Sunday, May 12, 2013

Sunday Morning Insight: Computational Cooking, you won't see food the same way anymore.

This entry will be short because there is so much to say and so much to think about that I won't do it justice if I am supposed to write something all encompassing. 

Herve This is known in France for advocating Molecular Gastronomy (go read the wikipedia entry, it's worth it). This is no small feat in a country that counts as many cheese as there are regions, where wine is celebrated around the world and whose talented chefs makes a large portion of the three stars catalog in the Michelin guide. My interest in this blog entry is not in the issues raised with a scientific approach to cooking, rather I am interested in showing a parallel between Note by Note Cooking and  issues we have mentioned here on Nuit Blanche. In order to do this, I first recently asked Herve to point me to a reference and he kindly pointed me to this one [1] :

Solutions are solutions, and gels are almost solutions by Herve This

Abstract: Molecular gastronomy is the scientific discipline that looks for mechanisms of phenomena occurring during dish preparation and consumption. Solutions are studied because most foods, being based on animal and plant tissues, are gels, with a liquid fraction and a continuous solid phase. This is why food can be studied in situ using liquid NMR spectroscopy in the frequency domain (isq NMR). Using such tools, processes of the kind F@M → F' @ M' (where F stands for the food matrix, M for its environment, and @ for inclusion) were investigated for various processes as classified using the complex disperse system/non-periodical organization of space formalism (“disperse systems formalism”, DSF). As an application of these studies, “note by note cuisine” was promoted as a new paradigm for culinary activities.
Here are the elements of interest (but go read the whole paper, I'll wait). While describing the possibilities of dishes, here it is how it is described:
A simple calculation shows the immensity of the world to be discovered. If we assume that the number of traditional food ingredients is about 1000 and if we assume that a traditional recipe uses 10 food ingredients, the number of possibilities is 1000 to the power 10, or 10 to the power 30. However, if we assume that the number of compounds present in food ingredients is about 1000, and if we assume that the number of compounds that will be used in note by note cuisine is of the order of 100, then the number of possibilities is about 10 to the power 3000... And, in this calculation we did not consider that the concentration of each compound can be adjusted, which indeed means that a whole new flavor continent can be discovered.
I can bet that only a few of these dishes can actually be made or are stable and I would venture that this a sparse set. That sparse set is still larger than the set we currently know. Herve This goes on to describe the manner with which one can describe the complexity of dishes as the composition of operations on different elements:  
With this clearer scientific program, it is convenient to come back now to the question of “bioactivity”. Most formulated products, and in particular food, are systems (often of colloidal nature [31]) which display “bioactivity”: this means that they can exchange “bioactive” compounds (bc), i.e. compounds that have the potentiality to interact with biological receptors. In some cases, a physical binding is needed to trigger physiological effects (olfaction, taste, trigeminal effects...) but for vision the effect is indirect, and for receptors inside tissues, a transfer into the blood system is needed, sometimes after modifications during digestion. Here let us remark that even compounds that would be trapped by the swallowed bits of food are encompassed by the given definition, such as tasty ions adhering surfaces, complexed salivary proteins on some compounds in food; after all, absorption is a negative release.
Bioactivity being defined, one of the main issues of molecular gastronomy (and other scientific disciplines) can be investigated, i.e., the relationship between the modifications of the structure of food and the actual bioactivity of food (Fig. 1). In order to study this question, a formal description (“disperse systems formalism”, DSF) of the structure of food systems was proposed [42]. Recently, modifications of this formalism were introduced as shown below. 
In order to give a formal description of formulated products, DSF considers “objects” and “operators”, after the definition of a “reference size” (rs), as an order of magnitudes of considered objects (the same kind of description can be made at any order of magnitude for sizes) [44]. For example, if the diameter of a plate is chosen as rs(~3 × 10–1 m), then only the various food items of diameter (defined as the largest dimension) between rs/5 and 5 rs should be considered, all objects of different diameters being excluded from the description.  “Objects” can be of various “physical dimensions”: D0, D1, D2, and D3. Here, D0 stands for objects of physical dimension equal to 0 (“dots”), i.e., objects whose size in the three directions of space is more than one order of magnitude lower than rs; D1 stands for “lines” (with only one dimension of the same order of magnitude as rs), D2 for surfaces (with two dimensions of the same order of magnitude as rs), D3 for volumes (all dimensions of the same order of magnitude as rs). If necessary, Dx objects could be considered, x being non-integer, and these objects then being fractals [45].
If necessary, the rs can be explicitly added in brackets at the end of the formula (units should of course be in International System of Units). For example, D1 [10–5] would indicate a linear structure whose length is of the order of magnitude of 10–5 m (and, accordingly, whose radius is more than one order of magnitude lower). 
The organization of a system at a particular scale is then described from the various parts Dk (k ∈{1,2,3}) with some spatial relationships between these parts described using operators: the operator “/” represents random dispersion; the operator “×” represents intermixing of two continuous phases; the operator “+” represents coexistence of phases; the operator “@” represents inclusion; geometrical operators such as “σx ”, “σy”, “σz” represent respectively superposition in the directions x, y, and z(but any particular direction could be given by the Cartesian coordinates of a vector, such as in (u, v, w), or even other coordinates systems such as {r, θ, ϕ} for spherical organization). Other operators may be added when needed.
Such systems are frequent in food [48], in particular because plant and animal tissues, made of cells whose smallest dimension is of the order of 1 μm, are colloids according to IUPAC definition: cell aggregation in tissues makes formally non-connected gels, contrary to gelatin gels, which are connected gels, water forming a continuous phase in the continuous solid phase due to collagen molecular associations by triple helixes. Emulsions are also frequent in the kitchen (mayonnaise, aioli, wine sauces with butter…) [49]. When complex systems are considered (e.g., multiple emulsions), physics generally focuses on the interface [50], i.e., local descriptions of macroscopic systems, or on some thermodynamic properties. However, this has two main disadvantages. First, the global description of the systems is lost. Then, in more complex—but familiar—systems, such as potato tissue or ice cream, the denominations are rather complex. Potatoes, for example, are mainly “suspensions dispersed in gels”, as amyloplasts (solidstarch granules having a radius lower than 20 μm), are dispersed in the cytoplasm of cells (water or gel, depending on the description level), this phase being itself dispersed in the network of cell walls responsible for the “solid” behavior of the whole potato [51,52]. Ice cream is another example of a complex food system that should be called “multiple suspension/foam/emulsion”, as gas (air) bubbles, ice crystals, protein aggregates, sucrose crystals, fat (either crystals or liquid droplets), etc. (depending on the “recipes” and on the process used) are dispersed in an aqueous solution [53]. On the other hand, the names “potato” or “ice cream” are probably not admissible names in physics textbooks because they are imprecise and restricted to a particular food.  This is why DSF was introduced, based on the same idea as the one proposed by Lavoisier for chemistry [54]. In colloids, the phases are gases, solids, or liquids, and in food, liquids are mostly water and mixtures of triacylglycerols. Accordingly, the symbols G, O, W, S, respectively, stand for “gas”, “oil”, “water”, “solid”; of course, other symbols such as E (for ethanol) could be added if necessary (this would be useful in fields other than food). Adding some rules gives more coherence to the formalism. First, some formula can be simplified.
For example, D0(G)/D3(G) or D1(W)/D3(W) are respectively reduced to D3(G) or D3(W). Then, the various components of a sum (operator +) must be written in alphabetical order. For example, custard {which is not an emulsion D0 (O)/D3 (W), contrary to what is frequently published in culinary textbooks [55]} is made of oil droplets O (from milk), air bubbles G (introduced during the initial whipping of sugar and egg yolks) and small solid particles S (due to egg coagulation during thermal processing), and should be described as [D0 (G) + D0 (O) + D0 (S)]/D3 (W). This rule is the key to the uniqueness of formulas associated with physical systems. Repetitions can be described by exponents. For example, egg yolks are made of concentric layers called light and deep yolk, deposited, respectively, during the day and the night; their number is about 9, as shown on ultrasound scan pictures. As each layer is composed of granules (S) dispersed into a plasma (W) [56], the full yolk could be described as [D0 (S)/D2 (W)]@9 .The basic formalism can be enhanced to give more precise descriptions of systems. For example, the quantity of each phase can be added as a subscript. For example, D0 (O)200/D3 (W)5 would describe an oil into water emulsion at the limit of failure, with 95 g of oil dispersed in 5 g of water (the oil droplets would have a polyhedral shape) [57]. Using such subscripts, conservation laws can be used. For example, the overall making of a mayonnaise could be written as  
O95 + W5– E W→ O95/W5(11) where E W stands for mechanical energy
How helpful is DSF for studying real biological structures with a high level of complexity? First, the reference to a rsis one way to bring some order into the complexity of real systems. Then even when a complex object such as a living cell is considered, the fact that some particular structures are detected with various observational tools demonstrates that the formalism applies. Of course, the description can sometimes be cumbersome, such as for the Golgi apparatus [43], but the use of “random distribution operators” such as “/” is a way to get at least a partial description of such complex systems. Moreover the number of operators could be increased when new particular cases are encountered (up to now, even in such complex fields as sauces, only the four operators given above were useful). Specifically for food and solutions, DSF has value in differentiating the various kinds of gels.....
This is important because, as said before, plant tissues and animal tissues are gels, as well as gelatin or pectin gels. However, gelatin gels are very different from plant tissues, as their liquid phase is continuous, which is not the case for the liquid phase of plant tissues, where it is localized in the cells (of course, this is an approximation: some continuous liquid phase exists in the vascular tissues); and plant tissues are very different from animal tissues in muscles, as muscle fibers are elongated cells of length up to 20 cm. The formulae for these three different gels are therefore different: D3(W) × D3(S) for gelatin gels; D0(W)/D3(S) for plant tissues; D1(W)/D3(S) for muscular tissues. The usefulness of DSF has been shown many times for the description of systems and for inventing new systems, but how is it useful for science, specifically for the study of bioactivity and matrix effects? A good way to use it is first to consider that bioactivity can be considered as phenomena occurring in the F @ M or F σ M. As one should have rational reasons to study particular systems, ranking systems according to the “complexity” of their formula in DSF is a strategy. Of course, the number of different phases is to be considered first for this ranking, but also the organization of phases within the system. This organization is described by the various operators, which means that a ranking of operators by order of “complexity” is needed.
So in this burgeoning science, there seems to be this ability to decompose a complex scene (dish) through the composition of few operators with complexity being a factor. We have seen something very similar recently in Matrix Factorizations and the Grammar of Life. I also note that in some other parts of the paper, Herve points to an issue of inverse problerm familiar to computational biology ( see Sunday Morning Insight: The extreme paucity of tools for blind deconvolution of biochemical networks ).

What about the sensing and the data being generated so as to provide some quantification that can eventually be used in the matrix decomposition highlighted above ? In other words, what sort of sensing allows one to explore dishes and thereby enable support for the assumption of low complexity of the unit operators ?
Now, with a means of ranking the various systems by order of “complexity”, the question arises: how can the transfers of bc in matrices be followed? Recently, a new analytical method called in situ quantitative NMR (isq NMR) was proposed for this purpose. It is based on the fact that gels enclose a liquid component. This new method was first studied for the analysis of saccharides and amino acids, and organic acids of plant tissues [58,59]different: D3(W) × D3(S) for gelatin gels
Oh! isq NMR, which sounds a lot like fingerprinting NMR

In short, we have a large set of potential dishes, a potential smaller set of viable ones, a set of rules describing a grammar and an innocuous way of producing larger amount of data through NMR or image processing) than before. Notwithstanding the ability to probably study some of these dishes through hyperspectral cameras (which is an issue we come back often to). All this seems to be ripe for new kind of exploration. 

In response to queries by interested readers, Herve recently wrote this blog entry [3] and wrote a page summarizing the different questions to be elucidated (it's in French [4] and in English here). I see very little in the way of data processing (besides image processing of elephoresis at the end of the list).

Herve's blogs and sites are: 

Credit photo: Ça vient d’avoir lieu !

Join the CompressiveSensing subreddit or the Google+ Community and post there !
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.


Leslie N. Smith said...

IBM is playing in this field - I recently saw at article at "Try The First Recipe Devised By IBM’s Supercomputer Chef". After describing the magnitude of possible recipes, it says:

“You generate a million new ideas, but of course a million isn’t useful,” Varshney says. “You want to rank them by which will be perceived as flavorful and which are novel.”

Igor said...


This is very interesting. I think though that Herve's effort aims at a bottom up discovery process. But I am sure that with a good use of AI one could get there as well.