The Deep Dojo Machine Learning Blog

A Mac lover's guide to the robot apocalypse.

Exporting from PyTorch to Core ML

Oleg Poyaganov has implemented a bridge from PyTorch to Core ML using coremltools. The Python package listing can be found here.

Only the nn module from Torch is currently supported but this still converts a range of different layers.

Continue reading →

Godfather of AI

Joe Shute interviews Geoffrey Hinton about the bumpy road at the beginning of his career.

“In a sensibly organised society, if you improve productivity there is room for everybody to benefit. The problem is not the technology, but the way the benefits are shared out.”

Compiling Core ML Models on Device

Camilla Dahlstrøm dives under the hood and reveals some dynamic details of Core ML.

“This post will have a look at the case where the developer wishes to update or insert a model into the application, but avoiding the process of recompiling the entire application. … Several approaches will be discussed, from Apple’s preferred method to less conventional tricks which trade speed and storage efficiency alike.”

Matthijs Hollemans dives into even greater detail.

On the surface, Core ML has the appearance of being fairly static. You bring a trained model file into your project and let Xcode turn it into a runtime solution when your app is built. The details, it turns out, are a bit more dynamic.

Some of these details have even made their way up to public API.

let compiledModelURL = try MLModel.compileModel(at: modelURL)
let model = FlowerClassifier(contentsOf: compiledModelURL)
Aug 16, 2017

Under the Hood of Torch

Back in 2015 interest in Torch was peaking. Around that time, I took a look under the hood to try and map out how the framework was put together.

Continue reading →

The Data Show Podcast Discusses PyTorch

Ben Lorica interviews Soumith Chintala from Facebook who provides some background on frameworks used for training deep learning models.

“Around pre-2014, there were three main frameworks. … They all had their nitch.

Theano was really good as a symbolic compiler. Torch was a framework that would try to be out of your way if you’re a C programmer. You could write your C programs and then just interface it into Lua’s interpreter language. Caffe was very suited to computer vision models. So if you wanted a conv net and you wanted to train it on a large vision dataset, Caffe was your framework.

All three of these frameworks had aging designs. These frameworks were about six or seven years old. It was evident that the field was moving their research in a certain direction and these frameworks and their abstractions weren’t keeping up.

In late 2015, TensorFlow came out. Tensorflow was one of the first professionally built frameworks from the ground up to be open source. … I see Tensorflow as a much better Theano-style framework.”

… [Before that] Deep Mind was using Torch. Facebook. Twitter. Several university labs. The year of 2015 was Torch. The year of 2014 was Caffe. The year of 2016 was TensorFlow in terms of getting the large set of audiences.”

… Keras is a fantastic front end for TensorFlow and Theano and CNTK. You can build neural networks quickly. … It’s a very powerful tool for data scientists who want to remain in Python and never want to go into C or C++.”

Soumith was a significant contributor to Torch and started working on its successor in July 2016.

PyTorch is both a front end and a back end. You can think of PyTorch as something that gives you the ease of use of Keras, or probably more in terms of debugging. And power users can go all the way down to the C level and do hand coded optimizations.

It takes the whole stack of a front end calling a back end to create a neural network. And that back end in turn calls some underlying GPU code or CPU code. And we make that whole stack very flat without many abstractions so that you have a superior user experience.”

How to Break Into Machine Learning

Pete Warden offers advice for turning machine learning into a career.

“I took a very random path to focusing on deep learning full time, but so did most of the people I work with.”

The Story Behind ImageNet

Dave Gershgorn covers the story of Fei-Fei Li and her motivation for creating the dataset.

“In 2009, Li and her team published the ImageNet paper with the dataset — to little fanfare. Li recalls that CVPR, a leading conference in computer vision research, only allowed a poster, instead of an oral presentation, and the team handed out ImageNet-branded pens to drum up interest. People were skeptical of the basic idea that more data would help them develop better algorithms.”

— Dave Gershgorn

Within three years, everything would change.

“If the artificial intelligence boom we see today could be attributed to a single event, it would be the announcement of the 2012 ImageNet challenge results.

Geoffrey Hinton, Ilya Sutskever, and Alex Krizhevsky from the University of Toronto submitted a deep convolutional neural network architecture called AlexNet — still used in research to this day — which beat the field by a whopping 10.8 percentage point margin.”

Machine Learning Resources for iOS

A curated list of over 250 links across 20+ categories from Alex Sosnovshchenko.

“Some of the resources are awesome, some are great, some are fun, and some can serve as an inspiration.”

The most recent update was this month. If you have a GitHub account you can watch the list here.

Jul 12, 2017

Python Notes

Just enough syntax to get you started in Python. Many details will look familiar to Swift programmers.

Continue reading →

A Firsthand History of Deep Learning

Imagine you had Geoffrey Hinton, Yoshua Bengio and Yann LeCun in the same room talking about deep learning.

“In the 90’s other machine learning methods, that were easier for a novice to apply, did as well or better than neural nets on many problems. Interest in them died.

The three of us all knew they were ultimately going to be the answer. When we got better hardware and more data and a slight improvement in the techniques, they suddenly took off again.”

— Geoffrey Hinton

The interview starts 11 minutes in but the rest of the episode (and the Talking Machines podcast in general) has great content and production value.

“We had small data sets in computer vision that only have a few thousand training samples. If you train a convolutional net of the type that we had in the late 80’s and early 90’s, the performance would be very much lower than what you would get with classical vision systems. Mostly because those networks with many parameters are very hard to train. They learn the training set perfectly but they over-fit on the test set.

We devised a bunch of architectural components like rectification, contrast normalization and unsupervised pre-training that seemed to improve the performance significantly, which allowed those very heavy learning-based systems to match the performance or at least come close to the performance of classical systems. But it turns out all of this is rendered moot if you have lots of data and you use very large networks running on very fast computers.”

— Yann LeCun

“In the late 90’s and early 2000’s it was very, very difficult to do research in neural nets. In my own lab, I had to twist my students’ arms to do work on neural nets. They were afraid of seeing their papers rejected because they were working on the subject. Actually it did happen quite a bit for all the wrong reasons like, ‘Oh. We don’t do neural nets anymore.’

… I tried to even show mathematically why [the alternatives] wouldn’t work for the kinds of ambitious problems we wanted to solve for AI. That was how I started contributing towards the new wave of deep learning that CIFAR has promoted.”

— Yoshua Bengio

Correction: The original version of this post misspelled Yoshua Bengio’s name.