This article was published in Scientific American’s former blog network and reflects the views of the author, not necessarily those of Scientific American
Note: In the spirit of creativity, I've written this blog post in the style of an academic article. It is clearly not a true academic article. However, all of the information is factual and based on interviews with attendees of the 2013 Falling Walls conference and the creator of the artificial intelligence menu planning system himself - Lav R Varshney.
On supporting science journalism
If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.
ABSTRACT: To date most self-appointed chefs in U.S. households select recipes from dog-eared cookbooks, online print-outs or torn pages of expensive one-time-purchase cooking magazines. The act of searching for a menu that utilizes available ingredients on hand, while pleasing family members’ diverse palates, often results in extreme anxiety and a tendency to reach with exasperation for a thick-crust pizza in the freezer. The objective of this article was to assess the enjoyment of a multi-course dinner created entirely by a computer program and served on 9 November 2013 at the Museum Für Kommunikation located in Berlin. Six diners were present at a lavishly decorated table. Overall taste bud sensitivity and specificity varied according to each participant. Imaginations ran wild. Cookbook shelf space in future kitchens will likely be replaced by a computer screen with easy access millions of recipes heretofore known only to an algorithm.
Recipe planning is a common complaint among adult men and women. On 9 November 2013, two retired French biologists agreed that the evening’s menu, created by the computational creativity research program at the IBM Thomas J. Watson Research Center in New York City, was quite intricately planned, indeed. After choosing the 2011 Grauburgunder white from Hermann Dönnhoff Winery in Nache, Germany, the two women pored over the menu. They wondered aloud at the audaciousness of combining disparate food types in one dish. One potential explanation of recipe fatigue could be due to the current necessity of recycling specific foods in repetitive ways. An opportunity exists for computers to create a surprise in every meal and remove guesswork from the cooking equation. This article gathers information on diners’ reactions to a meal planned by artificial intelligence and offers some suggestions for the future.
Methods
Sampling
The study sample consisted of six diners, including the lead researcher (5 females, 1 male). Participants were informally interviewed about their reaction to the dinner. All study participants were affiliated with the Falling Walls conference that took place November 8 to 9 last year in Berlin. At the 1 hour time point, a seventh participant joined the table. Nobel Laureate and French biologist Jules A. Hoffman sat down with the group, which included his wife, and answered a few quick questions related to his talk earlier in the day. Due to his necessarily late arrival, his dinner responses were not included in this summary. All other participants seemed to be well-versed in descriptive words and possessing base knowledge of what constitutes quality food. They freely shared their thoughts and feelings about the meal. The computer program sampling consisted of nearly 40,000 recipes from other large databases, including Recipes Wiki and the Institute of Culinary Education. For its reactions, to be discussed in more detail below, the computer relied on a natural language algorithms.
Measures
Diagnostic Interview The HSI-I (highly skeptical interview version 1) was developed to probe participants into sharing honest thoughts about the menu and food items. The HSI-I is a loosely structured, reporter-administered instrument, based on what is commonly accepted as good food. The HSI-I begins with an open-ended question whereby the participant is asked to qualitatively describe his or her initial reaction to the menu. Follow-up questions focused on the tastiness of the actual food. Hard-working wait staff delivered multiple plates simultaneously to an estimated 100 total tables in a central space at the Museum for Communication that could have been confused with a set from Brazil (1985).
Surprise, Pleasantness, Chemical Pairing The mathematical algorithms for the measurement of “surprise” are rooted in Bayesian surprise theory, which essentially measures the difference between a person’s prior belief and the changed belief. The units are called “wows.” With the right amount of new combinations, wows values could go to infinity. The “pleasantness” factor is based on chemistry and psychology and measures what a person likes and dislikes, using the Hedonic scale. “Chemical pairings” are established norms in culinary science and differ across cultures. In most Western cuisine, the more chemicals food share, the more likely they will be tasty when combined. Asian cuisine follows a different philosophy: opposites taste better together. As an example, rice and soy sauce do not share any flavor compounds.
Results
Humans
A total of three in-depth interviews were conducted as part of the study. Among these participants, 3 women participated, including the lead author. The attrition rate was 50 percent. Participants were lost due to a noise barrier created by distance from one end of the table to the other and the cacophony of conversations in the echoing museum. The single male participant spoke rarely and only in French. The prevalence of fascination among the women was 100 percent. Interviewing that covered the topics of an artificial intelligence-created menu and food accounted for about 30 percent of total time. This was followed by 40 percent discussion of personal lives, languages, career accomplishments and aspirations. The remaining 30 percent of discussion occurred in French, and, as such, the English-speaking lead author had to wait patiently and gaze out over the crowd, and wish she had learned French in college. A pattern emerged during the interviewing process. Participants expressed extreme skepticism about the taste of a food item before it arrived at the table. However, skepticism dissipated after the third or fourth bite of food. During discussions of the menu, the topic that proved most popular consisted of the measures of “surprise, pleasantness and chemical pairings.” Participants noted that the values of each factor had been calculated, but no scale was provided. One participant expressed frustration at the lack of context. Another attempted to create a scale based on the values printed in the menu, which are included below.
Computer
Before the dinner event on 9 November, a team of researchers at the IBM Thomas J. Watson Research Center in New York City ran algorithms on a computer to determine a menu. They assigned the following parameters: no more than 13 ingredients per dish and focus on Spanish, American, Asian and German regional cuisine. After inputting this data, the computer produced a menu in 2 seconds. More human work was required to choose which of the recipes to actually use. The complete menu and values (Figure 1) were printed for each participant:
Finger Food
Grilled tomato on a saffron crouton // Surprise: 0.041 Pleasantness: 0.423 Chemical pairing: 2
Veal roulade with kefalotyri cheese and a nutmeg remoulade // Surprise: 0.022 Pleasantness: 0.391 Chemical pairing: 1
Shot of trout and sea bass with a beer-and-buttermilk foam // Surprise: 0.194 Pleasantness: 0.401 Chemical pairing: 22
Potato burger with beef and cottage cheese // Surprise: 0.0248 Pleasantness: 0.394 Chemical pairing: 13
Mini potato with a mushroom salad and cottage cheese // Surprise: 0.143 Pleasantness: 0.385 Chemical pairing: 22
Main Menu
1st course - Pumpkin muffin with date sauce, wild herb salad and capocollo // Surprise: 0.065 Pleasantness: 0.428 Chemical pairing: 5
2nd course - Rice wrap with cinnamon-smoked salmon, blackberries, celery and a Tabasco glaze // Surprise: 0.188 Pleasantness: 0.411 Chemical pairing: 18
Or
Fennel and saffron risotto with candied ginger and a Martini granitée (vegetarian) // Surprise: 0.191 Pleasantness: 0.402 Chemical pairing: 12
3rd course - Choice of desserts
Chocolate and banana cake with a saffron-apple ragout, yogurt and honey ice cream and cherry caviar // Surprise: 0.021 Pleasantness: 0.409 Chemical pairing: 20
Lime pie with mango and honey ice cream and ginger/pepper salsa // Surprise: 0.028 Pleasantness: 0.440 Chemical pairing: 16
Quark crème caramel with cranberry and caraway ice cream // Surprise: 0.429 Pleasantness: 0.404 Chemical pairing: 17
Fig and yogurt terrine on a Merlot zabaglione // Surprise: 2.250 Pleasantness: 0.399 Chemical pairing: 3
(Figure 1)
Discussion
These findings show that the fig and yogurt terrine on Merlot zabaglione, a dessert, registered highest on the “surprise” scale with 2.250. No other food item even comes close to this value. All other values, with the exception of the chemical pairings, did not vary tremendously. The chemical pairings ranged from 22 to 1. Your lead researcher, who is mostly vegetarian, found it interesting that the vegetarian dinner “pleasantness” score was 0.009 points lower than the meat option. However, this difference does not appear to be statistically significant and therefore could be due to chance. While the age ranges in the group are unknown, and cultural backgrounds differed, all of the participants expressed concern about a computer replacing the fulfilling and familial bonding opportunity known as “the home-cooked meal.” Such fears could be unfounded and more research in this area is needed. This was the very first public event in which IBM’s menu planning computer program was used for the entire meal. The strengths of this study include the observational perspective. The lead author was on-site during the event and was able to ask follow-up questions with the participants. Additional information was gathered, including body language and number of dinner rolls consumed, but these values are not included in the current analysis. There are limitations to this study. Without a translator, it could be that some important insights about the dinner were shared, but not recorded. Although the HSI-I diagnostic interview is a tried-and-true method for gathering information, it has not been validated. As such, all responses are opinion and therefore not generalizable beyond the limits of this study. Future studies should probably take place in a lab setting and far away from the distracting interiors of a beautiful museum and techno-dance music.
Conclusions
A computer algorithm that plans meals could be a helpful addition to future kitchens by lessening the anxiety and frustration that comes with meal planning. Based on the dinner event included in this study, the algorithm produces fascinating and tasty meals. However, the author is concerned that a computer software program, while harnessing more creative power than is humanly possible, could paradoxically remove the creativity and spontaneity that comes with cooking dishes and adding salt and pepper as needed. One could describe the current human-based recipe-finding process as “hunting” and “gathering.” When primed in this mindset, a human might also seek simple, affordable, whole ingredients to use in cooking. On the other hand, a person who touches a single button may also seek to use a combination of processed foods that do not require fastidious chopping, slicing, sautéing, braising, mixing, molding, kneading, dicing and fluting, to reach the endpoint. But the lead author is also rather biased: she views cooking as a time to sip wine, talk with family, think through the recipe process, and be completely away from computers and smart phones. Like cookbooks, she may soon be a relic of earlier times.
FINANCIAL DISCLOSURE: The author, K.R., discloses that she attended the 2013 Falling Walls conference on a science writing fellowship in the company of excellent journalists. Her airfare, lodging and event food were paid for by Falling Walls.
ACKNOWLEDGEMENTS: Aatish Bhatia over at Wired Blogsalso wrote on this subject. I can recommend eating up his post, too.
PHOTO CREDIT: Falling Walls