NYU Data Science newsletter – September 25, 2015

GROUP CURATION: N/A

For the third consecutive year, O’Reilly Media conducted an anonymous survey to expose the tools that successful data scientists and engineers use, and how those tool choices might relate to their salary.

In this blog post, I want to discuss how we at Element-Research implemented the recurrent attention model (RAM) described in [1]. Not only were we able to reproduce the paper, but we also made of bunch of modular code available in the process. You can reproduce the RAM on the MNIST dataset using this training script. We will use snippets of that script throughout this post. You can then evaluate your trained models using the evaluation script.

light, it’s sometimes easy to forget about the humble book! … So here’s a rundown of 15 books which I think every data scientist should have on their shelf. Some are technical and will only be of interest to programmers or analysts, others will be interesting to anyone interested in the wider implications of our Big Data society.

Data analysis in astronomy has traditionally fallen into two categories: work done in high-level dynamic programming languages, such as IDL or Python, and work done in low-level compiled languages, such as C, C++, and Fortran. Astronomers love high-level dynamic languages for their shallow learning curve and rapid development cycle, but on very large scale or for computationally intensive problems, they often become inadequate. Compiled low-level languages give the best performance but require a higher level of programming skill and have a slower development cycle with fewer easily available tools.

Enter Julia. Julia is a new programming language that promises the dynamic programming experience of a language like Python with the performance of a low-level compiled language like C. Programming in Julia, one can scale up a problem without ever switching languages or leaving familiar tools and libraries behind. As astronomers ask more complex questions with increasing amounts of data, this capability will become critical.

The Consortium for Systemic Risk Analytics (CSRA), a not-for-profit Delaware Corporation, and the MIT Sloan School of Management announced today that it will become part of a joint collaboration among three MIT research centers—the Laboratory for Financial Engineering (LFE), the Center for Finance and Policy (CFP), and the newly launched Institute for Data, Systems, and Society (IDSS)—to continue its mission to foster interdisciplinary research in the measurement of systemic risk in the financial system.

Back in 2012, we announced that Google voice search had taken a new turn by adopting Deep Neural Networks (DNNs) as the core technology used to model the sounds of a language. These replaced the 30-year old standard in the industry: the Gaussian Mixture Model (GMM). DNNs were better able to assess which sound a user is producing at every instant in time, and with this they delivered greatly increased speech recognition accuracy.

Today, we’re happy to announce we built even better neural network acoustic models using Connectionist Temporal Classification (CTC) and sequence discriminative training techniques. These models are a special extension of recurrent neural networks (RNNs) that are more accurate, especially in noisy environments, and they are blazingly fast!

Are you ready to join the inspection of the data refineries and interview their operators? Fri Oct 2, 2015 is the date, and our bus will leave Hearst Mining Circle at 9am and be back around 9pm (after dinner at Google).

Homework/Application due: Sunday, September 27 at 5pm