So long and thanks for the PhD!
Friends, colleagues, students, mentors - I'm off to a new start!
In the last couple of years, the Hebrew University, in particular - the cognitive sciences department, ELSC, the CS school, and the center for the study of rationality, have been a second home to me. Through joys and celebrations, successes and failures, alongside the talented colleagues I had the honor of working with and humbly call my friends, my PhD was a profoundly shaping experience. That is to say, my PhD was really nice.
In fact, it was so nice that it became obvious that I just had to get (at least a little) more. In the foreseeable future, I'll be hanging around Boston (MA) and New-Haven (CT), keeping up the fun with some exciting new research. If you are in the neighborhood, come say hi :)
Do deep neural networks generate visual categories like humans? The case of flower classification
Image - Gilboa Iris (Iris haynei) passed through several deep dream generators.
I recently completed a course given by Prof. Yair Weiss ("Lab in computational neuroscience and learning") . In this course, I developed several implementations for classification algorithms of flower images using deep neural networks (DNNs). In the first phase of the project, I present an implementation that successfully classifies a novel dataset with 90% accuracy. In the second phase, I note that both humans and DNNs are excellent classifiers. Then, I use the trained classification networks of the first phase to ask whether the success humans and machines share in terms of function (both are great classifiers) also implies that the two also rely on similar mechanisms for such classification. The short answer - using the limited dataset and methods I checked - is no.
Read the full report here.
The great Tahini Experiment
In previous semester I had the honor, and pleasure, of supervising a very talented group of students. The venue, is a hands-on course in research where the students should bring into action all the theoretical knowledge they accumulated from previous years - from theory-driven hypotheses to the statistical definition of H0 distributions across their data. As for a research domain, the guidance was to form a research question that could be hosted under the flexible roof of the cognitive sciences while stressing external validity (we called it "cognition in the wild"). The students, very talented but not without a sense of humor, decided to study a question which is apparently at the heart of the Jerusalem-student-pop-culture: which Tahini brand is truly the best? or rather - how our preferences, specifically preferences of low-level sensory stimuli, are shaped? 150 surveyed subjects and nights of statistical processing later, we now know the answer.
-- research by: Shifra Avital, Yotam Beit-Yannai, Shir Hoter, Matan Lotem, Noa Schneider and Aviya Ben-Melech Chovev
As part of our B.A studies, we were granted the opportunity to examine an interesting cognitive question of our choice. After a short deliberation we all agreed that there was only one thing that we all desperately wanted to know - what tahini is the best? And you must understand - tahini is fundamental.
As students, tahini makes the great majority of our menu. Tahini can save a bad sandwich, it can upgrade significantly any dinner - be it vegetables cooked in an oven, schnitzel, falafel, salad or pretty much any meal. Tahini is accessible, easy to prepare and constantly features in students' dinner parties. And there's also a great expertise associated with tahini making.
A good tahini maker, a tahini-barista if you will, knows just the right temperature, how much water to add, the exact color of the lemon from which a juice squeezed into the raw tahini, and the careful mixture of spices and herbs that should be added top optimized the whole tahini experience.
Naturally, many different tahini brands are present in the market. This great variety of brands is the subject of an endless debate between students regarding the identity of the absolute best tahini; and like the best beer, soda or car-brand, these debates never lead to an ultimate consensus. While some brands are considered great or just okay, the elusive question of which is the best tahini, forever remained undecided. But behold, this is where we step in.
When pondering this question ourselves, we stumbled upon an interesting phenomenon – although most of us hadn’t tasted all brands, we all had solid opinions about their taste and quality. This fact made us think - how can we have such a clear opinion about different types of tahini without even tasting them?... Horror crept into our hearts when we thought that maybe our preference for tahini is not only a function of its taste. Perhaps our preferences are influenced by other irrelevant factors. Our hypothesis is that the brand of the tahini is such a factor. For example, we assume that store brands are perceived as less tasty, regardless of their taste.
When dealing with a major question such as 'Which tahini is the best?', It is important to ascertain the factors that influence our answer
Therefore, we wanted to test and see - Does the brand of the tahini have an impact on people's preference? Could it be that the best tahini is actually a simple underrated store brand? Or is the student’s perception of branding compatible to the taste of the tahini?
We planned an experiment to test and determine- which tahini is the best? What impact does the branding have on people’s preference?
Will "Har-Bracha" prevail and justify its status as one of the best or fail to deliver? Could this be the break "Barake" needed, or the proof that it doesn't belong in the best tahini discussion? Will "Shufersal" the underdog surprise us or fall as predicted to the bottom of the ranking? May the best tahini win!
We collected data from 54 students by distributing questionnaires in which subjects rated the three tahinis without tasting them, in order to verify our preliminary opinions regarding which tahini is considered the best, using the following questionnaire:
Afterwards we prepared the tasting trails. We set up tahini stands next to campus’s library, fishing for passersby and making them taste and rank the three tahini. Importantly, we repeated the experiment two independent times (same venue, different dates) with one major distinction: While at the first experiment the brands of tahini were clearly visible, in the second experiment they were not.
During the trial run, we had to deal with challenging subjects such as-
Luckily the starving students jumped at the opportunity to participate in the once in a lifetime experiment, not very surprising given the love of students for free food and for tahini in particular. Anyway, in spite of dismal predictions of having to collect data over months upon months, we got all the data we needed in 15 minutes - 57 in the revealed stand and 65 in the blind.
At first glance, the results do not show huge differences between the taste ratings in the two different conditions (blind vs. revealed brand).
Indeed, when we performed a statistical test (t-test for independent samples) between the ranking of each tahini in the blind test and in the revealed test, we found no significant difference (n revealed =57, n blind =65, df=120, p>0.05). One can also see in the second graph that in both groups, "Har-Bracha" was ranked first as the best tahini (second graph, green bars).
We did, however, notice that the rating of the "Shufersal" tahini rose significantly when the subjects tasted the tahini compared to the dry questionnaire (first graph, right bars). In addition, when comparing the tahini of "Shufersal" with the tahini of "Barake", it became apparent that the average rating of "Barake" is higher than that of "Shufersal" both in the questionnaire and in revealed test, where the brands were exposed, while on the test when the brands were hidden, the ranking was reversed, but the difference was not significant (second graph blue and yellow bars). Lastly, we performed a statistical test (Chi-square for Independence) to see if the subjects' preferences were significantly altered between the revealed test and the blind test, and voilà! we found a marginally significant difference between the ratings in the two conditions (n revealed =57, n blind =65, degrees of freedom =8, p=0.054).
Overall our results show that there are small differences between subjects’ preferences when they see the brand and when the brand is hidden. We saw that "Har-bracha" is ranked in the highest place in both groups, so it seems to be really tastier.
On the other hand, the rating of the "Shufersal" was different in the different tests, which suggests that the cause for its low ranking in the open questionnaire is the branding rather than its taste.
Given the proven effect of branding on preference and the solid opinion of students regarding the ranking of tahini even without tasting, the results are quite surprising. We expected that the brand would affect the preferences so that the tahini ratings would be very different in the open test compared with the blind test. In practice, we did not find any significant differences between the two groups.
We offer a few possible explanations for the lack of significant differentiation between the two groups. The first, straightforward, explanation is that indeed the stereotypes hold true- “Har-Bracha” is the better tahini and the participants were consistent in their ratings accordingly.
However, in science it is important to consider alternative explanations. The first alternative explanation is that students from Givat Ram (where we performed the study, and where we study– best campus ever) are simply superior to the general population. Along with their advantage in intelligence, creativity, social skills, internal and external beauty, tennis and other gymnastics skills, they are also resistant to the effects of branding, and have a most accurate, sense of taste. A second alternative explanation is that the effect does indeed exist in the general population however, we inadvertently surveyed an expert-population in which this effect is not manifested. It seems that the opinion of the surveyed students about the ranking of tahini is derived from a deep and meaningful acquaintance with tahini. They know from experience which tahini is tastier and therefore there is no significant difference in taste ratings when exposed and not exposed to brands.
In our opinion, it would be interesting to repeat the experiment. Specifically, to test audiences that are less familiar with tahini and see if their taste ratings will be affected by branding.
In conclusion, this research provides strong evidence that lays a scientific rational to an already heavily reinforced behavior, that could have been otherwise suspected as irrational – we will keep on paying more for our favorite, branded tahini as it is now scientifically proven, that it just tastes the best.