MLBP 9: ONNX Shakes up the Deep Learning Landscape and Numpy Drops Support for Python 2.7

• Machine Learning Blueprint Facebook Group
• Subscribe Here
• Forward to Friends

Spotlight Machine Learning Articles

ONNX: Open Neural Network Exchange Format AWS, Facebook, Microsoft and several other major players in the deep learning space announce ONNX an open source neural network format for Caffe2, MSF Cognitive, MXNet and Pytorch.

Machine Learning Blueprint's Take

Google’s TensorFlow ecosystem has dominated the deep learning world for the past several years. While different frameworks now integrated under the ONNX umbrella like Pytorch have offered advantages over TF, as discussed in the last MLBP issue, Google is quickly bridging any gaps through the release of new dynamic eager execution modes and investment in Keras as a front end. ONNX standardization promises to enable quick development combining the advantages (and libraries) across the diverse frameworks of most major TF competitors. Coupled with promised hardware optimizations, we might be seeing a reification of a true rival to Google dominance.

Numpy Drops Support for Python 2.7 Until December 31, 2018, all NumPy releases will fully support both Python2 and Python3. Starting on January 1, 2019, any new feature releases will support only Python3. On January 1, 2020 community support for the last Python2 supporting release will end.

High-Fidelity Speech Synthesis with WaveNet | DeepMind DeepMind releases details of their “probability density distillation” (PDD) technique, which enables high speed real time execution of the WaveNet algorithm. PDD uses an architecture similar to GAN networks, where a pre-trained “teacher” model functions as the discriminative algorithm and a “student” model functions as the generative algorithm. The student model learns to generate waveforms that match the teacher model but can move at high speed since the waveform distribution does not have to be generated sequentially. A related discussion on visualizing Wavenet with dimensionality reduction techniques can also be found here.

Learning Machine Learning

Understanding SSD MultiBox — Real-Time Object Detection In Deep Learning An intuitive explanation of the Single Shot Detector (SSD), an deep image classifier for drawing multiple bounding boxes that performs inference at 59 frames/second.

Uncovering the Intuition behind Capsule Networks and Inverse Graphics: Part I

Understanding the Information Bottleneck Interpretation of Deep Learning

Installing TensorFlow 1.4.0 on macOS with CUDA Support With macOS High Sierra supporting external GPU’s via thunderbolt 3, but TensorFlow dropping GPU support on macs, Hackintosh power users and new MBP owners can use this guide to backport TensorFlow to train DNN’s with whatever Nvidia setup they might have.

Expressivity In Machine Learning

How to use ML to Count Crowds and Lines

Training Machine Learning Models on the Device

Interesting Research

[1711.04837] Improving Factor-Based Quantitative Investing by Forecasting Company Fundamentals Alberg and Lipton present models suggesting that deep learning techniques can be used to forecast future estimates of publicly traded company fundamentals. Results indicate that using these techniques they are able outperform a standard factor approach with over 18% lift over compounded annual return performance.

[1711.07655] Genetic Algorithms for Evolving Deep Neural Networks

[1711.06782] Leave no Trace: Learning to Reset for Safe and Autonomous Reinforcement Learning

Machine Learning News Links

This Ai Algorithm Probably Means The End Of High-end Art Forgeries Researchers have devised a way to use machine learning to detect art fakes by looking at the characteristics of the brush strokes, which have latent features that are akin to a signature and transcend compositional features.

Machine Learning Blueprint's Take

The interesting part of the cited research paper is in the featurization of the image input using graphs to identify the strokes that will be used for training and classification. While deep learning is known for doing its own feature extraction, by performing this preprocessing they can force the algorithms to focus on the individual strokes instead of the compositional features. In the end, comparing both traditional ML algorithms with handcrafted features (worth checking out) and deep learning methods, the traditional ML approach detects fakes the best. An interesting extension of this research might be to evaluate performance of detecting fakes generated by a GAN.

Scientists Look at How Humans Drive in Self-Driving Cars MIT releases their research plan for collecting data and understanding how humans adapt to driving-assistance technology. This longitudinal study will use camera footage of drivers coupled with all the self-driving sensors to improve the technology-user relationship.

Microsoft Extends Airsim To Include Autonomous Car Research

Using Machine Learning to Unveil Congress’ Priorities. ProPublica showcases applied machine learning with topic modeling through Doc2Vec on press releases filed by Congress members. Results reveal current topic-labeling systems may incorrectly categorize documents (e.g. abortion bills may fall under “Health”, or “Law & Crime”). Furthermore, the current topics may not reflect important topics to a granular enough level. These press releases are assumed to be more telling of the Congress Person’s priorities.

Can A.I. Be Taught To Explain Itself?

How Machine Learning Is Helping Us To Understand The Brain

Deep Learning Techniques Opening up a New World of Artistic Expression

AI Can Help Hunt Down Missile Sites in China. University of Missouri uses GoogleNet and ResNet to tackle the challenge of identifying missile sites from satellite imagery. This problem is confounded by not only a lack of training data, but a huge class imbalance issue.

An On-device Deep Neural Network for Face Detection From Apple