Neurohackademy 2018: A wrap-up

It’s become something of a truism in recent years that scientists in many fields find themselves drowning in data. This is certainly the case in neuroimaging, where even small functional MRI datasets typically consist of several billion observations (e.g., 100,000 points in the brain, each measured at 1,000 distinct timepoints, in each of 20 subjects). Figuring out how to store, manage, analyze, and interpret data on this scale is a monumental challenge–and one that arguably requires a healthy marriage between traditional neuroimaging and neuroscience expertise, and computational skills more commonly found in data science, statistics, or computer science departments.

In an effort to help bridge this gap, Ariel Rokem and I have spent part of our summer each of the last three years organizing a summer institute at the intersection of neuroimaging and data science. The most recent edition of the institute–Neurohackademy 2018–just wrapped up last week, so I thought this would be a good time to write up a summary of the course: what the course is about, who attended and instructed, what everyone did, and what lessons we’ve learned.

What is Neurohackademy?

Neurohackademy started its life in Summer 2016 as the somewhat more modestly-named Neurohackweek–a one-week program for 40 participants modeled on Astrohackweek, a course organized by the eScience Institute in collaboration with data science initiatives at Berkeley and NYU. The course was (and continues to be) held on the University of Washington’s beautiful campus in Seattle, where Ariel is based (I make the trip from Austin, Texas every year–which, as you can imagine, is a terrible sacrifice on my part given the two locales’ respective summer climates). The first two editions were supported by UW’s eScience Institute (and indirectly, by grants from the Moore and Sloan foundations). Thanks to generous support from the National Institute of Mental Health (NIMH), this year the course expanded to two weeks, 60 participants, and over 20 instructors (our funding continues through 2021, so there will be at least 3 more editions).

The overarching goal of the course is to give neuroimaging researchers the scientific computing and data science skills they need in order to get the most out of their data. Over the course of two weeks, we cover a variety of introductory and (occasionally) advanced topics in data science, and demonstrate how they can be productively used in a range of neuroimaging applications. The course is loosely structured into three phases (see the full schedule here): the first few days feature domain-general data science tutorials; the next few days focus on sample neuroimaging applications; and the last few days consist of a full-blown hackathon in which participants pitch potential projects, self-organize into groups, and spend their time collaboratively working on a variety of software, analysis, and documentation projects.

Who attended?

Admission to Neurohackademy 2018 was extremely competitive: we received nearly 400 applications for just 60 spots. This was a very large increase from the previous two years, presumably reflecting the longer duration of the course and/or our increased efforts to publicize it. While we were delighted by the deluge of applications, it also meant we had to be far more selective about admissions than in previous years. The highly interactive nature of the course, coupled with the high per-participant costs (we provide two weeks of accommodations and meals), makes it unlikely that Neurohackademy will grow beyond 60 participants in future editions, despite the clear demand. Our rough sense is that somewhere between half and two-thirds of all applicants were fully qualified and could have easily been admitted, so there’s no question that, for many applicants, blind luck played a large role in determining whether or not they were accepted. I mention this mainly for the benefit of people who applied for the 2018 course and didn’t make it in: don’t take it personally! There’s always next year. (And, for that matter, there are also a number of other related summer schools we encourage people to apply to, including the Methods in Neuroscience at Dartmouth Computational Summer School, Allen Institute Summer Workshop on the Dynamic Brain, Summer School in Computational Sensory-Motor Neuroscience, and many others.)

The 60 participants who ended up joining us came from a diverse range of demographic backgrounds, academic disciplines, and skill levels. Most of our participants were trainees in academic programs (40 graduate students, 12 postdocs), but we also had 2 faculty members, 6 research staff, and 2 medical residents (note that all of these counts include 4 participants who were admitted to the course but declined to, or could not, attend). We had nearly equal numbers of male and female participants (30F, 33M), and 11 participants came from traditionally underrepresented backgrounds. 43 participants were from institutions or organizations based in the United States, with the remainder coming from 14 different countries around the world.

The disciplinary backgrounds and expertise levels of participants are a bit harder to estimate for various reasons, but our sense is that the majority (perhaps two-thirds) of participants received their primary training in non-computational fields (psychology, neuroscience, etc.). This was not necessarily by design–i.e., we didn’t deliberately favor applicants from biomedical fields over applicants from computational fields–and primarily mirrored the properties of the initial applicant pool. We did impose a hard requirement that participants should have at least some prior expertise in both programming and neuroimaging, but subject to that constraint, there was enormous variation in previous experience along both dimensions–something that we see as a desirable feature of the course (more on this below).

We intend to continue to emphasize and encourage diversity at Neurohackademy, and we hope that all of our participants experienced the 2018 edition as a truly inclusive, welcoming event.

Who taught?

We were fortunate to be able to bring together more than 20 instructors with world-class expertise in a diverse range of areas related to neuroimaging and data science. “Instructor” is a fairly loose term at Neurohackademy: we deliberately try to keep the course non-hierarchical, so that for the most part, instructors are just participants who happen to fall on the high-experience tail of the experience distribution. That said, someone does have to teach the tutorials and lectures, and we were lucky to have a stellar cast of experts on hand. Many of the data science tutorials during the first phase of the course were taught by eScience staff and UW faculty kind enough to take time out of their other duties to help teach participants a range of core computing skills: Git and GitHub (Bernease Herman), R (Valentina Staneva and Tara Madhyastha), web development (Anisha Keshavan), and machine learning (Jake Vanderplas), among others.

In addition to the local instructors, we were joined for the tutorial phase by Kirstie Whitaker (Turing Institute), Chris Gorgolewski (Stanford), Satra Ghosh (MIT), and JB Poline (McGill)–all veterans of the course from previous years (Kirstie was a participant at the first edition!). We’re particularly indebted to Kirstie and Chris for their immense help. Kirstie was instrumental in helping a number of participants bridge the (large!) gap between using git privately, and using it to actively collaborate on a public project. As one of the participants elegantly put it:

Chris shouldered a herculean teaching load, covering Docker, software testing, BIDS and BIDS-Apps, and also leading an open science panel. I’m told he even sleeps on occasion.

We were also extremely lucky to have Fernando Perez (Berkeley)–the creator of IPython and leader of the Jupyter team–join us for several days; his presentation on Jupyter (videos: part 1 and part 2) was one of the highlights of the course for me personally, and I heard many other instructors and participants share the same sentiment. Jupyter was a critical part of our course infrastructure (more on that below), so it was fantastic to have Fernando join us and share his insights on the fascinating history of Jupyter, and on reproducible science more generally.

As the course went on, we transitioned from tutorials focused on core data science skills to more traditional lectures focusing on sample applications of data science methods to neuroimaging data. Instructors during this phase of the course included Tor Wager (Colorado), Eva Dyer (Georgia Tech), Gael Varoquaux (INRIA), Tara Madhyastha (UW), Sanmi Koyejo (UIUC), and Nick Cain and Justin Kiggins (Allen Institute for Brain Science). We continued to emphasize hands-on interaction with data; many of the presenters during this phase spent much of their time showing participants how to work with programmatic tools to generate the kinds of results one might find in papers they’ve authored (e.g., Tor Wager and Gael Varoquaux demonstrated tools for neuroimaging data analysis written in Matlab and Python, respectively).

The fact that so many leading experts were willing to take large chunks of time out of their schedule (most of the instructors hung around for several days, facilitating extended interactions with participants) to visit with us at Neurohackademy speaks volumes about the kind of people who make up the neuroimaging data science community. We’re tremendously grateful to these folks for their contributions, and hope they’ll return to teach at future editions of the institute.

What did we cover?

The short answer is: see for yourself! We’ve put most of the slides, code, and videos from the course online, and encourage people to interact with, learn from, and reuse these materials.

Now the long(er) answer. One of the challenges in organizing scientific training courses that focus on technical skill development is that participants almost invariably arrive with a wide range of backgrounds and expertise levels. At Neurohackademy, some of the participants were effectively interchangeable with instructors, while others were relatively new to programming and/or neuroimaging. The large variance in technical skill is a feature of the course, not a bug: while we require all admitted participants to have some prior programming background, we’ve found that having a range of skill levels is an excellent way to make sure that everyone is surrounded by people who they can alternately learn from, help out, and collaborate with.

That said, the wide range of backgrounds does present some organizational challenges: introductory sessions often bore more advanced participants, while advanced sessions tend to frustrate newcomers. To accommodate the range of skill levels, we tried to design the course in a way that benefits as many people as possible (though we don’t pretend to think it worked great for everyone). During the first two days, we featured two tracks of tutorials at most times, with simultaneously-held presentations generally differing in topic and/or difficulty (e.g., Git/GitHub opposite Docker; introduction to Python opposite introduction to R; basic data visualization opposite computer vision).

Throughout Neurohackademy, we deliberately placed heavy emphasis on the Python programming language. We think Python has a lot going for it as a lingua franca of data science and scientific computing. The language is free, performant, relatively easy to learn, and very widely used within the data science, neuroimaging, and software development communities. It also helps that many of our instructors (e.g., Fernando Perez, Jake Vanderplas, and Gael Varoquaux) are major contributors to the scientific Python ecosystem, so there was a very high concentration of local Python expertise to draw on. That said, while most of our instruction was done in Python, we were careful to emphasize that participants were free to work in whatever language(s) they like. We deliberately include tutorials and lectures that featured R, Matlab, or JavaScript, and a number of participant projects (see below) were written partly or entirely in other languages, including R, Matlab, JavaScript, and C.

We’ve also found that the tooling we provide to participants matters–a lot. A robust, common computing platform can spell the difference between endless installation problems that eat into valuable course time, and a nearly seamless experience that participants can dive into right away. At Neurohackademy, we made extensive use of the Jupyter suite of tools for interactive computing. In particular, thanks to Ariel’s heroic efforts (which built on some very helpful docs, similarly heroic efforts by Chris Holdgraf, Yuvi Panda, and Satra Ghosh last year), we were able to conduct a huge portion of our instruction and collaborative hacking using a course-wide Jupyter Hub allocation, deployed via Kubernetes, running on the Google Cloud. This setup allowed Ariel to create a common web-accessible environment for all course participants, so that, at the push of a button, each participant was dropped into a Jupyter Lab environment containing many of the software dependencies, notebooks, and datasets we used throughout the course. While we did run into occasional scaling bottlenecks (usually when an instructor demoed a computationally intensive method, prompting dozens of people to launch the same process in their pods), for the most part, our participants were able to drop into a running JupyterLab instance within seconds and immediately start interactively playing with the code being presented by instructors.

Surprisingly (at least to us), our total Google Cloud computing costs for the entire two-week, 60-participant course came to just $425. Obviously, that number could have easily skyrocketed had we scaled up our allocation dramatically and allowed our participants to execute arbitrarily large jobs (e.g., preprocessing data from all ~1,200 HCP subjects). But we thought the limits we imposed were pretty reasonable, and our experience suggests that not only is Jupyter Hub an excellent platform from a pedagogical standpoint, but it can also be an extremely cost-effective one.

What did we produce?

Had Neurohackademy produced nothing at all besides the tutorials, slides, and videos generated by instructors, I think it’s fair to say that participants would still have come away feeling that they learned a lot (more on that below). But a major focus of the institute was on actively hacking on the brain–or at least, on data related to the brain. To this effect, the last 3.5 days of the course were dedicated exclusively to a full-blown hackathon in which participants pitched potential projects, self-organized into groups, and then spent their time collaboratively working on a variety of software, analysis, and documentation projects. You can find a list of most of the projects on the course projects repository (most link out to additional code or resources).

As one might expect given the large variation in participant experience, project group size, and time investment (some people stuck to one project for all three days, while others moved around), the scope of projects varied widely. From our perspective–and we tried to emphasize this point throughout the hackathon–the important thing was not what participants’ final product looked like, but how much they learned along the way. There’s always a tension between exploitation and exploration at hackathons, with some people choosing to spend most of their time expanding on existing projects using technologies they’re already familiar with, and others deciding to start something completely new, or to try out a new language–and then having to grapple with the attendant learning curve. While some of the projects were based on packages that predated Neurohackademy, most participants ended up working on projects they came up with de novo at the institute, often based on tools or resources they first learned about during the course. I’ll highlight just three projects here that provide a representative cross-section of the range of things people worked on:

1. Peer Herholz and Rita Ludwig created a new BIDS-app called Bidsonym for automated de-identification of neuroimaging data. The app is available from Docker Hub, and features not one, not two, but three different de-identification algorithms. If you want to shave the faces off of your MRI participants with minimal fuss, make friends with Bidsonym.

2. A group of eight participants ambitiously set out to develop a new “O-Factor” metric intended to serve as a relative measure of the openness of articles published in different neuroscience-related journals. The project involved a variety of very different tasks, including scraping (public) data from the PubMed Central API, computing new metrics of code and data sharing, and interactively visualizing the results using a d3 dashboard. While the group was quick to note that their work is preliminary, and has a bunch of current limitations, the results look pretty great–though some disappointment was (facetiously) expressed during the project presentations that the journal Nature is not, as some might have imagined, a safe house where scientific datasets can hide from the prying public.

3. Emily Wood, Rebecca Martin, and Rosa Li worked on tools to facilitate mixed-model analysis of fMRI data using R. Following a talk by Tara Madhyastha  on her Neuropointillist R framework for fMRI data analysis, the group decided to create a new series of fully reproducible Markdown-based tutorials for the package (the original documentation was based on non-public datasets). The group expanded on the existing installation instructions (discovering some problems in the process), created several tutorials and examples, and also ended up patching the neuropointillist code to work around a very heavy dependency (FSL).

You can read more about these 3 projects and 14 others on the project repository, and in some cases, you can even start using the tools right away in your own work. Or you could just click through and stare at some of the lovely images participants produced.

So, how did it go?

It went great!

Admittedly, Ariel and I aren’t exactly impartial parties–we wouldn’t keep doing this if we didn’t think participants get a lot out of it. But our assessment isn’t based just on our personal impressions; we have participants fill out a detailed (and anonymous) survey every year, and go out of our way to encourage additional constructive criticism from the participants (which a majority provide). So I don’t think we’re being hyperbolic when we say that most people who participated in the course had an extremely educational and enjoyable experience. Exhibit A is this set of unsolicited public testimonials, courtesy of twitter:

The organizers and instructors all worked hard to build an event that would bring people together as a collaborative and productive (if temporary) community, and it’s very gratifying to see those goals reflected in participants’ experiences.

Of course, that’s not to say there weren’t things we could do better; there were plenty, and we’ve already made plans to adjust and improve the course next year based on feedback we received. For example, some suggestions we received from multiple participants included adding more ice-breaking activities early on in the course; reducing the intensity of the tutorial/lecture schedule the first week (we went 9 am to 6 pm every day, stopping only for an hourlong lunch and a few short breaks); and adding designated periods for interaction with instructors and other participants. We’ve already made plans to address these (and several other) recommendations in next year’s edition, and expect it to looks slightly different from (and hopefully better than!) Neurohackademy 2018.

Thank you!

I think that’s a reasonable summary of what went on at Neurohackademy 2018. We’re delighted at how the event turned out, and are happy to answer questions (feel free to leave them in the comments below, or to email Ariel and/or me).

We’d like to end by thanking all of the people and organizations who helped make Neurohackademy 2018 a success: NIMH for providing the funding that makes Neurohackademy possible; the eScience Institute and staff for throwing their wholehearted support behind the course (particularly our awesome course coordinator, Rachael Murray); and the many instructors who each generously took several days (and in a few cases, more than a week!) out of their schedule, unpaid, to come to Seattle and share their knowledge with a bunch of enthusiastic strangers. On a personal note, I’d also like to thank Ariel, who did the lion’s share of the actual course directing. I mostly just get to show up in Seattle, teach some stuff, hang out with great people, and write a blog post about it.

Lastly, and above all else, we’d like to thank our participants. It’s a huge source of inspiration and joy to us each year to see what a group of bright, enthusiastic, motivated researchers can achieve when given time, space, and freedom (and, okay, maybe also a large dollop of cloud computing credits). We’re looking forward to at least three more years of collaborative, productive neurohacking!

yet another Python state machine (and why you might care)

TL;DR: I wrote a minimalistic state machine implementation in Python. You can find the code on GitHub. The rest of this post explains what a state machine is and why you might (or might not) care. The post is slanted towards scientists who are technically inclined but lack formal training in computer science or software development. If you just want some documentation or examples, see the README.

A common problem that arises in many software applications is the need to manage an application’s trajectory through a state of discrete states. This problem will be familiar, for instance, to almost every researcher who has ever had to program an experiment for a study involving human subjects: there are typically a number of different states your study can be in (informed consent, demographic information, stimulus presentation, response collection, etc.), and these states are governed by a set of rules that determine the valid progression of your participants from one state to another. For example, a participant can proceed from informed consent to a cognitive task, but never the reverse (on pain of entering IRB hell!).

In the best possible case, the transition rules are straightforward. For example, given states [A, B, C, D], life would be simple if the the only valid transitions were A –> B, B –> C, and C –> D. Unfortunately, the real world is more complicated, and state transitions are rarely completely sequential. More commonly, at least some states have multiple potential destinations. Sometimes the identity of the next state depends on meeting certain conditions while in the current state (e.g., if the subject responded incorrectly, the study may transition to a different state than if they had responded correctly); other times the rules may be probabilistic, or depend on the recent trajectory through state space (e.g., a slot machine transitions to a winning or losing state with some fixed probability that may also depend on its current position, recent history, etc.).

In software development, a standard method for dealing with this kind of problem is to use something called a finite-state machine (FSM). FSMs have been around a relatively long time (at least since Mealy and Moore’s work in the 1950s), and have all kinds of useful applications. In a nutshell, what a good state machine implementation does is represent much of the messy logic governing state transitions in a more abstract, formal and clean way. Rather than having to write a lot of complicated nested logic to direct the flow of the application through state space, one can usually get away with a terse description of (a) the possible states of the machine and (b) a list of possible transitions, including a specification of the source and destination states for each transition, what conditions must be met in order for the transition to execute, etc.

For example, suppose you need to write some code to transition between different phases in an online experiment. Your naive implementation might look vaguely like this (leaving out a lot of supporting code and focusing just on the core logic):

if state == 'consent' and user_response == 'Agree':
    state = 'demographics'
elif state == 'demographics' and validate_demographics(data):
    state = 'personality'
elif state == 'personality':
    if not has_more_questions():
        state = 'task'
elif state == 'task':

This is a minimalistic example, but already, it illustrates several common scenarios–e.g., that the transition from one state to another often depends on meeting some specified condition (we don’t advance beyond the informed consent stage until the user signs the document), and that there may be some actions we want to issue immediately before or after a particular kind of transition (e.g., we save survey responses before we move onto the next phase).

The above code is still quite manageable, so if things never get any more complex than this, there may be no reason to abandon a (potentially lengthy) chain of conditionals in favor of a fundamentally different approach. But trouble tends to arises when the complexity does increase–e.g., you need to throw a few more states into the mix later on–or when you need to move stuff around (e.g., you decide to administer the task before the demographic survey). If you’ve ever had the frustrating experience of tracing the flow of your app through convoluted logic scattered across several files, and being unable to figure out why your code is entering the wrong state in response to some triggered event, the state machine pattern may be right for you.

I’ve made extensive use of state machines in the past when building online studies, and finding a suitable implementation has never been a problem. For example, in Rails–which is what most of my apps have been built in–there are a number of excellent options, including the state_machine plugin and (more recently) Statesman. In the last year or two, though, I’ve begun to transition all of my web development to Python (if you want to know why, read this). Python is a very common language, and the basic FSM pattern is very simple, so there are dozens of Python FSM implementations out there. But for some reason, very few of the Python implementations are as elegant and usable as their Ruby analogs. This isn’t to say there aren’t some nice ones (I’m partial to Fysom, for instance)–just that none of them quite meet my needs (in particular, there are very few fully object-oriented implementations, and I like to have my state machine tightly coupled with the model it’s managing). So I decided to write one. It’s called Transitions, and you can find the code on GitHub, or install it directly from the command prompt (“pip install transitions”, assuming you have pip installed). It’s very lightweight–fewer than 200 lines of code (the documentation is about 10 times as long!)–but still turns out to be quite functional.

For example, here’s some code that does almost exactly the same thing as what we saw above (there are much more extensive examples and documentation in the GitHub README):

from transitions import Machine

# define our states and transitions
states = ['consent', 'demographics', 'personality', 'task']
transitions = [
        'trigger': 'advance',
        'source': 'consent',
        'dest': 'demographics',
        'conditions': 'user_agrees'
        'trigger': 'advance', 
        'source': 'demographics', 
        'dest': 'personality', 
        'conditions': 'validate_demographics', 
        'before': 'save_demographics'
        'trigger': 'advance',
        'source': 'personality',
        'dest': 'task',
        'conditions': 'no_more_items',
        'before': 'save_items'

# Initialize the state machine with the above states and transitions, and start out life in the solid state.
machine = Machine(states=states, transitions=transitions, initial='consent')

# Let's see how it works...
> 'consent'
machine.advance() # Trigger methods are magically added for us!
> 'demographics'

That’s it! And now we have a nice object-oriented state machine that elegantly transitions between phases of matter, triggers callback functions as needed, and supports conditional transitions, branching, and various other nice features, all without ever having to write a single explicit conditional or for-loop. Understanding what’s going on is as simple as looking at the specification of the states and transitions. For example, we can tell at a glance from the second transition that if the model is currently in the ‘demographics’ state, calling advance() will effect a transition to the ‘personality’ state–conditional on the validate_demographics() function returns True. Also, right before the transition executes, the save_demographics() callback will be called.

As I noted above, given the simplicity of the example, this may not seem like a huge win. If anything, the second snippet is slightly longer than the first. But it’s also much clearer (once you’re familiar with the semantics of Transitions), scales much better as complexity increases, and will be vastly easier to modify when you need to change anything.

Anyway, I mention all of this here for two reasons. First, as small and simple a project as this is, I think it ended up being one of the more elegant and functional minimalistic Python FSMs–so I imagine a few other people might find it useful (yes, I’m basically just exploiting my PageRank on Google to drive traffic to GitHub). And second, I know many people who read this blog are researchers who regularly program experiments, but probably haven’t encountered state machines before. So, Python implementation aside, the general idea that there’s a better way to manage complex state transitions than writing a lot of ugly logic seems worth spreading.

The homogenization of scientific computing, or why Python is steadily eating other languages’ lunch

Over the past two years, my scientific computing toolbox been steadily homogenizing. Around 2010 or 2011, my toolbox looked something like this:

  • Ruby for text processing and miscellaneous scripting;
  • Ruby on Rails/JavaScript for web development;
  • Python/Numpy (mostly) and MATLAB (occasionally) for numerical computing;
  • MATLAB for neuroimaging data analysis;
  • R for statistical analysis;
  • R for plotting and visualization;
  • Occasional excursions into other languages/environments for other stuff.

In 2013, my toolbox looks like this:

  • Python for text processing and miscellaneous scripting;
  • Ruby on Rails/JavaScript for web development, except for an occasional date with Django or Flask (Python frameworks);
  • Python (NumPy/SciPy) for numerical computing;
  • Python (Neurosynth, NiPy etc.) for neuroimaging data analysis;
  • Python (NumPy/SciPy/pandas/statsmodels) for statistical analysis;
  • Python (MatPlotLib) for plotting and visualization, except for web-based visualizations (JavaScript/d3.js);
  • Python (scikit-learn) for machine learning;
  • Excursions into other languages have dropped markedly.

You may notice a theme here.

The increasing homogenization (Pythonification?) of the tools I use on a regular basis primarily reflects the spectacular recent growth of the Python ecosystem. A few years ago, you couldn’t really do statistics in Python unless you wanted to spend most of your time pulling your hair out and wishing Python were more like R (which, is a pretty remarkable confession considering what R is like). Neuroimaging data could be analyzed in SPM (MATLAB-based), FSL, or a variety of other packages, but there was no viable full-featured, free, open-source Python alternative. Packages for machine learning, natural language processing, web application development, were only just starting to emerge.

These days, tools for almost every aspect of scientific computing are readily available in Python. And in a growing number of cases, they’re eating the competition’s lunch.

Take R, for example. R’s out-of-the-box performance with out-of-memory datasets has long been recognized as its achilles heel (yes, I’m aware you can get around that if you’re willing to invest the time–but not many scientists have the time). But even people who hated the way R chokes on large datasets, and its general clunkiness as a language, often couldn’t help running back to R as soon as any kind of serious data manipulation was required. You could always laboriously write code in Python or some other high-level language to pivot, aggregate, reshape, and otherwise pulverize your data, but why would you want to? The beauty of packages like plyr in R was that you could, in a matter of 2 – 3 lines of code, perform enormously powerful operations that could take hours to duplicate in other languages. The downside was the intensive learning curve associated with learning each package’s often quite complicated API (e.g., ggplot2 is incredibly expressive, but every time I stop using ggplot2 for 3 months, I have to completely re-learn it), and having to contend with R’s general awkwardness. But still, on the whole, it was clearly worth it.

Flash forward to The Now. Last week, someone asked me for some simulation code I’d written in R a couple of years ago. As I was firing up R Studio to dig around for it, I realized that I hadn’t actually fired up R studio for a very long time prior to that moment–probably not in about 6 months. The combination of NumPy/SciPy, MatPlotLib, pandas and statmodels had effectively replaced R for me, and I hadn’t even noticed. At some point I just stopped dropping out of Python and into R whenever I had to do the “real” data analysis. Instead, I just started importing pandas and statsmodels into my code. The same goes for machine learning (scikit-learn), natural language processing (nltk), document parsing (BeautifulSoup), and many other things I used to do outside Python.

It turns out that the benefits of doing all of your development and analysis in one language are quite substantial. For one thing, when you can do everything in the same language, you don’t have to suffer the constant cognitive switch costs of reminding yourself say, that Ruby uses blocks instead of comprehensions, or that you need to call len(array) instead of array.length to get the size of an array in Python; you can just keep solving the problem you’re trying to solve with as little cognitive overhead as possible. Also, you no longer need to worry about interfacing between different languages used for different parts of a project. Nothing is more annoying than parsing some text data in Python, finally getting it into the format you want internally, and then realizing you have to write it out to disk in a different format so that you can hand it off to R or MATLAB for some other set of analyses*. In isolation, this kind of thing is not a big deal. It doesn’t take very long to write out a CSV or JSON file from Python and then read it into R. But it does add up. It makes integrated development more complicated, because you end up with more code scattered around your drive in more locations (well, at least if you have my organizational skills). It means you spend a non-negligible portion of your “analysis” time writing trivial little wrappers for all that interface stuff, instead of thinking deeply about how to actually transform and manipulate your data. And it means that your beautiful analytics code is marred by all sorts of ugly open() and read() I/O calls. All of this overhead vanishes as soon as you move to a single language.

Convenience aside, another thing that’s impressive about the Python scientific computing ecosystem is that a surprising number of Python-based tools are now best-in-class (or close to it) in terms of scope and ease of use–and, in virtue of C bindings, often even in terms of performance. It’s hard to imagine an easier-to-use machine learning package than scikit-learn, even before you factor in the breadth of implemented algorithms, excellent documentation, and outstanding performance. Similarly, I haven’t missed any of the data manipulation functionality in R since I switched to pandas. Actually, I’ve discovered many new tricks in pandas I didn’t know in R (some of which I’ll describe in an upcoming post). Considering that pandas considerably outperforms R for many common operations, the reasons for me to switch back to R or other tools–even occasionally–have dwindled.

Mind you, I don’t mean to imply that Python can now do everything anyone could ever do in other languages. That’s obviously not true. For instance, there are currently no viable replacements for many of the thousands of statistical packages users have contributed to R (if there’s a good analog for lme4 in Python, I’d love to know about it). In signal processing, I gather that many people are wedded to various MATLAB toolboxes and packages that don’t have good analogs within the Python ecosystem. And for people who need serious performance and work with very, very large datasets, there’s often still no substitute for writing highly optimized code in a low-level compiled language. So, clearly, what I’m saying here won’t apply to everyone. But I suspect it applies to the majority of scientists.

Speaking only for myself, I’ve now arrived at the point where around 90 – 95% of what I do can be done comfortably in Python. So the major consideration for me, when determining what language to use for a new project, has shifted from what’s the best tool for the job that I’m willing to learn and/or tolerate using? to is there really no way to do this in Python? By and large, this mentality is a good thing, though I won’t deny that it occasionally has its downsides. For example, back when I did most of my data analysis in R, I would frequently play around with random statistics packages just to see what they did. I don’t do that much any more, because the pain of having to refresh my R knowledge and deal with that thing again usually outweighs the perceived benefits of aimless statistical exploration. Conversely, sometimes I end up using Python packages that I don’t like quite as much as comparable packages in other languages, simply for the sake of preserving language purity. For example, I prefer Rails’ ActiveRecord ORM to the much more explicit SQLAlchemy ORM for Python–but I don’t prefer to it enough to justify mixing Ruby and Python objects in the same application. So, clearly, there are costs. But they’re pretty small costs, and for me personally, the scales have now clearly tipped in favor of using Python for almost everything. I know many other researchers who’ve had the same experience, and I don’t think it’s entirely unfair to suggest that, at this point, Python has become the de facto language of scientific computing in many domains. If you’re reading this and haven’t had much prior exposure to Python, now’s a great time to come on board!

Postscript: In the period of time between starting this post and finishing it (two sessions spread about two weeks apart), I discovered not one but two new Python-based packages for data visualization: Michael Waskom’s seaborn package–which provides very high-level wrappers for complex plots, with a beautiful ggplot2-like aesthetic–and Continuum Analytics’ bokeh, which looks like a potential game-changer for web-based visualization**. At the rate the Python ecosystem is moving, there’s a non-zero chance that by the time you read this, I’ll be using some new Python package that directly transliterates my thoughts into analytics code.


* I’m aware that there are various interfaces between Python, R, etc. that allow you to internally pass objects between these languages. My experience with these has not been overwhelmingly positive, and in any case they still introduce all the overhead of writing extra lines of code and having to deal with multiple languages.

** Yes, you heard right: web-based visualization in Python. Bokeh generates static JavaScript and JSON for you from Python code, so  your users are magically able to interact with your plots on a webpage without you having to write a single line of native JS code.

the Neurosynth viewer goes modular and open source

If you’ve visited the Neurosynth website lately, you may have noticed that it looks… the same way it’s always looked. It hasn’t really changed in the last ~20 months, despite the vague promise on the front page that in the next few months, we’re going to do X, Y, Z to improve the functionality. The lack of updates is not by design; it’s because until recently I didn’t have much time to work on Neurosynth. Now that much of my time is committed to the project, things are moving ahead pretty nicely, though the changes behind the scenes aren’t reflected in any user-end improvements yet.

The github repo is now regularly updated and even gets the occasional contribution from someone other than myself; I expect that to ramp up considerably in the coming months. You can already use the code to run your own automated meta-analyses fairly easily; e.g., with everything set up right (follow the Readme and examples in the repo), the following lines of code:

dataset = cPickle.load(open('dataset.pkl', 'rb'))
studies = get_ids_by_expression("memory* &~ ("wm|working|episod*"), threshold=0.001)
ma = meta.MetaAnalysis(dataset, studies)

…will perform an automated meta-analysis of all studies in the Neurosynth database that use the term ‘memory’ at a frequency of 1 in 1,000 words or greater, but don’t use the terms wm or working, or words that start with ‘episod’ (e.g., episodic). You can perform queries that nest to arbitrary depths, so it’s a pretty powerful engine for quickly generating customized meta-analyses, subject to all of the usual caveats surrounding Neurosynth (i.e., that the underlying data are very noisy, that terms aren’t mental states, etc.).

Anyway, with the core tools coming along, I’ve started to turn back to other elements of the project, starting with the image viewer. Yesterday I pushed the first commit of a new version of the viewer that’s currently on the Neurosynth website. In the next few weeks, this new version will be replacing the current version of the viewer, along with a bunch of other changes to the website.

A live demo of the new viewer is available here. It’s not much to look at right now, but behind the scenes, it’s actually a huge improvement on the old viewer in a number of ways:

  • The code is completely refactored and is all nice and object-oriented now. It’s also in CoffeeScript, which is an alternative and (if you’re coming from a Python or Ruby background) much more readable syntax for JavaScript. The source code is on github and contributions are very much encouraged. Like most scientists, I’m generally loathe to share my code publicly because I think it sucks most of the time. But I actually feel pretty good about this code. It’s not good code by any stretch, but I think it rises to the level of ‘mostly sensible’, which is about as much as I can hope for.
  • The viewer now handles multiple layers simultaneously, with the ability to hide and show layers, reorder them by dragging, vary the transparency, assign different color palettes, etc. These features have been staples of offline viewers pretty much since the prehistoric beginnings of fMRI time, but they aren’t available in the current Neurosynth viewer or most other online viewers I’m aware of, so this is a nice addition.
  • The architecture is modular, so that it should be quite easy in future to drop in other alternative views onto the data without having to muck about with the app logic. E.g., adding a 3D WebGL-based view to complement the current 2D slice-based HTML5 canvas approach is on the near-term agenda.
  • The resolution of the viewer is now higher–up from 4 mm to 2 mm (which is the most common native resolution used in packages like SPM and FSL). The original motivation for downsampling to 4 mm in the prior viewer was to keep filesize to a minimum and speed up the initial loading of images. But at some point I realized, hey, we’re living in the 21st century; people have fast internet connections now. So now the files are all in 2 mm resolution, which has the unpleasant effect of increasing file sizes by a factor of about 8, but also has the pleasant effect of making it so that you can actually tell what the hell you’re looking at.

Most importantly, there’s now a clean, and near-complete, separation between the HTML/CSS content and the JavaScript code. Which means that you can now effectively drop the viewer into just about any HTML page with just a few lines of code. So in theory, you can have basically the same viewer you see in the demo just by sticking something like the following into your page:

 viewer = Viewer.get('#layer_list', '.layer_settings')
 viewer.addView('#view_axial', 2);
 viewer.addView('#view_coronal', 1);
 viewer.addView('#view_sagittal', 0);
 viewer.addSlider('opacity', '.slider#opacity', 'horizontal', 'false', 0, 1, 1, 0.05);
 viewer.addSlider('pos-threshold', '.slider#pos-threshold', 'horizontal', 'false', 0, 1, 0, 0.01);
 viewer.addSlider('neg-threshold', '.slider#neg-threshold', 'horizontal', 'false', 0, 1, 0, 0.01);
 viewer.addDataField('voxelValue', '#data_current_value')
 viewer.addDataField('currentCoords', '#data_current_coords')
 viewer.loadImageFromJSON('data/MNI152.json', 'MNI152 2mm', 'gray')
 viewer.loadImageFromJSON('data/emotion_meta.json', 'emotion meta-analysis', 'bright lights')
 viewer.loadImageFromJSON('data/language_meta.json', 'language meta-analysis', 'hot and cold')

Well, okay, there are some other dependencies and styling stuff you’re not seeing. But all of that stuff is included in the example folder here. And of course, you can modify any of the HTML/CSS you see in the example; the whole point is that you can now easily style the viewer however you want it, without having to worry about any of the app logic.

What’s also nice about this is that you can easily pick and choose which of the viewer’s features you want to include in your page; nothing will (or at least, should) break no matter what you do. So, for example, you could decide you only want to display a single view showing only axial slices; or to allow users to manipulate the threshold of layers but not their opacity; or to show the current position of the crosshairs but not the corresponding voxel value; and so on. All you have to do is include or exclude the various addSlider() and addData() lines you see above.

Of course, it wouldn’t be a mediocre open source project if it didn’t have some important limitations I’ve been hiding from you until near the very end of this post (hoping, of course, that you wouldn’t bother to read this far down). The biggest limitation is that the viewer expects images to be in JSON format rather than a binary format like NIFTI or Analyze. This is a temporary headache until I or someone else can find the time and motivation to adapt one of the JavaScript NIFTI readers that are already out there (e.g., Satra Ghosh‘s parser for xtk), but for now, if you want to load your own images, you’re going to have to take the extra step of first converting them to JSON. Fortunately, the core Neurosynth Python package has a img_to_json() method in the imageutils module that will read in a NIFTI or Analyze volume and produce a JSON string in the expected format. Although I’m pretty sure it doesn’t handle orientation properly for some images, so don’t be surprised if your images look wonky. (And more importantly, if you fix the orientation issue, please commit your changes to the repo.)

In any case, as long as you’re comfortable with a bit of HTML/CSS/JavaScript hacking, the example/ folder in the github repo has everything you need to drop the viewer into your own pages. If you do use this code internally, please let me know! Partly for my own edification, but mostly because when I write my annual progress reports to the NIH, it’s nice to be able to truthfully say, “hey, look, people are actually using this neat thing we built with taxpayer money.”