Words:Ritesh Kumar, Rishemjit Kaur & Amol P. Bbhondekar Photo:Stian Gregersen ArtDirection:Amalie Winther & Mikal Strøm
The sense of smell has been the least understood of all our senses. We use it inadvertently or advertently, to check the spoilage of food, danger signs (gas-leaks, fire etc.), and in relishing food and beverages. There are many animals in our eco-system which also use it for their mate identification. In spite of many attempts at understanding the smell, we still don’t have an objective way of telling how a molecule will smell like. The part of problem lies in the large number of varying verbal descriptors used to describe the ‘olfactory-perception’. Till date an intuitive, useful and objective way of representing the perceptual descriptors has not been done. The potential purpose of a certain odor system or olfactory perceptual space classification is diverse, ranging from allocation of odors in classes with appropriate labels to a hierarchical representation by which they may be distinguished.
The other part is that, we do not know how many types of different smells exist and also we do not know how odors are arranged in a perceptual space. Also, smell perception is arguably more complex than the color and audio perception. On one hand, for example, the perception of light, identified by a certain wavelength can be described by the perceived hue, brightness, and saturation and on the other, perception of sound can be identified by frequency distribution of sound waves. Olfactory perception has not had such objective representations and is likely to have more than three dimensions (opposed to vision). It has been assumed that the perceptual space of olfaction is much more complex and has many more dimensions than those of the other senses.
We set out to answer these complex questions by at first gathering a large number of odor evoking molecules (Leon and Johnson, GoodScents, SuperScent, Flavornet and Sigma-Aldrich) and their smells (which are actually, words like “green”, “sweet”, “fat” etc.) and projecting these large number of verbal descriptors according to their frequency of use on a graph as it is generally done in social network analysis. Some of the important questions that we sought to answer were related to its structural organisation and semantic positioning. We therefore put forward some general charactersitics of the graphs and their analysis. The distribution of the verbal descriptor on the graph seems to follow a “power law”. The power law does not have a well-defined mean or variance which makes it very difficult to assess the charactersitics of the distributions. It seems to be ubiqutous in natural and man-made systems.
The usage of the words in English, the internet, cities ranked by population and wealth of individuals etc. also follow this law. Figure 1 shows the degree distribution of the graphs in log-log plot. This distribution leads to a phenomenon of “scale-free”, which has been observed in distribution of internet. This leads to having hubs in the network. In our case, it means the odour space is dominated by few most frequently occurring perceptual descriptors or hubs in a network, thereby suggesting preferential attachment of the perceptual descriptors.
The dominance of these hubs in almost all the networks suggests universality of the olfactory perception. A similar kind of observation has been reported in linguistic research in which the frequency of occurrence of a particular word has been seen to be inversely related to the rank of that word in the corpora. The underlying origin of such phenomenon has been given the name as “theory of least action”, i.e. people tend to speak those words which they think will convey the broadest of information on a given topic, which is true for expert or a layman. So, in the event of smell reporting, the subject would tend to speak first the broader meaning words and then she/he would tend to speak more specific and related words. Moreover, odour perception in general is associated with objects which we encounter and it is an integration of inherent odour characteristics with the psychophysical condition of a perceiver. Odour association with objects becomes difficult for a perceiver in absence of a visual cue’
Overall, we found that English as a language is not descriptive enough to represent the olfactory perception due to the presence of some perceptual descriptors which are hubs or leaders in the graphs, implying, “if you can’t understand, at least speak this”. The use of very specific words are rare, but when used, they are mostly used in combination with the more obvious descriptors like “fruit”, “floral”, “meat”, “green”, “wood” etc. It could confirm the “theory of least action” which suggests people tend to speak those words which they think will convey the broadest of information on a given topic, which is true for expert or a layman . So, in the event of smell reporting the subject would tend to speak first the broader meaning words and then if asked or pressed more she/he would tend to speak more specific related words.
The descriptors are then grouped together using community detection technique which is often done in social networks. The smells very interestingly fell into very specfic groups which could be related to their origin and perception similarity (figure 2). For example, one of the perceptual groups is of “cream”, “butter”, “cheese”, “caramel”, “animal” etc. Similarly, the notes “medicinal”, “mint”, “smoke” and “phenol” are observed together along with “leather”, “camphor” and “pine”. The descriptors “fatty”, “wax”, come along with “tallowy”, “aldehydic” and “soapy”. The descriptor “fatty”, is generally used to describe smells that relate to oil or wax, “fatty” along with “tallowy”, “aldehydic” and “soapy” are produced by short-chained aliphatic aldehydes and hence can be regarded as belonging to the same group (figure 2).
The community of “fish”, “onion”, “sulfur”, “garlic”, “meat”, “alliceous” along with “earth”, “burnt, “coffee”, “roast”, and “bread” suggests the culinary part of the perceptual space (figure 2). We believe this part of human odour space is not biased by perfumers’ observations, but generalized perceptions. These smells if presented individually, will usually elicit unpleasant notions from the subjects, but, when presented with cues e.g. food with garlicky smell will usually result in a pleasant response. Similarly, for “onion”, “meat”, “beef” etc., these smells are learnt along with the other objects and are mostly part of our cooked food items.
Most interestinly, the togetherness of these perceptual descriptors are not because of their semantic relatedness in English language. People generally don’t use the verbal descriptors, used in reporting smell, together in their writing and speaking. This fact was statistically proven by running similar kind of analysis on a large corpus of English text comprising of books, conversations etc.
Besides this, the molecules when projected in a non-linear space separately from the perceptual and physico-chemical side overlap significantly in a non-linear dimension. This fact was shown to be correct by using a amthematical measure called as Hubert Index. It affirms that, the molecules occupy similar clusters positions in different spaces. It affirms the structure preserving property of the odour spaces. This also gave us the encouragement to cluster molecules based on perceptual qualities and define categories to the molecules.
We also developed a machine learning technique to predict the smell of a molecule based on its physico-chemical property i.e. you present our method the physcico-chemical properties of the molecules or present the molecules itself, it will spell out the perceptual class it belongs to and hence the smells that it can possibly elicit.
It should be noted that the smell is some times a very individual specific experience. We need still larger data and larger verbal descriptions to be able to reach a very specific conclusion. The addition of neurological data would add still better element to it. The study of verbal descriptors present a very interesrting aspect of olfcatory perception. We believe such kind of analysis on different languages could present some more insights into understanding of smell.
figure 1: The degree distribution of the networks in log-log plot along with the fitted truncated power law. For all the networks, except SuperScent, the probability that a given perceptual descriptor connects with k other perceptual descriptors follows a power law (p(x) = x-┒) with┒Є [2, 3] or┒~2.
figure 2(a-g). Odour Network. The communities detected in the odour network of databases using community detection algorithm. The colours indicate the different communities.
Reference: Ritesh Kumar, Rishemjit Kaur, Benjamin Auffarth, and Amol P. Bhondekar. “Understanding the Odour Spaces: A Step towards Solving Olfactory Stimulus-Percept Problem.” PloS one 10, no. 10 (2015): e0141263.