At this year's SciPy 2020, a talk and proceedings paper caught my eye: "Software Engineering as Research Method". (Here are links to the paper and the talk.) In it, the authors, Mridul Seth and Sebastian Benthall, detailed the benefits that software skills bring to the academic and industrial research communities, both from the perspective of making scientific progress and from the perspective of pedagogy.
I've been doing some reflection of my own on how software skills have helped me tremendously, as a data scientist. Here's my thoughts on them, using the example of random number generators and probability distributions.
In building scientific software,
we are recognizing pragmatically useful categories around us,
and formalizing them in a language.
Let's look at a really simple example of drawing
from a Bernoulli distribution.
If we were to write a Bernoulli draws generator in Python code, we might implement it as follows:
from random import random p = 0.5 draws =  for i in range(num_draws): draws.append(int(random() > p))
This is purely imperative style programming. Without going too deep (pun intended) into definitions, by using mostly built-in primitives of the Python language, we are operating at a fairly low-level paradigm.
Now, let's imagine a world where the Bernoulli distribution object does not exist. I have a collaborator who wants to build on top of Bernoulli distributions to make other things. They would have to copy/paste this imperative block of code into their programs.
What can we do to alleviate this, then?
One thing we can try is to encapsulate it in a function! The implementation might look like this:
from random import random def bernoulli(num_draws, p): draws =  for i in range(num_draws): draws.append(int(random() > p)) return p
But wait, what are the semantics of
Is it the probability of obtaining class 1, or class 0?
The behaviour isn't documented in the class,
so let's add some of it in.
from random import random def bernoulli(num_draws: int, p: float) -> List[int]: """ Return a sequence of Bernoulli draws. :param p: The probability of obtaining class 0. :num_draws: The number of Bernoulli draws to make. Should be greater than zero. :returns: A list of 1s and 0s. """ draws =  for i in range(num_draws): draws.append(int(random() > p)) return draws
Now, we can do a few things with this implementation. Firstly, we can write a test for it. For example, we can assert that numbers drawn from the Bernoulli are within the correct support:
def test_bernoulli(): draws = bernoulli(10, 0.3) assert set(draws).issubset([1, 0])
Secondly, we can compose it in bigger programs, such as a Binomial generator:
def binomial(num_draws: int, n: int, p: float) -> List[int]: draws =  for i in range(num_draws): num_ones = n - sum(bernoulli(n, p)) draws.append(num_ones) return draws
Please excuse my lack of docstrings for a moment!
This is all good and such, but there is another thing we need to consider. A random number generator exists logically as a sub-category of things that a probability distribution can do. There are other things that we do with probability distributions, such as evaluating data against a probability distribution's density/mass function. After all, generating numbers isn't necessarily the only action we want to take with a probability distribution.
The probability density function is unique property of a probability distribution; it's also an invariant property. As such, it is something we can instantiate once and forget about later. Perhaps using an object-oriented paradigm might be good here. Let's see this in action:
class Bernoulli: def __init__(self, p): self.p = p def pdf(self, data: int): if data not in [0, 1]: raise ValueErrror("`data` must be either 0 or 1!") if data == 0: return p return 1 - p def rvs(self, n: int): draws =  for i in range(num_draws): draws.append(int(random() > p)) return draws
Now, with this object, we've made using the category of Bernoulli distributions much easier!
b = Bernoulli(0.5) draws = b.rvs(10) # draw 10 numbers pdfs = [b.pdf(d) for d in draws] # calculate pdf of data
An object-oriented paradigm here also works well,
because the way
.pdf() have to be implemented
for each probability distribution is different,
but the inputs to those functions are similar
across the entire category of things
called "probability distributions".
I want to first state up-front that this essay is not one championing object-oriented programming. Those who know me actually know that I much prefer functional programming. Rather, this essay is about how leveraging the natural programming paradigms and data structures for a problem or model class gives us a way to mastering that problem or model class. After all, building useful and verifiable models of the world is a core activity of the quantitative sciences.
By exploring and structuring probability distributions in a logical fashion, my pedagogical exploration of probability distributions as a modelling tool is reinforced with the following main ideas:
And by structuring a probability distribution as an object with class methods, our mental model of the world of probability distributions is made much clearer.
Yes, research discipline, and in both ways. Here’s what I mean.
Discipline (a la "rigour") in thought means thinking clearly about how the thing we are working with fits most naturally in relation to other things. Nudging ourselves to make better software makes us better thinkers about a problem class. Bringing this level of discipline gives us an opportunity to organize the little subject matter discipline that we inhabit.
As a discipline (i.e. "way of thought"), well-written software also enables us to build on top of others' thoughtfully laid out tooling. As such we are
Contrary to a generation of us that might think that "choice in everything is the ultimate", a stable base is what a field really needs in order for a discipline to flourish and build on top. That means agreed-upon foundational definitions, that have stood up to the test of time. The base being well-tested means we can rely on it; having longevity is also a very good prior for innate invariance over time.
If we did not all agree on using shared primitives from the language, we would never have been able to build on top of the shared primitives to define a probability distribution the way we did. And without defining a probability distribution in a uniform but useful way, we leave no easy path to build on top of that to make other things, such as probabilistic programming languages!
As such, shared stability in the base gives us confidence to build on top of it.
In the SciPy talk, and in the proceedings paper, Sebastian and Mridul make the point that building software enables "constructionist" learning, i.e. a learner learns when playing with software. I think it is plausible to extend the idea, in that a learner can learn well when building software.
The act of organizing the world around us into categories with properties and relationships is made concrete when we build software. By organizing the conceptual world into layers of abstractions and sequences of actions, we bring a clarity of thought to the flow of work. I would argue that this can only help with the construction of a consistent and productive view of a field.
In addition to these conceptual benefits, learning how to structure a codebase logically and write software tests for your code helps with reproducibility. Through tests, our code is known to run correctly, and hence can be depended upon as stable. If structured logically, they can be creatively composed to fit a variety of workflows that may show up, which can be reliably automated as well.
Some of the work I engage in at NIBR
involves fairly non-standard models.
These are the models for which
might not necessarily be available,
and for which I myself might not necessarily have
prior training in to rely on.
As such, there is a process
of "learning on the job"
each time I encounter a new model class.
Over the past two years, for me, that has been recurrent models,
such as recurrent neural networks (RNNs)
and autoregressive hidden Markov models (AR-HMMs).
In both cases, I have worked with interns
to build a software library for the model core.
In there, we not only worry
about the implementation of the core,
we also worry about how we will use the model,
and strive to make its API semantics
match and leverage that of
the surrounding PyData ecosystem,
such as following SciPy's API for distributions,
scikit-learn's API for model construction,
jax for speed and expressiveness.
In both model’s cases, we started by implementing the core of the model. This refers to the key algorithmic steps. The exact process of how we get there usually involves some trial and error, but the general idea is to work out the algorithm for a single “unit” of data, and work out how that unit of data passes through the model until it reaches the end point that we want, whether that is a “prediction” or a “likelihood” calculation. Implementing the core forces us to more clearly learn the semantics of the model class and the problem we are trying to solve, and hence helps us chart out what kind of internal API we might need to be productive.
The act of writing tests for the library helps with cross-checking assumptions. By checking against "well-known" or "trivial" cases, we can check that our understanding of how the model ought to work is correct at a basic level. By leveraging property-based testing, we can also check that our understanding of the model is also more generally correct.
We invest time up-front working on documentation. By documentation, I mean at least the following:
Investing time up-front for documentation is actually a great thing to do. By putting in workflow examples, we actually leave for ourselves copy/paste-able workflows that can greatly speed up the ramp-up phase of a new project that leverages what we have built.
In building dog-fooded, usable software around a topic, I believe we are empowered to leverage the act of organizing knowledge into a usable software package to also learn that new topic. I also build it with a view towards an open source release. That helps increase the motivation level for learning it well, because I am placing reputation points on the line! (And not just my own: usually my interns’ reputations are on the line too!) Having the public audience is something personally motivating. Having written all that, I will acknowledge too that this is not uniformly true for everybody!
Whenever I go to my tribe of data scientists at work and advocate for better software skills, I often get a myriad of responses. A small sliver agree with me and actively put it into practice. A much larger slice agree, other factors stop them from getting started; these might be lack of knowledge, or too much external pressure to deliver results, and as such no action is taken.
With this essay, I attempted to extend what Mridul Seth and Sebastian Benthall elaborated on. In particular, I wanted to show that “writing good code” is a natural extension of “thinking clearly about a problem”, and thus suggest that leveraging software skills gives us superpowers in our data science work.
I started with the concrete task of "generating random numbers", showing how thinking clearly about how to structure probability distributions as classes with innate properties gives us clarity in how to interact with them. I then showed, through a personal example, how common and simple software practices can help the practice of data science.
It is easy to over-read this essay and start suggesting that data science teams should be run like an engineering team. This conclusion would be far from the spirit of my intent. Rather, I am merely suggesting that it would be beneficial for data scientists to learn better software skills, as it would help us think more clearly about the problems we encounter, and help us accelerate our work. My hope is that you, a data scientist who reads this essay, will consider adopting and learning basic software skills and incorporate them into your work.