Somewhat recovered from the sudden onset of Doctorhood at the end of last year and the associated rush of productivity (photographically speaking) that comes with intense procrastination, I’m taking a brief moment to pause and think about my upcoming adventures in bubble and photo, both will involve a smattering of artificial intelligence.
Machine learning has always been a relatively pop-science-y buzzword for me, its power I didn’t fully appreciate until a few weeks ago when I got my hands dirty. In the space of a few days, I (re)trained a convolution-neutral-network to recognise photos I took which are stylistically “me” using thousands of my old photos and carefully selecting 35 hashtags on instagram where I crawled ~70k photos. Each of the hashtags selected has a particular look to them, which is useful when categorising something intangible and intrinsically artistic. The result was illuminating. The AI-algorithm managed an accuracy rate of 31.7%, which is not bad. [Bearing in mind that, by randomly guessing, the accuracy would be 1/36 ~ 2.8%]
More interestingly, what is “me”?
Unashamedly existential, I’m personally quite curious about such an “objective” measure since I’m consistently inconsistent with my photography and whether the machine can recognise this under my numerous transformations is definitely of interest. A set of photographic eigenvalues if you will, in other words.
Without further ado, here are six distinct examples with their associated Li numbers (L), where L = 0 correspond to nothing like me and L = 1 means almost surely me.
Comments. The Li number is surprisingly accurate as a measure of me-ness. The two which scored the lowest, numbers 4 and 5 (with L scores 0.038 and 0.069, respectively), do indeed correspond to something I don’t usually do. Number 4 because I was trying something new on purpose (i.e. film simulation + wide angle + night) and the machine picked up on that. On the other hand, number 5 is a distinctly tourist-y snapshot, which also isn’t very me, hence a low score. Number 3 with the highest score (with L = 0.795) is interesting in itself because I distinctly remembered that I was being “very me” when I took the photograph as I often like stark and abstract signs. The remaining photos between ~0.2 to ~0.5 can be described as a normal day in the office, being typically me-ish without being a copycat of earlier work.
So, this is obviously a first look at what a simple algorithm can accomplish but with time I think it will become a powerful tool to study the intangible objects such as photography which are often in the gray area and certainly not very mathematical in nature.
It is also quite fun and so in the future I will embed a “Li-ness” score dressed up as #thedroidlife on my instagrams posts :)