Chapter 1 endnote 48, from Lisa Feldman Barrett.
Some context is:
My collaborators and I applied pattern classification to our meta-analysis of brain-imaging studies of emotion.
The experimental method
- We started with a database of 148 brain imaging studies from various researchers, involving over 2000 participants. In each of these studies, test subjects were asked to perform some task intended to invoke happiness, sadness, anger, fear or disgust, such as looking at photos, remembering something, or watching a film. Each study computed a kind of summary of its participants' brain activity. This summary is known as a map. So we had 148 maps.
- We then trained a computer program, the classifier, to distinguish each category of emotion from these maps, in the following way:
- Isolate one map from the database. Call it the "left-out" map.
- Train the classifier on the remaining 147 maps, using a statistical analysis.
- Use the classifier to try to classify the left-out map, and see if it's classified correctly (with respect to the intent of its associated study). For example, if the map's study was about cultivating anger, hopefully our classifier will classify the map as anger.
- Repeat step 2 for each of the 148 maps in the database.
- Repeat step 3 thousands of times.
In the end, we had a highly trained classifier that had produced a statistical pattern for each emotion category. We correctly classified the maps into one of five emotion categories — fear, anger, disgust, sadness, or happiness — with 66% accuracy (43-86% across categories). So far so good.
Now, the next step is where the problem lies in many pattern classification studies. Some scientists look at the statistical pattern for one emotion category, such as happiness, and say, “Eureka, we have found a neural signature: the elusive biological essence of happiness!" Various studies have made these claims as if their pattern classification study had discovered a "signature," "fingerprint," or "biomarker" for an emotion category.
Well, no, they haven’t.
A "happiness" pattern produced by a classifier is a statistical summary of the huge variety found in everyone's happy maps. It is not the brain state for happiness, nor the neural fingerprint for happiness. It's more like a sophisticated mathematical average.
When you compute the “distinct statistical picture of brain activity” for happiness, any given map is not necessarily similar to that picture. The two might have a lot in common, or a little bit in common, or even nothing at all! All we know is that a map which is diagnosed as coming from a study of "happiness" is mathematically closer to the “happy” classifier than to any other classifier that was built from emotion maps in the database. In fact, each study's “happy map” could be entirely distinct from everyone else’s. This is true for every emotion (and every mental category) ever studied with pattern classification.
A helpful analogy
By analogy, let’s think about animals rather than emotions. Suppose you had 148,000 photos of animals, you tabulated all their physical features, and you created a classifier to distinguish between dogs, cats, birds, guinea pigs and fish. You'd end up with five distinct, statistical pictures representing a dog, a cat, a bird, a guinea pig and a fish. The classifier for a dog is in no way an "essence" of a dog — Darwin would be horrified — but a statistical summary of the vast variety of dogs out there and how they differ, on average, from cats, birds, guinea pigs and fish. It's Darwin's idea of population thinking: a species is a population of individuals who vary from one another and who have no essence, but you can summarize a species in statistical terms. Classification allows you to predict (or diagnose) the category of an animal, not compute the animal’s essence. Likewise for emotions: for a given map, a classifier tells you which groups of maps it is most similar to (relative to some other groups), but the sum total of all that computation is neither the brain state for an emotion nor an essence of emotion.
Notes on the Notes
- Wager, Tor D., Jian Kang, Timothy D. Johnson, Thomas E. Nichols, Ajay B. Satpute, and Lisa Feldman Barrett. 2015. "A Bayesian model of category-specific emotional brain responses." PLOS Computational Biology 11 (4): e1004066.
- Kragel, Philip A., and Kevin S. Labar. 2015. "Multivariate neural biomarkers of emotional states are categorically distinct." Social Cognitive and Affective Neuroscience 10 (11): 1437-1448.
- Kragel, Philip A., and Kevin S. Labar. 2016. "Decoding the Nature of Emotion in the Brain." Trends in Cognitive Sciences 20 (6): 444-455.
- Saarimäki, Heini, Athanasios Gotsopoulos, Iiro P. Jääskeläinen, Jouko Lampinen, Patrik Vuilleumier, Riitta Hari, Mikko Sams, and Lauri Nummenmaa. 2016. "Discrete Neural Signatures of Basic Emotions." Cerebral Cortex 26 (6), 2563-2573.
- Kassam, Karim S., Amanda R. Markey, Vladimir L. Cherkassky, George Loewenstein, and Marcel Adam Just. 2013. "Identifying emotions on the basis of neural activation." PloS One 8 (6): e66032.
- Clark-Polner, Elizabeth, Timothy D. Johnson, and Lisa Feldman Barrett. 2016. "Multivoxel pattern analysis does not provide evidence to support the existence of basic emotions." Cerebral Cortex doi: 10.1093/cercor/bhw028.
- Mayr, Ernst. 2004. What Makes Biology Unique?: Considerations on the Autonomy of a Scientific Discipline. Cambridge University Press.
- Kragel, Philip A., and Kevin S. LaBar. 2013. "Multivariate pattern classification reveals autonomic and experiential representations of discrete emotions." Emotion 13 (4): 681-690.
- Park, Byoung-Jun, Eun-Hye Jang, Myoung-Ae Chung, and Sang-Hyeob Kim. 2013. "Design of prototype-based emotion recognizer using physiological signals." ETRI Journal 35 (5): 869-879.
- Yuen, Kenneth SL, Stephen J. Johnston, Federico De Martino, Bettina Sorger, Elia Formisano, David EJ Linden, and Rainer Goebel. 2012. "Pattern classification predicts individuals’ responses to affective stimuli." Translational Neuroscience 3 (3): 278-287.