Tura.io

We'll get together to discuss questions you may have about deep learning. So don't be shy. Let us know what you are working on, what you've learned and what you'd like help with.

I also propose we read and discuss "A Survey of the Usages of Deep Learning in Natural Language Processing" https://arxiv.org/abs/1807.10854 Note: There is no single person presenting this paper we'll all read it and discuss what we learned and what we didn't understand.

Probabilistic Programming will be a powerful addition to anyone's AI/ML tool belt and can be particularly useful in low data situations (and you never have enough data), to take advantage of hard learned domain knowledge and to reason under uncertainty. Kind of the opposite of deep learning. But I see them as complementing rather than competing technologies.

In this first meeting we will:

• get to know each other
• discuss our projects, interests and experiences regarding probabilistic programming
• get started learning

We'll start the first few meetings by going through the free book Bayesian Methods for Hackers by Cameron Davidson-Pilon (link below). Go ahead and read the first few chapters if get a chance before the meeting.

There is a slack channel too.

I hope you'll join us.

This will be a hands on Python and computer vision machine learning so bring a laptop. We will be using the Fruits360 dataset from Kaggle to demonstrate how to prepare images for a variety of computer vision methods. The Gap framework uses a OOP style interface which should be familiar to application developers. Andy will cover how to use the framework to preprocess images, store/retrieve, regularization and mini-batch generation.

Brad will lead a discussion on managing uncertainty in ML. Topics may include:

• Tests of statistical significance
• Confidence intervals
• Credible intervals
• Central limit theorem
• Monte Carlo simulation
• Density estimation
• Model selection
• Uncertainty in deep learning
• Probabilistic calibration
• and others ...

Can one lift the style from one image and transfer it to another? We will discuss the first paper to use neural networks to do this:

A Neural Algorithm of Artistic Style Leon A. Gatys, Alexander S. Ecker, Matthias Bethge https://arxiv.org/abs/1508.06576

We will go through the algorithm. I will present some example artistic transfers and discuss when it seems to work well and when it doesn't. I'd also like to get ideas from people about other use cases and generalizations.

If people are interested, we can also discuss practical implementation issues that may be glossed over by the paper. One implementation is available in the official pytorch tutorials. While it is a good place to start, there are some errors to look out for:

https://pytorch.org/tutorials/advanced/neural_style_tutorial.html

There also seem to be many TensorFlow tutorials available.

Bio: Currently Bill Dirks is a full time new dad. On the side, he is making art using machine learning. Previously, he worked as a software engineer for such companies as Google, New Relic, and Amazon.

We'll focus on probabilistic models and their relation to machine learning and deep learning.

Brush up by checking out "Probabilistic Programming & Bayesian Methods for Hackers", Cam Davidson-Pilon http://nbviewer.jupyter.org/github/CamDavidsonPilon/Probabilistic-Programming-and-Bayesian-Methods-for-Hackers/blob/master/Prologue/Prologue.ipynb

This week Zach will walk us through The YOLO https://pjreddie.com/darknet/yolo/ neural net, how it works and how it can be used in your apps.

Lets get together to discuss all aspects of deep learning and any questions you may have on the Fast.AI or Andrew Ng's coursera course.

Fast.Ai's Deep Learning For Coders Part 1 Version 2 has been soft launched and the videos, notebooks and code are available. So now is a great time to watch the videos and work through the exercises and come and discuss them. http://forums.fast.ai/t/unofficial-release-of-part-1-v2/9285

Lets get together to discuss all aspects of deep learning and any questions you may have on the Fast.AI or Andrew Ng's coursera course.

Fast.Ai's Deep Learning For Coders Part 1 Version 2 has been soft launched and the videos, notebooks and code are available. So now is a great time to watch the first few and work through the exercises and come and discuss them. http://forums.fast.ai/t/unofficial-release-of-part-1-v2/9285

All levels of experience are welcome. We rarely have presentations, though we may someone lead a discussion. Bring your questions, projects and ideas, basic or advanced, and we'll try to help.

Also, if you want more structure and more content check out our own Brice's "AI Saturdays" https://www.meetup.com/AI-Saturdays-PDX/

Don't forget the #deep-learning channel in the PDX Startups slack which you can join at https://pdx-startups-slack.herokuapp.com

Lets get together to discuss all aspects of deep learning and any questions you may have on the Fast.AI or Andrew Ng's coursera course.

We'll also focus on Hinton's capsules and see what we can make of them. Here are some useful links:

• The Arxiv paper

• A great youtube overview

• A Medium Piece

Can't wait to hear your experiences with capsules.

All levels of experience are welcome. We rarely have presentations, though we may someone lead a discussion. Bring your questions, projects and ideas, basic or advanced, and we'll try to help.

Lets get together to discuss all aspects of deep learning and any questions you may have on the Fast.AI or Andrew Ng's coursera course.

Also our own Matt Boyd will lead a discussion on:

"What I've found that works and can still be improved"

Matt will do a brief overview of the general approach to setting up a deep learning Kaggle project, touch on 2 or 3 good practices to follow, and throw around some ideas of what to improve.

Mat will cover how to index the data and build train/validation sets, finding the right balance of augmentation, and strategies for keeping results for 50+ different models straight.

All levels of experience are welcome. We rarely have presentations, though we may someone lead a discussion. Bring your questions, projects and ideas, basic or advanced, and we'll try to help.

Lets get together to discuss all aspects of deep learning.

In particular, this week I'm interested in learning more about ensemble methods for deep learning. If you get a chance and are so inclined check out The Relative Performance of Ensemble Methods with Deep Convolutional Neural Networks for Image Classification

Also, bring any questions you may have on the Fast.AI or Andrew Ng's coursera course.

All levels of experience are welcome. We rarely have presentations, though we may someone leads a discussion. Bring your questions, projects and ideas, basic or advanced, and we'll try to help.

Don't forget the #deep-learning channel in the PDX Startups slack which you can join at https://pdx-startups-slack.herokuapp.com