It was great taking part in the Sex Tech Hack 2016.
Hosted at Goldsmiths University of London by the Hacksmiths society, the hackathon explored the themes of intimacy, companionship and sexuality in an increasingly technological and computer-based society.
I went alone to the event and I was really happy to find out that there were other people interested in A.I. and Machine Learning, just like myself. So, it was not hard to team up with three other nice chaps and put together a plan for the weekend.
My group was particularly interested in finding out connections between our instincts, desires, and mental drives and what our tastes can tell us about ourselves.
We thought that pornography, and the basic instincts expressed through its consuming in our society, would have been a perfect starting point for our hack.
We split into two sub-units: a research pair and a coding pair. Our research pair didn’t take long to start finding connections between certain geographical areas or population groups and the most searched porn tags within those sections of the population.
They then found apparently irrelevant “most searched” words or phrases used by those very same groups of people and drew statistical links. Furthermore, they used published papers and statistics from companies with access to large users’ tastes datasets to validate such connections. Just as an example, we discovered that the dating website Cupid had found a statistically backed connection between being fond of beer and being open to sex on the first date.
Once we had both a framework for our software and the information we wanted, we created a custom learning model in Clarifai and we trained it with hundreds of different porn pictures. We decided to have at least 80 picture per porn category, then we trained the algorithms by telling them which “concepts” to see in each picture.
We discovered, for instance, that people from areas where the website PornHub claimed “Lesbians” was the most searched porn tag were also likely to search for McDonalds when it came to restaurants. Other interesting links were between people searching for “Black Women” or “Asian Housemaid” in areas with either a history of racial discrimination or a strong tendency to vote for politicians opposing racial integration.
We kept feeding the algorithms with pairs of images and concepts until the tool was ready to recognise apparently unrelated tags in porn images of all kinds. Even images it had never seen before.
Just before the presentation of our mock website, we tested our software and checked connections both ways using a restricted custom search engine.
We tried to select “appealing” or “intriguing” sex pictures based on individual, mostly irrational, tastes and saw what “tags” the machine would spit out in response. It was definitely hilarious to see that, after selecting a bunch of hard-core “gangbang” images, people were told they probably liked a certain type of food. However, it was also insightful because we saw a strangely disquieting relationship between sexual/pornographic tastes and other, apparently random, human desires.
The reverse was also interesting, although we could not spend a long time testing it. We hacked a few google searches of the words/concepts we had used for our tagging model and inserted a few pornographic images among the results. A few pictures that our software had classified as relevant for the concept “Jeremy Corbin”, for example, were placed among the results of a search using that tag.
Unfortunately, due to lack of time and people for testing, it was hard to come out with strong conclusions. In a way, we had set out to discover how the very same online media tagging system, which is responsible for the “social bubble” of taste-based feed suggestions in which most of us live, could be turned on itself and used to suggest apparently (but really not-so-much) random search results. In the end, due to lack of rigorous testing and large population samples, we couldn’t say for certain whether we had found a way to so.
The fact that most people reported an odd sense that the machine was guessing “right”, though, was a very positive indicator. Ultimately, our intuition could have well been correct and a machine learning based pattern-recognition approach like the one we explored could actually help us in three tangible ways if correctly implemented:
- It could show us hidden links between desires we might have in apparently unrelated areas, as well as -perhaps- links between what arouses us and our personality.
- It could help our online searches and product-suggestion tools provide us with apparently random results, which in fact we may find interesting even though we had never searched for them directly.
- It could help media providers -in this case porn providers, but others may use the same technique- tag their media with apparently unrelated words, which would, in fact, help their target users find such content whilst searching for something else.
After the Hackathon, I continued to study machine learning algorithms and decided to implement a few from scratch myself.
Here you have two pictures of two C++ programs I made using openframeworks:
In the first one (GitLab repo), I equipped a physical agent (the black triangle) with a perceptron-based “brain” and watched it adapt to its environment in its pursue for a prey (the red square).
In the second one (GitLab repo), I created an animated structure for a Neural Network, which can help visualise how a neural network works and can be used to develop evolving networks using Genetic Algorithms -the network can, in fact, generate itself using specific instructions that can be encoded in a “genotype”.
One question ultimately begs for an answer, and I don’t think anyone really knows it: will my “seeing” software, “steering” triangles. and “evolving” digital brains truly become intelligent one day?
By this I mean, will they ever exhibit not just an “algorithmically” smart behaviour, but an actually creative one?
I think so, but there is only one way to find out: keep calm and carry on coding 😉