just the model to a centralized location. By taking the averages of the edge models, you can protect privacy and distribute processing of building models.
*Probabilistic Programming is a family of programming languages that make statistical problems easier to describe and solve.
*It is heavily influenced by Bayesian Inference or an approach to experimentation that turns what you know before the experiment and the results of the experiment into concrete answers on what you should do next.
* The Bayesian approach to unsupervised learning comes with the ability to measure uncertainty (or the ability to quantify risk).
* Most of the tooling used for Probabilistic Programming today is highly declarative. “You simply describe the world and press go.”
* If you have a practical, real-world problem today for Probabilistic Programming, Stan and PyMC3 are two languages to consider. Both are relatively mature languages with great documentation.
* Prophet, a time-series forecasting library built at Facebook as a wrapper around Stan, is a particularly approachable place to use Bayesian Inference for forecasting use cases general purpose.

Show Notes

Mike Lee Williams on Probabilistic Programming, Bayesian Inference, and Languages like PyMC3

On this week’s podcast, Wes Reisz talks to Emmanuel Ameisen, head of AI for Insight Data Science, about building a semantic search system for images using convolution neural networks and word embeddings, how you can build on the work done by companies like Google, and then explores where the gaps are and where you need to train your own models.  The podcast wraps up with a discussion around how you get something like this into production.

Why listen to this podcast:

 - A common use case is the ability to search for similar things - I want to find another pair of sunglasses like these, or I want a cat that looks like this picture, or even a tool like Google’s Smart Reply, can all be considered broadly the domain of semantic search.

 - For image classification you generally want a convolutional neural network. You typically use a model pre-trained with a public data set like Imagenet pre-trained to generate embeddings, using the pre-trained model up to the penultimate layer, and storing the value of the activations.

 - From here the idea is to mix image embeddings with word embeddings.  The embeddings, whether for words or images, are just a vector that represents a thing.  There are many approaches to getting vectors for words, but the one that started it is word2vec.

 - For both image embeddings and word embeddings you can typically use pre-trained models, meaning that you only need to train the final step of bringing the two models together.

 - Before deploying to production it is important that you validate the model against biases such as sexism, typically using outside people to a carry out a through audit.

More on this: Quick scan our curated show notes on InfoQ https://bit.ly/2RAEUrV

You can also subscribe to the InfoQ newsletter to receive weekly updates on the hottest topics from professional software development. bit.ly/24x3IVq

Subscribe: www.youtube.com/infoq
Like InfoQ on Facebook: bit.ly/2jmlyG8
Follow on Twitter: twitter.com/InfoQ
Follow on LinkedIn: www.linkedin.com/company/infoq
Check the landing page on InfoQ: https://bit.ly/2RAEUrV

Emmanuel Ameisen, Head of AI at Insight, on Building a Semantic Search System for Images

Today on the InfoQ Podcast, Wes speaks with ThirdLove’s Megan Cartwright. Megan is the Director of Data Science for the personalized bra company. In the podcast, Megan first discusses why their customers need a more personal experience and how their using technology to help. She focuses quite a bit of time in the podcast discussing how the team got to an early MVP and then how they did the same for getting to an early machine learning MVP for product recommendations. In this later part, she discusses decisions they made on what data to use, how to get the solution into production quickly, how to update/train new models, and where they needed help. It’s a real early stage startup story of a lean team leveraging machine learning to get to a practical recommendations solution in a very short timeframe.

Why listen to this podcast:

 - The experience for women selecting bras is poor experience characterized by awkward fitting experiences and an often uncomfortable product that may not even fit correctly. ThirdLove is a company built to serve this market.
 - ThirdLove took a lean approach to develop their architecture. It’s built with the Parse backend. The leveraged Shopify to build the site. The company’s first recommender system used a rules engine embedded into the front end. After that, they moved to a machine learning MVP with a Python recommender service that used a Random Forest algorithm in SciKit-Learn. 
 - Despite having the data for 10 million surveys, the first algorithms only need about 100K records to be trained. The takeaway is you don’t have to have huge amounts of data to get started with machine learning.
 - To initially deploy their ML solution, ThirdLove first shadowed all traffic through the algorithm and then compared it to what was being output by the rules engine. Using this along with information on the full customer order lifecycle, they validated the ML solution worked correctly and outperformed the rules engine.
 - ThirdLove’s machine learning story shows that you move towards a machine learning solution quickly by leveraging your own network and using tools that may already familiar to your team.

More on this: Quick scan our curated show notes on InfoQ https://bit.ly/2G9RnQn
You can also subscribe to the InfoQ newsletter to receive weekly updates on the hottest topics from professional software development. bit.ly/24x3IVq

Subscribe: www.youtube.com/infoq
Like InfoQ on Facebook: bit.ly/2jmlyG8
Follow on Twitter: twitter.com/InfoQ
Follow on LinkedIn: www.linkedin.com/company/infoq
Check the landing page on InfoQ: https://bit.ly/2G9RnQn

Landing Page

Megan Cartwright on Building a Machine Learning MVP at an Early Stage Startup

Today on The InfoQ Podcast, Wes Reisz speaks with Grady Booch. Grady is well known as the co-creator of UML, an original member of the design patterns movement, and now work he’s doing around Artificial Intelligence. On the podcast today, the two discuss what today’s reality is for AI. Grady answers questions like what does an AI mean to the practice of writing software and around how he seems it impact delivering software. In addition, Grady talks about AI surges (and winters) of over the years, the importance of ethics in software, and host of other related questions.

Why listen to this podcast:

 - There have been prior ages of AI that has lead to immediate winters of where reality set in. It stands to reason, there will be a version of an AI winter that follows today’s excitement around deep learning.

 - AIs are beginning to look at the code for testing edge cases in software and do things such as looking over your shoulder and identifying patterns in the code that you write.

 - AIs will remove tedium for software developers; however, software developer is (and will remain) a labor-intensive activity for decades to come.nAI is another bag of tools in a larger systems activity.

 - Much of the AI developers are young white men from the United States. That has a number of inherent biases in this fact. There are several organizations that are focused on combating some of these biases and bringing ethical learning into the field. This is important for us to be aware of and encourage.

 - The traditional techniques of systems engineering we know for building non-AI systems will still apply. AI’s are pieces of larger systems. That might be really interesting parts, but it’s just a part of a larger system that requires a lot of non-AI engineering use cases.

 - Early machine learning systems were mostly learn and forget systems. You teach them, you deploy them, and you walk away. Today, we do continuous learning and we need to integrate these new models into the delivery pipeline.

More on this: Quick scan our curated show notes on InfoQ https://bit.ly/2SjJOsq
You can also subscribe to the InfoQ newsletter to receive weekly updates on the hottest topics from professional software development. bit.ly/24x3IVq

Subscribe: www.youtube.com/infoq
Like InfoQ on Facebook: bit.ly/2jmlyG8
Follow on Twitter: twitter.com/InfoQ
Follow on LinkedIn: www.linkedin.com/company/infoq
Check the landing page on InfoQ: https://bit.ly/2SjJOsq

Grady Booch on Today’s Artificial Intelligence Reality and What it Means for Developers

Today on The InfoQ Podcast, Wes talks with Katharine Jarmul about privacy and fairness in machine learning algorithms. Katharine discusses what’s meant by Ethical Machine Learning and some things to consider when working towards achieving fairness. Katharine is the Co-Founder at KIProtect a machine learning security and privacy firm based in Germany and is one of the three keynotes at QCon.ai.

Why listen to this podcast:

- Ethical machine learning is about practices and strategies for creating more ethical machine learning models. There are many highly publicized/documented examples of machine learning gone awry that show the importance of the need to address ethical machine learning.
- Some of the first steps to prevent bias in machine learning is awareness. You should take time to identify your team goals and establish fairness criteria that should be revisited over time. This fairness criteria then can be used to establish the minimum fairness criteria allowed in production.
- Laws like GDPR in the EU and HIPAA in the US provide privacy and security to users and have legal implications if not followed.
- Adversarial examples (like the DolphinAttack that used subsonic sounds to activate voice assistants) can be used to fool a machine learning model into hearing or seeing something that’s not there. More and more machine learning models are becoming an attack vector for bad actors.
- Machine learning is always an iterative process.
- Zero-Knowledge Computing (or Federated Learning) is an example of machine learning at the edge and is designed to respect the privacy of an individual’s information.

More on this: Quick scan our curated show notes on InfoQ https://bit.ly/2TD3nSd
You can also subscribe to the InfoQ newsletter to receive weekly updates on the hottest topics from professional software development. bit.ly/24x3IVq

Subscribe: www.youtube.com/infoq
Like InfoQ on Facebook: bit.ly/2jmlyG8
Follow on Twitter: twitter.com/InfoQ
Follow on LinkedIn: www.linkedin.com/company/infoq
Check the landing page on InfoQ: https://bit.ly/2TD3nSd

Katharine Jarmul and Ethical Machine Learning

Software engineers, architects and team leads have found inspiration to drive change and innovation in their team by listening to the weekly InfoQ Podcast. They have received essential information that helped them validate their software development map. We have achieved that by interviewing some of the top CTOs, engineers and technology directors from companies like Uber, Netflix and more. Over 1,200,000 downloads in the last 3 years.