Quines regions del cervell del gos permeten la comprensió de paraules humanes?

Quines regions del cervell del gos permeten la comprensió de paraules humanes?

Un grup d’investigadors d’Estats Units d’Amèrica ha realitzat un estudi per determinar les regions del cervell que intervenen en la comprensió de paraules humanes.

Font de l’enllaç: https://www.frontiersin.org/articles/10.3389/fnins.2018.00737/full

Us passem el resum i conclusions de l’estudi (en anglès):

Resum

How do dogs understand human words? At a basic level, understanding would require the discrimination of words from non-words. To determine the mechanisms of such a discrimination, we trained 12 dogs to retrieve two objects based on object names, then probed the neural basis for these auditory discriminations using awake-fMRI. We compared the neural response to these trained words relative to “oddball” pseudowords the dogs had not heard before. Consistent with novelty detection, we found greater activation for pseudowords relative to trained words bilaterally in the parietotemporal cortex. To probe the neural basis for representations of trained words, searchlight multivoxel pattern analysis (MVPA) revealed that a subset of dogs had clusters of informative voxels that discriminated between the two trained words. These clusters included the left temporal cortex and amygdala, left caudate nucleus, and thalamus. These results demonstrate that dogs’ processing of human words utilizes basic processes like novelty detection, and for some dogs, may also include auditory and hedonic representations.

Discussion

Using awake-fMRI in dogs, we found neural evidence for auditory novelty detection in the domain of human speech. The hallmark of this finding was greater activation in parietotemporal cortex to novel pseudowords relative to trained words. Thus, even in the absence of a behavioral response, we demonstrate that dogs process human speech at least to the extent of differentiating words they have heard before from those they have not. The mechanism of such novelty detection may be rooted in either the relatively less frequent presentation of the pseudowords (oddball detection) or the lack of meaning associated with them (lexical processing).

The activation observed in the parietotemporal cortex to pseudowords relative to trained words meets current standards of human fMRI analyses concerning up-to-date methods for cluster thresholds. Specifically, to address concerns raised by Eklund et al. (2016), present analyses for cluster inferences address the former Gaussian-shaped assumption about spatial structure in the residuals of fMRI data and provide more accurate false positive rates compared to previous methods (Eklund et al., 2016Cox et al., 2017Slotnick, 2017). As the identified cluster was significant at P ≤ 0.005, corrected for whole-brain FWE, the result does not appear to be a false positive. However, as the study was limited to 11 participants, future studies with an increased number of participants could produce a more robust finding.

In humans, real words typically result in more activation than pseudowords, evoking activity in what has been called a general semantic network (Binder et al., 2009). Although such activations are generally bilateral, they tend to lateralize more to the left and cluster around the angular gyrus, but that anatomical location is ill-defined and is sometimes called the TPJ or temporal-parietal-occipital cortex (TPO) (Graves et al., 2017). It is clear that the nature of the task and the relative frequency of words and pseudowords can influence the relative activation to words and pseudowords. For example, stronger activation to pseudowords depends on whether the pseudoword resembles a known word or is so unlike known words as to prevent any semantic retrieval. When the pseudoword is similar to a known word, more processing has been observed in the superior temporal gyri, presumably to disambiguate it from known words (Kotz, 2002Raettig and Kotz, 2008). Thus, in dogs, the greater activation to the pseudowords could be due to the acoustic similarity between pseudowords and words that the dogs “knew” and their attempt to resolve the ambiguity. This would be a form of low-level lexical processing. However, previous research has shown that dogs can discriminate between altered phonemes of well-known commands (Fukuzawa et al., 2005), suggesting that it is unlikely that the dogs in our study were confused by acoustic similarity of words and pseudowords.

More likely, a novel word resulted in increased processing to facilitate learning the association with the novel object that followed. A dog’s behavioral bias for novelty is often described as an explanation for performance otherwise labeled as learning by exclusion (Bloom, 2004Markman and Abelev, 2004Zaine et al., 2014). As such, a dog may select a novel item because it is novel among other stimuli, but not because she has learned all other stimuli and associated a new word with the novel item. A bias for novelty would therefore be reflected in the dog’s brain as with her behavior.

Auditory stimuli can be difficult to discriminate in the scanner. We used a continuous scanning protocol because that is what the dogs were accustomed to. The simple contrast of all words vs. all objects showed bilateral activation of the superior temporal lobe, indicating that the dogs heard something. However, the main effect of pseudowords vs. trained words showed that the majority of dogs discriminated well enough to tell the difference. The predominant location in the auditory pathway also suggests that the effect was not based on non-verbal cues from the handler (i.e., Clever Hans effect).

The manner in which dogs learn words is different than humans do, and this undoubtedly affects their performance on behavioral tests and the patterns of brain activation we observed. Humans acquire nouns as early as 6 months of age and differentiate between nouns prior to their ability to use verbs (Bergelson and Swingley, 2012Waxman et al., 2013). In contrast, dogs do not typically have much experience with nouns because humans tend to train them on actions/verbs (e.g., sit and fetch). Consequently, even the trained words in our study were novel for the dogs in comparison to years of experience with verbs as commands. Prior studies have shown only three dogs that consistently retrieved objects given a verbal referent (Kaminski et al., 2004Pilley and Reid, 2011). Additionally, those dogs had been trained to retrieve from a young age (<11 months), and in most cases rarely attained 100 percent accuracy. Object retrieval training for the current experiment was modeled from these studies; however, because the dogs’ owners conducted training at home on a voluntary basis, training rigor could not be enforced.

Although humans readily generalize the meaning of words to a variety of contexts, this may not be the case for dogs. The environment in which the dogs learned the words was different than both the testing and scanning environments (Mills, 2015). In addition, although human fMRI language studies do not typically repeat the spoken word each trial, as is common in oddball paradigms, it was necessary for the dogs to make sure that they heard each word. Trials also did not include a condition in which a spoken pseudoword was followed by a trained object, or trials in which a trained object was mismatched to a trained word. These types of trials would have provided additional evidence for violation of expected semantic content; however, these types of trials have the potential to confuse the dogs and result in extinction of the words already learned. Lastly, dogs might have habituated to the continued presentation of trained words followed by trained objects, as opposed to the single trial presentations of pseudowords and the accompanying novel objects.

So what do words mean to dogs? Even though our findings suggest a prominent role for novelty in dogs’ processing of human words, this leaves the question of what the words represent. One possibility is that the words had no further representation other than the relative hedonic value of the objects. While some dogs showed a behavioral preference for one object over the other, this preference was not reflected in whole brain analyses. Admittedly, the somewhat arbitrary designation of word1/word2 and object1/object2 could explain the non-significant results in the univariate analysis. Indeed, the MVPA of word1 vs. word2, which identified regions that classified the words above chance regardless of directionality, showed one cluster in the left caudate. However, the MVPA also identified clusters in the left TPJ, anterior temporal lobe and amygdala, and posterior thalamus. The TPJ was located just posterior to the region in the univariate analysis, which would take it out of the area of cortex associated with low-level acoustic processing. Its location appears similar to human angular gyrus. This region and its extension to the anterior temporal lobe look remarkably similar to the conceptual component of the general semantic network in humans (Dahl, 2004Binder et al., 2009Binder and Desai, 2011). If so, these could be potential sites for receptive word processing and even semantic representation in dogs, but future work would need to verify this.

Evaluating classifier performance for MVPA remains a complex task. We used MVPA as an exploratory analysis to identify brain regions that potentially discriminate between trained words across dogs. But classification using the whole brain may result in a high classification accuracy that is not generalizable across subjects. Indeed, the regions identified using MVPA were of marginal statistical significance, especially given the small sample size. Further, it should be noted that only a subset of dogs contained informative voxels in the TPJ region. Although all dogs had informative voxels somewhere in the brain, only seven dogs had informative voxels in the TPJ area. Thus, even though all the dogs were cleared for scanning by reaching performance criterion, they may have used different mechanisms to process the words. Like our previous fMRI studies, heterogeneity seems to be the rule (Cook et al., 2016a,b). Even so, the accuracy of the classifier was not correlated with a dog’s performance. This suggests that performance on such tasks may be influenced by factors other than word discrimination alone.

These results highlight potential mechanisms by which dogs process words. Word novelty appears to play an important role. The strong response of the parietotemporal region to pseudowords suggests that dogs have some basic ability to differentiate words with associations from those that do not. Future studies may reveal whether these representations remain in the auditory domain or whether such representations are invariant to modality.

Share

Deixa un comentari

L'adreça electrònica no es publicarà. Els camps necessaris estan marcats amb *