Machine Learning Blueprint Newsletter, Edition 10, 12/8/17

 View Only

Machine Learning Blueprint Newsletter, Edition 10, 12/8/17 

Wed June 19, 2019 03:11 PM

B
This newsletter is written and curated by Mike Tamir and Mike Mansour. 

December 8, 2017

Spotlight Machine Learning Articles
Turi releases a new library democratizing specific ML algorithms and task-oriented deep learning solutions that abstracts away complexities of well-solved problems with direct rails to deployment on a number of different platforms. Turi leverages their original SFrame data format for efficient operations with medium-size data on disk while boasting interoperability with Apple’s CoreML.
Machine Learning Blueprint's Take
Turi has a proven track record with making industry level machine learning toolkits in their past lives as Dato & GraphLab, but this toolkit is now aimed at a wider range of software and application developers. The direct rails for deployment on end-Apple devices (iOS, macOS, watchOS, and tvOS apps) is powerful, since deploying models that perform inference on the device is a hurdle for machine learning engineers who don’t develop applications, and application developers who don’t deeply understand machine learning. Both groups now have an easier way to deliver machine learning to market.
The winner of NIPS 2017 best paper leverages parallels between text construction and molecule construction in organic chemistry. The authors cast task of predicting chemical reactions as a translation problem, applying sequence-to-sequence modeling techniques that have worked so well ML translation. Results demonstrate a significant margin on the top-1 accuracy.
Google Brain’s team releases DeepVariant, one of the most powerful open source genomic sequencing algorithms that goes beyond simple statistical methods with deep learning. It won an award from the FDA for best performance, and is additionally available as a service through Google Cloud.
Learning Machine Learning
A guided walkthrough with a Jupyter Notebook stepping through some of the rudimentary aspects of running TensorFlow with multiple machines or on a cluster. A disclaimer: It is more practical in nature and does not cover any actual DL or RL routines.
Stochastic Gradient Descent is a great workhorse for model optimization, but this review covers some tweaks to adaptive learning methods,like ADAM, that further speed up these newer methods for a variety of application spaces.
4e3e36bb-e262-4e08-a4a0-f2796d076821.png
Used for modeling other functions or black-boxes, Gaussian Processes are a powerful tool that also model uncertainty, which is useful when the training data may not entirely cover the prediction space. This tutorial steps through some of the concepts, math, and python code for trying them out.
Interesting Research
The team at Google Brain posits that because a number of high performance indexing schemes, like B-Trees and Bloom Filters, are probabilistic in nature under certain conditions and an efficient representation is “learned” over the data, the problem space is is a good application for deep-learned index schemes. The learned index structures would be high optimized over the data and offer potentially constant time lookups, as opposed to a O(log(n)) lookup in a B-Tree. The proposed work focuses on read-only applications, but they show how it could be extended for making write-operations also more efficient.
86c222fb-20b5-409e-8f26-9399a3df0335.png

Machine Learning News Links
Teaching Machines to Understand User Interfaces. A novel idea to use deep learning in the user interface design workflow to remove non-creative and redundant aspects when moving between sketches/concepts to wireframes, and from wireframes to functional code. The goal of the concept is not to replace anyone, but to augment and speed up their workflows, enabling them to focus on the more interesting parts. The author shares their working basic implementation for automating the wireframe code workflow step here.
55bc7246-4585-4d4e-afdc-cb2895a50c29.png
Addressing One of the More Controversial NIPS Presentations: Alchemy, Rigour and Engineering. A review and commentary on a controversial talk at NIPS by Ali Rahimi, who claims much of the recent advances in deep learning are alchemy, and response by Yann LeCun defending the current state of feature engineering.

#GlobalAIandDataScience
#GlobalDataScience

Statistics

0 Favorited
6 Views
0 Files
0 Shares
0 Downloads