Well, I guess it’s time for this graduate student to emerge from hibernation, the long sit, social torpor, which ever you’d like to refer to as the ‘data’ season – especially as I come up on my 1-year Durhamiversary.
Since it’s been too long since my last appearance on this blog, I guess I should catch you guys up on where the acceleration project is at. Following the holidays and attending a few conferences to present my prior work in January, it was time to get cracking. In the past 9 months, I have finished all the focal video behavioural analysis, as you have likely heard loads about in past years on this blog regarding behavioural observation approaches, and have become deeply entrenched in the accelerometer data. Other than a fun break to help teach the undergraduate inter-tidal field course in the middle of the summer, I have been sat squarely at my desk refining my approach and coding away. Given that each individual accelerometer file is so large, as explained in a prior post, this task is a bit more like wading through an ocean. The key to tackling and coding the analysis for such large datasets? Sub-sampling! By turning a large file of, say, 43 million sample points into a much more ‘manageable’ chunk of 2 or 3 million, sorting out the trial and error of even the most complex code becomes much less frustrating. This is especially true when the size of your dataset takes up a large proportion of your overall computer memory capacity.
My goals for this project include the use of machine learning to help identify behaviours of interest. But what exactly is this machine learning thing? Before I get into the specifics of the approach, the use of these techniques has become fairly commonplace in our everyday life. Have you ever used an app to monitor your sleep patterns? Track the number of steps you’ve taken in a day using your wristwatch? Well, apps like these tend to use machine learning techniques to interpret signals that come in through built-in sampling devices in your phone or accessory of wearable technology. These signals can be used to interpret your movement and classify that movement signal to what you are likely to be doing at that moment in time.
We are applying a similar technique with our work with the seals! But how do these devices manage such a feat? Let’s walk through a simple machine learning workflow:
The first step is to train a model to interpret a signal in the way that you have labeled it as. Essentially, this is what would be referred to as supervised learning, in which one would input a training data set in which you have already classified each segment or signal. While there are many approaches to doing this, once you have what is considered your best model, one generates a decision tree-type set of criteria to classify your data with (think of a series of flowchart questions of yes/no).
The second step is to then use these criteria developed in the previous step and apply them to new data. The better your model is as developed in the first step, the better the predicted results will be – and the less confusion they will generate in what the true signal was telling the model (e.g. not mixing up running with sitting down).
Pretty simple, eh? Well, at least in theory. But, fear not, I have made great strides in the last few months and am still on track for my original analysis goals prior to…
~ THE UPCOMING FIELD SEASON ~
That’s right boys and girls, that season that you all know and love will soon be coming back to the blog. The seal crew will be heading out toward the end of October to begin another season on the Isle of May! CAPARs, tags, heart rate monitors, angry mums, and fluffy pups – oh my! There will be a new face around as well as we welcome our new MBIOL, Jodi, at the start of term (we will hear from her about herself and her project once term starts back up). The lab is slowly turning into well-organized piles of kit and to-do lists – categorizing not only tasks to accomplish before the field season, but meticulous lists of chocolate preferences to purchase. Stay tuned for more exciting action as the season progresses and we head out into the fray.
[I am also hoping to do some video/GIF segments discussing what seal behaviours look like as well as life on the colony and our day-to-day comings and goings in the field – after all, the kids love the videos/GIFs – stay tuned!]