More and more startups are looking to hire data scientists who can work autonomously to derive valuable insights from data. In principle, this sounds great: engineers and designers build the product, while data scientists crunch the numbers to gain insights. In practice, finding these data scientists and enabling them to be productive are very challenging tasks.

Before diving further, it's useful to note a few trends in data and product development that have emerged over the past decade: Companies such as Google, Amazon and Netflix have shown that proper storage and analysis of data can lead to tremendous competitive advantages.

It’s now feasible for startups to instrument and collect vast amounts of usage data. Mobile is ubiquitous, and apps are constantly emitting data. Big data infrastructure has matured, which means large-scale data storage and analysis are affordable.

The widely adopted lean startup philosophy has shifted product development to be much more data-driven. Startups now face the challenges of defining Key Performance Indicators (KPI), designing and implementing A/B tests, understanding growth and engagement funnel conversion, building machine learning models, etc.

Because of these trends, startups are eager to develop in-house data science capabilities. Unfortunately many of them have the wrong ideas about how to build such a team. Let me describe three popular misconceptions.

Misconception #1: It's okay to compromise the engineering bar for statistical skills.

For data scientists to work productively and independently, they must be able to navigate the entire technical stack and work effectively with existing systems to extract relevant data. The only exception is if a startup has already built out its data infrastructure. But in reality, very few startups have their infrastructure in place before building a data science team. In these cases, a data science team without strong engineering skills or engineering support will have a hard time doing their job. At best, they will produce suboptimal solution that will be rewritten by another team for production.

To illustrate this, take the example of building the KPI dashboard at Codecademy. Before visualizing the data in d3, I had to extract and join (a) user data from MongoDB, (b) cohort data from Redis, and (c) pageview data from Google Analytics. The data collection alone would've been near impossible without an engineering background, let alone making the dashboard real-time, modularized and reusable.

Misconception #2: It's okay to compromise the statistics bar for engineering skills.

Proper interpretation of data is not easy, and misinterpreted data can do more damage than data that's not interpreted at all (check out "Statistics done wrong"). Building useful machine learning (ML) models is trickier than most people expect. A popular but misguided view holds that ML problems can be solved either by applying some black box algorithm (a.k.a magic), or by hiring interns who are PhD students. In practice, hundreds of decisions and tradeoffs are made in solving such problems, and knowing which decisions to make requires a lot of experience. (I’ll expand upon this more in a future post titled “Machine learning done wrong”.) For a given ML problem, there are tens if not hundreds of way to solve it. Each solution makes different assumptions and it's not obvious how to navigate and identify which assumptions are reasonable and which model should be used. Some would argue: why not just try all different approaches, cross validate, and see which one works the best? In reality, you never have the bandwidth to do so. A strong data scientist might be able to produce a sensible model right off the bat, while a weak one might spend months optimizing the wrong model without knowing where the problem is.

Misconception #3: It's okay to hire data scientists who lack product thinking.

Imagine asking someone who doesn't have a holistic view of the product to optimize your business KPIs. They may prematurely optimize the sign up funnel before making sure the product has reasonable retention, which would lead to more unretained users. Some think data-driven product development is a local optimization. This criticism is only correct when those who drive product development with data fail to think about the product holistically.

To sum up, a productive data science team requires data scientists that are strong in engineering, statistics, and product thinking. It's hard. And it becomes even harder to look for the first data science hire who will be spearheading data efforts in a startup. For startups that don't have the luxury to wait and hire these rare data scientists, it's important to be aware of the compromises made especially in terms of the hiring bar. Before the data team is strong enough across all three areas, make sure they have strong support for the skills they lack, and don't expect them to work autonomously.

If you like the post, you can follow me (@chengtao_chu) on Twitter or subscribe to "ML in the Valley". Also, special thanks Ian Wong (@ihat), Leo Polovets, and Bob Ren (@bobrenjc93) for reading a draft of this.

More and more startups are looking to hire data scientists who can work autonomously to derive valuable insights from data. In principle, this sounds great: engineers and designers build the product, while data scientists crunch the numbers to gain insights. In practice, finding these data scientists and enabling them to be productive are very challenging tasks.

Before diving further, it's useful to note a few trends in data and product development that have emerged over the past decade: Companies such as Google, Amazon and Netflix have shown that proper storage and analysis of data can lead to tremendous competitive advantages.

It’s now feasible for startups to instrument and collect vast amounts of usage data. Mobile is ubiquitous, and apps are constantly emitting data. Big data infrastructure has matured, which means large-scale data storage and analysis are affordable.

The widely adopted lean startup philosophy has shifted product development to be much more data-driven. Startups now face the challenges of defining Key Performance Indicators (KPI), designing and implementing A/B tests, understanding growth and engagement funnel conversion, building machine learning models, etc.

Because of these trends, startups are eager to develop in-house data science capabilities. Unfortunately many of them have the wrong ideas about how to build such a team. Let me describe three popular misconceptions.

Misconception #1: It's okay to compromise the engineering bar for statistical skills.

For data scientists to work productively and independently, they must be able to navigate the entire technical stack and work effectively with existing systems to extract relevant data. The only exception is if a startup has already built out its data infrastructure. But in reality, very few startups have their infrastructure in place before building a data science team. In these cases, a data science team without strong engineering skills or engineering support will have a hard time doing their job. At best, they will produce suboptimal solution that will be rewritten by another team for production.

To illustrate this, take the example of building the KPI dashboard at Codecademy. Before visualizing the data in d3, I had to extract and join (a) user data from MongoDB, (b) cohort data from Redis, and (c) pageview data from Google Analytics. The data collection alone would've been near impossible without an engineering background, let alone making the dashboard real-time, modularized and reusable.

Misconception #2: It's okay to compromise the statistics bar for engineering skills.

Proper interpretation of data is not easy, and misinterpreted data can do more damage than data that's not interpreted at all (check out "Statistics done wrong"). Building useful machine learning (ML) models is trickier than most people expect. A popular but misguided view holds that ML problems can be solved either by applying some black box algorithm (a.k.a magic), or by hiring interns who are PhD students. In practice, hundreds of decisions and tradeoffs are made in solving such problems, and knowing which decisions to make requires a lot of experience. (I’ll expand upon this more in a future post titled “Machine learning done wrong”.) For a given ML problem, there are tens if not hundreds of way to solve it. Each solution makes different assumptions and it's not obvious how to navigate and identify which assumptions are reasonable and which model should be used. Some would argue: why not just try all different approaches, cross validate, and see which one works the best? In reality, you never have the bandwidth to do so. A strong data scientist might be able to produce a sensible model right off the bat, while a weak one might spend months optimizing the wrong model without knowing where the problem is.

Misconception #3: It's okay to hire data scientists who lack product thinking.

Imagine asking someone who doesn't have a holistic view of the product to optimize your business KPIs. They may prematurely optimize the sign up funnel before making sure the product has reasonable retention, which would lead to more unretained users. Some think data-driven product development is a local optimization. This criticism is only correct when those who drive product development with data fail to think about the product holistically.

To sum up, a productive data science team requires data scientists that are strong in engineering, statistics, and product thinking. It's hard. And it becomes even harder to look for the first data science hire who will be spearheading data efforts in a startup. For startups that don't have the luxury to wait and hire these rare data scientists, it's important to be aware of the compromises made especially in terms of the hiring bar. Before the data team is strong enough across all three areas, make sure they have strong support for the skills they lack, and don't expect them to work autonomously.

If you like the post, you can follow me (@chengtao_chu) on Twitter or subscribe to "ML in the Valley". Also, special thanks Ian Wong (@ihat), Leo Polovets, and Bob Ren (@bobrenjc93) for reading a draft of this.

Dissertation Defense

Feature Extraction and Fusion for Supervised and Semi-supervised
Classification: Application to fMRI and LTM Data

Wei Du

2:00pm Thursday, 24 April 2014, ITE 325B

Extracting powerful features from high dimensional noisy data promises to significantly improve the effectiveness of further analysis, especially of classification. Since there is no single feature selection and extraction method or classifier that works best on all given problems, developing effective and efficient feature selection and extraction methods and classifiers for specific applications has became one of the most active areas in the machine learning field. The aim of this dissertation is to develop novel data-driven methods for extracting and selecting the most distinguishing features for performing classification using functional magnetic resonance imaging (fMRI) and laser tread mapping (LTM) tire data.

FMRI data have the potential to characterize and classify various brain disorders including schizophrenia. However, the high dimensionality and unknown nature of fMRI data present numerous challenges to accurate analysis and interpretation. Independent component analysis (ICA), as a data-driven method, has proven very useful for fMRI analysis in extracting spatial components as multivariate features used in classification, and more recently, for the analysis of fMRI data in its native complex-valued form. In this dissertation, we first present a novel framework to extract powerful features from components estimated by ICA, allowing us to remove the redundancy and retain the most discriminative activation patterns from multivariate ICA features. We apply the proposed three-phase feature extraction framework to two real-valued fMRI data sets, and achieve high classification rates in discriminating healthy controls from patients with schizophrenia. Second, due to the iterative nature of ICA algorithms, typically independent components (ICs) are not estimated consistently during different ICA runs, and hence it is not clear which result to use further. We present a statistical framework that utilizes an objective criterion to select the best of multiple ICA runs such that the multivariate ICA features from the best run can be used for further analysis and inference. Using the proposed framework, we study the performance of a novel complex ICA algorithm for fMRI analysis, entropy rate bound minimization, which takes all three types of diversity into account, including non-Gaussianity, sample dependence and noncircularity that are present in the complex-valued fMRI data. We show that CERBM leads to significant improvement in ICs that provide higher classification accuracy, and thus is a promising ICA algorithm for the analysis of complex-valued fMRI data.

Classification using LTM data is another problem we address where we first study the use of highly multivariate solutions such as ICA and then note the advantages using lower-level features for classification. In this case, an important problem is the selection of best set of features for the best classification performance. Additionally, there are a large amount of unlabeled tire data that are easy to collect but only a few of them can be easily labeled by expert. In this dissertation, we propose a novel mutual information (MI) based approach to achieve feature splits for co-training, a practical and powerful data-driven method in semi-supervised learning. Inspired by the idea of dependent component analysis, the proposed MI-based approach presents feature splits that are maximally independent between- or within- subsets, and thus selects and fuses features more effectively than other feature split methods. Experimental results on both simulated study and LTM tire data indicate that co-training with MI-based feature splits yields significantly higher accuracy than supervised classification.

Committee: Profs. Tulay Adali (Chair), Joel Morris, Janet Rutledge, Charles E. Laberge, Vince D. Calhoun (University of New Mexico and the Mind Research Network), and Dr. Matthew Anderson (Northrop Grumman Corp.)

Celtra Inc., today announced the new AdCreator with the industry first HTML5 development technology designed for mobile display advertising across multiple screens.

Intel’s updated cross-platform UI library is built for mobile HTML5 apps and the new App Framework JavaScript library 2.1 is loaded with new features.

The front end GNU Compiler for languages such as C, C++, Objective-C, and Java, has now been upgraded with improvements for devirtualization and bottlenecks fixes.

Marie desJardins, computer science and electrical engineering, has been selected as a participant in the American Council on Education’s (ACE) Fellows Program. desJardins was one of just 31 faculty and administrators chosen from across the United States this year.

The ACE Fellows Program is the premier program for “identifying and preparing the next generation of senior leadership for the nation’s colleges and universities.” More than 300 past ACE fellows have served as chief executive officers of colleges or universities and over 1,300 have served as provosts, vice presidents and deans.

During the year-long program, desJardins will work with the president and senior officials at a host institution, while also completing a project of pressing interest to UMBC. Click here to read more about the ACE Fellows Program and here to see the full list of fellows.

This article walks web developers through the basics of working with the channel messaging APIs in HTML5.

As President Obama begins a week-long visit to Asia, Asian Studies Program Director Constantine Vaporis writes in an Al Jazeera America op-ed that the trip is aimed at reassuring allies in the region that they remain top priorities on his agenda. Vaporis writes adding South Korea and Japan to the itinerary has made the trip even more important as tension between the two countries lingers over the history wars.

“The U.S. administration has made clear that it will not act as mediator between the two sides. Nevertheless, Obama has acted assertively to try to prevent a further erosion of relations between them,” Vaporis writes.

As Obama visits Japan, South Korea, the Philippines and Malaysia, Vaporis offers suggestions for the president for how he should approach all four visits, but writes it will be important to keep one overarching theme in mind.

“Obama’s four-nation trip sends an important message about the U.S.’s commitment to Asia. Only through cooperation and mutual respect will these countries meet the regional challenges they face from China and North Korea,” he adds.

To read the full column titled “Obama seeks to rebalance ‘pivot to Asia,'” click here.

In his latest column published in Fox News Latino, Public Policy Ph.D. student Justin Vélez-Hagan writes historians may view this as Puerto Rico’s “lost decade” if no changes are made to its economic policies. Vélez-Hagan is executive director of the National Puerto Rican Chamber of Commerce. He writes Puerto Rico’s economic woes may be rooted in production.

“Many believe that Puerto Rico’s weak economy is partly attributable to a lack of investment in the production of exports, often considered one of the major factors behind the original ‘lost decade,'” Vélez-Hagan writes.

In the article, Vélez-Hagan also contends that even with its economic struggles, it is unlikely Puerto Rico will go through a substantial economic collapse.

“Although it won’t be bailed out by the IMF or World Bank, its economy still has something the rest of the world envies: the financial backstop of the greatest economic powerhouse that ever existed, one that is unlikely to allow a U.S. territory with nearly four million of its citizens to suffer through a Latin American-style economic collapse,” he adds.

To read the full column in Fox News Latino, click here.