top of page
Search

What is an Emotion?

Writer's picture: m.t. wilson, phdm.t. wilson, phd

What is an Emotion? A Connectionist Perspective



Researchers often disagree as to whether emotions are largely consistent across people and over time, or whether they are variable. They also disagree as to whether emotions are initiated by appraisals, or whether they may be initiated in diverse ways. We draw upon Parallel-Distributed-Processing to offer an algorithmic account in which features of an emotion instance are bi-directionally connected to each other via conjunction units. We propose that such indirect connections may be innate as well as learned. These ideas lead to the development of the Interactive Activation and Competition framework for Emotion (IAC-E) which allows us to specify when emotions are consistent and when they are variable, as well as when they are appraisal-led and when they are input-agnostic.







SUPPLEMENTARY MATERIALS

What is an Emotion? A Connectionist Perspective

Gaurav Suri and James J. Gross


Simulation #3: Facial Feedback

According to the facial-feedback theory of emotion, the facial activity associated with particular emotional expressions can influence people’s affective experiences. This simulation describes an IAC-based mechanism underlying facial feedback.

Target Experiments: In an influential study related to the facial feedback effect, Strack and colleagues (1988), asked participants to watch amusing cartoons while holding a pen either between their lips or between their teeth. Their reasoning was that holding a pen between their lips would not facilitate natural smiling whereas holding a pen would do so. This manipulation was an elegant test of the facial feedback hypothesis since participants were frequently not aware of the effect of the manipulation. They found that participants in the lip condition rated the cartoons as being less funny than participants in the teeth condition.

Subsequent work (Noah, Schul, & Mayo, 2018) found that the above experiment replicated provided participants were not monitored, observed or recorded (which was the case in the original experiment, but not in subsequently attempted replications). A recent meta-analysis (Coles, Larsen, & Lench, 2019) confirmed the facial feedback effect, but not the effect of moderation by observation.

Empirical Results: Here we report on the replication results in the no-camera condition obtained by Noah and colleagues (2018) which are similar to the original results obtained by Strack and colleagues (1988). Participants in the lips condition evaluated the cartoons to have an amusement rating of 4.92, whereas participants in the teeth condition evaluated the cartoons to have an amusement rating of 5.75 (t(162 = 2.48, p = 0.01). This experiment was not constrained by multiple conditions which made it straightforward for a simulation to capture the precise numbers obtained in the experiment via an appropriately chosen probability density function (see Assumption 7 in Table 2) that translated activation values to the empirical results. Nevertheless, the empirical data does afford an opportunity to examine how experience and facial feedback may interact with each other in the context of the IAC framework. Importantly, the simulation must explain how amusement ratings –that are an input into the network – are altered over time due to facial feedback.

Network Structure: The network for this simulation is predicated on the idea that humans have many exemplars of instances in which they were amused at a high level and were (fully) smiling, and many exemplars of instances in which they were amused at a low level and were not smiling. These relationships are depicted in Figure SM1a. An experience input pool represents the subjective feeling of being amused. It has units that stand for varying levels of amusement (greyed out), and central to this simulation has two units that we have designated to stand for ‘high’ experienced amusement and ‘low’ experienced amusement. These units maximally inhibit each other (-1 weight). A second facial input pool contains units representing smiles that range from a unit standing for the ‘no smile’ condition and a unit standing for the ‘full smile’ condition. These units maximally inhibit each other (-1 weight). The ‘high’ amusement unit is connected to the ‘full smile’ unit and the ‘low’ amusement unit is connected to the ‘no smile’ unit via two conjunction units in the hidden pool. These weights are maximally excitatory (+1).


Figure SM1: Network structure and activation dynamics for Simulation 3. Panel (a) depicts the network structure in which instances of low amusement are generally connected with (via conjunction units) instances of not smiling, whereas instances of high amusement are generally connected with (via conjunction units) instances of smiling. The left graph in Panel (b) depicts activation in the ‘teeth condition’ in which smiling is possible. Here, the cartoons are generally perceived to be highly amusing. The right graph in Panel (b) depicts activation in the ‘lip condition’ in which smiling is not possible. Here, the cartoons are initially perceived to be highly amusing, but then to be less amusing due to interactive dynamics.


Network Dynamics and Simulation Results: We assumed that the cartoons reliably produced an input (=0.1) into the ‘high’ amusement unit (corresponding to a modest salience). We modeled the ‘teeth’ condition by allowing an input (=1) in the ‘full smile’ unit, and we modeled the ‘lip’ condition by allowing an input (=1) in the ‘no smile’ unit. In both conditions, the input into the amusement was assumed to be identical (= 1 in the ‘high’ amusement unit).

In the ‘teeth’ condition, the ‘full’ smile unit and the ‘high’ experienced amusement activated the same hidden unit and therefore reinforced each other. In the ‘lips’ condition ‘no smile’ unit and the ‘high’ amusement activated different hidden units that competed with each other. Thus, even though the cartoons were initially experienced as equally funny in both conditions, the effects of feedback made the ‘high’ amusement unit to dominate the ‘low’ amusement unit in the teeth condition, but the reverse was true in the lip condition.

Significance: The proposed network suggests that the principles of interactive activation provide a viable mechanistic explanation for the facial feedback phenomenon. Notably, the simulations also suggest that the impact of facial feedback requires some time to exert itself. In particular, the crossover-point in Figure SM1b (lips condition) occurred at approximately the 60th network cycle – suggesting that facial feedback effects may not be visible in an experimental condition that requires participants to provide amusement ratings very quickly after presentation (or a condition that limits the presentation time of the cartoon). Finally, a similar network architecture suggests that facial expression of surprise and disgust should increase the intensity of those emotions. There is gathering evidence supporting this prediction (Lewis, 2012).

Simulation #4: Emotion Words Shape Emotion Percepts

Emotion perceivers are known to use emotion concepts in the formation of visual percepts of emotion. In particular, access to emotion words may assist in the encoding of associated emotional faces. In this simulation, we propose an IAC-based mechanism for this phenomenon.

Target Experiment: Gendron and colleagues (2012) asked participants to complete two tasks: in the first task, they viewed 48 faces which – either weakly or intensely – depicted anger, sadness, fear, or disgust. This initial task was to ensure that participants would be faster to respond to stimuli upon a subsequent viewing in the second task (due to repetition priming). The second task began with participants being shown an emotion word (e.g. ‘anger’) 30 times in some trials, or three times, in other trials. The repeated exposure to a word (e.g. 30 times) is known to cause semantic satiation (Black, 2004) which leads to a temporary decrease in the accessibility of the word’s meaning. Showing a word three times is thought to lead to a temporary increase in the accessibility of the meaning of the word. Participants were next presented with a single image (i.e. the prime) from the study phase for 50ms followed by a mask for 50ms. Finally, in a forced choice task, participants were asked to identify the prime they had just seen from two images – one of which was the prime, and the other was the same face depicting the same emotion at a different level of intensity (the alternate image was weakly intense if the prime was strongly intense, and vice versa).

Empirical Results: Participants were significantly slower to judge which face had been shown as the prime after the relevant emotion word was satiated (30x repetition) versus when it was primed (3x repetition). The average reaction time for the satiation trials was 1720ms, whereas for the primed trials it was 1650 ms (F(1, 59) = 8.166, p = .006). This difference in reaction time was observed, even though emotion words were not required to perform the forced choice task. This suggests that this such conceptual knowledge of emotion words is routinely active during emotion perception.


Figure SM2: Network structure and activation dynamics for Simulation 4. Panel (a) depicts the network structure so that emotion category words in the primed condition the words are connected (via conjunction units) to face units depicting high or low intensity emotion. In the suppressed condition no such connection exists. Panel (b) assumes (without loss of generality) that the high intensity face was the right response. The top graph shows that activation in the ‘face high’ feature crosses threshold at the 18th network cycle compared to the 23rd cycle in the suppressed condition in the lower graph.


Network Structure: The network for this simulation (Figure SM2a) consists of a Face feature pool and an Emotion Word Semantic feature pool. The Face feature pool has a unit representing high intensity faces and another unit representing low intensity faces. Notably, the empirical design here does not require us to distinguish units among different emotions – and the network does not do so. These two units inhibited each other (-1). The Emotion Word Semantic feature pool has two units, one for the primed condition (in which the emotion word was displayed three times to increase accessibility) and one for the suppressed condition (in which the emotion word was displayed 30 times to decrease accessibility). These two units also inhibited each other (-1). The feature units are connected to two units in the Hidden pool as shown in Figure SM1a representing different intensities of emotion. One of these was connected with the high intensity face unit and the other was connected with the low intensity face unit (excitatory connections, +1). The ‘primed’ Emotion Word Semantic unit was connected with both the hidden units – since an activated emotion word was consistent with high intensity as well as low intensity emotions.

Network Dynamics and Simulation Results: Without loss of generality we assumed that the high intensity face unit was the prime and constituted the ‘correct’ choice (i.e. a similar argument would apply if the low intensity face unit was the prime). We therefore provided the high intensity face unit with an input activation of 0.3, and the low intensity face unit with an input of 0.2. This difference reflected the fact that (in the modeled scenario) the high intensity face had been seen during the study phase, then as the prime, and finally as one of the faces displayed in the forced choice. The low intensity face was only seen as one of the faces displayed in the forced choice. In the primed condition we provided activation to the ‘primed’ Emotion Word Semantic unit, and in the suppressed condition, we provided input to the ‘suppressed’ Emotion Word Semantic unit.

We assumed that a face was recognized when the activation level in the corresponding face feature unit exceeded a certain threshold (a free parameter set at 0.5). In the primed condition activation from the ‘primed’ Emotion Word Semantic unit activated both hidden units, and this activation then further activated both feature units. Inhibition caused the high intensity unit (in the modeled scenario) to cross threshold, whereas the low intensity unit did not. In the suppressed condition, the hidden units received no activation from the Emotion Word Semantic units, and therefore the face feature units relied upon their input to cross threshold.

As shown in Figure SM2b, in the primed condition the threshold was crossed in the 18th network cycle, and in the suppressed condition, this occurred in the 23rd network cycle. Using the linear transformation (1.4N + 140, where N represents the Network Cycle), these network cycles represent the RTs observed in the experiment (i.e. 165ms in the primed condition and 172ms in the suppressed condition).

Significance: Gendron and colleagues (2012) and colleagues proposed that their empirical findings were consistent with the constructed view of emotion. We note that the above simulation shows that their findings are also consistent with the IAC-E.



Suri, G., & Gross, J. J. (2022). What is an Emotion? A Connectionist Perspective. Emotion Review. https://doi.org/10.1177/17540739221082203


13 views0 comments

Comments


  • Instagram
  • Instagram
  • Twitter
  • LinkedIn
  • LinkedIn
bottom of page