An intriguing article by Serensinhe at al reports an experiment in ‘using deep learning to quantify the beauty of outdoor places’.
This struck a chord, because of Kant’s assertion that aesthetic judgements are a philosophical hybrid. When I like a view, my aesthetic judgement is purely subjective (there is no ‘objective’ reason for it) – but yet I expect most people to share my pleasure. Wikipedia says: “A pure judgement of taste is in fact subjective insofar as it refers to the emotional response of the subject and is based upon nothing but esteem for an object itself: it is a disinterested pleasure, and we feel that pure judgements of taste, i.e. judgements of beauty, lay claim to universal validity (§§20–22). It is important to note that this universal validity is not derived from a determinate concept of beauty but from common sense (§40). Kant also believed that a judgement of taste shares characteristics engaged in a moral judgement: both are disinterested, and we hold them to be universal.”
Does this new work represent a means of squaring this circle – that is, showing why a subjective judgement can be widely shared? Well, no. The paper relies on a database of UK photographs, which were rated by users for how ‘scenic’ they are. (“Scenic-Or-Not presents users with random geotagged photographs of Great Britain, which visitors can rate on an integer scale 1–10, where 10 indicates ‘very scenic’ and 1 indicates ‘not scenic’.) So first of all:
1. no definition of ‘scenic’ is given
2. this is not examining human judgement, just repeating it
3. the word ‘scenic’ is subtly different from ‘beautiful’ or Kant’s word ‘sublime’ and this must have influenced the respondents. It carries a whole load of baggage – overtones of the Leisure Industry. tourism, and bureaucracy. (Indeed one of the authors’ claims is that their work may “such advances can help us develop vital evidence necessary for policymakers, urban planners and architects to make decisions about how to design spaces that will most increase the well-being of their inhabitants.”)
This is not wild romantic beauty that stirs the soul, this is Town Planning.
The figures are interesting, The database has 217,000 images, which have been rated ‘over 1.5 million’ times. In other words, an average of 7 ratings per image. (The authors say: “We only include images in our analysis that have been rated more than three times”.) So there is no attempt to select by rating scores, eg which are the most selected, which are the least. Seven eccentrics are enough to involve the rating of an image. (For example, if it includes a steam train, railway enthusiasts might do this!). The raters themselves are self-selected from people who were aware of and chose to use the ‘Scenic or Not’ game – which suggests that they are people who use the internet a lot, and are interested in geographical sites. (I had never heard of it.) It also suggests that they are people who ‘care’ that there environment should be ‘scenic’, whatever this means.
However, given these caveats, the study then uses other databases to identify attributes in the images, (“spanning from materials to activities (e.g. ‘wire’, ‘vegetation’, ‘shopping’.”) and category classifications (such as ‘mountain’, ‘lake natural’, ‘residential neighbourhood’ and ‘train station platform’.)
It then uses these to classify images that are found to be ‘scenic’. The conclusions are interesting:
1. “Visual inspection of a sample of the most highly scenic images suggests that they conform to widely held notions of beautiful scenery, comprising rugged mountains, bodies of water, abundant greenery and sweeping views….The sample of least scenic images suggests that such images are often composed of primarily man-made objects such as industrial areas and highways.” (Of course they conform to widely held notions, since they are based on the widely held notions of the people who chose them.)
2. “the definition of scenicness in urban built-up settings is more varied than in rural areas”, because not only ‘natural beauty’ but also “man-made features such as historical architecture and bridge-like structures” are included, such as ” man-made elements, particularly historical architecture around
the city, including Big Ben and the Tower of London”. (I think Kant might argue that we admire bridges or clock towers partly for ‘objective’ functional reasons, or for historical or symbolic reasons, when ‘beauty’ is supposed to meet no such criteria, and to be ‘useless’ by definition, or, as Kant would put it, ‘disinterested’.)
3. “Interestingly, however, we also see feature associations that contradict the ‘what is natural is beautiful’ explanation. In both models, man-made elements can also lead to higher scenic ratings, including historical architecture such as ‘Church’, ‘Castle’, ‘Tower’ and ‘Cottage’, as well as bridge-like structures such as ‘Viaduct’ and ‘Aqueduct’. Large areas of greenspace such as ‘Grass’ and ‘Athletic Field’ appear to be unscenic in both models. We hypothesize that this might be due to the fact that images composed primarily of flat grass may lack other scenic features.”
The authors toy with an evolutionary explanation of why we find views ‘scenic’ – “Evolution might have conditioned us to dislike certain natural settings if they have attributes that are detrimental to our survival … For example, we seem to dislike certain natural settings if they appear to be drab
or neglected …, or simply uninteresting to explore…. We also find that ‘No Horizon’ and ‘Open Spaces’ are also associated with lower scenicness. This accords with Jay Appleton’s theory of ‘prospect and refuge’ …, which suggests that humans have evolved to prefer outdoor spaces where one can easily survey ‘prospects’ and which contain ‘refuge’ where one can easily hide and avoid potential dangers.”
They end up claiming that “Our findings demonstrate that the availability of large crowdsourced datasets, coupled with recent advances in neural networks, can help us develop a deeper understanding of what environments we might find beautiful.” I think this is too large a claim. It may allow us to decide which environments a selection of people are likely to find ‘scenic’, but it hardly gives us a deeper understanding. Many of its hard-won ‘findings’ are in fact statements of the obvious. Any attempt to probe the depths will likely lead to ‘overtraining’ assumptions by local planners, eg ‘this view would look 6.7% nicer if it had a bridge nearby and 3.2% nicer still if it had a castle.’
All photographs with this post are my own. Any excuse to use them, really.