Final Report Summary - COIN (Compositionality in infants)
Summary description of the project objectives
The COIN project aims at investigating compositionality in infants, using the concepts SAME and DIFFERENT as a case study. Precisely, we aim at understanding (1) whether infants have abstract representations of the relational concepts SAME and DIFFERENT prior to their understanding of the corresponding words same and different, (2) whether DIFFERENT is represented as the negation of the concept SAME (i.e. a case of compositionality: SAME = NOT DIFFERENT), and (3) whether language plays a role in the emergence of compositionlity. As part of the training, and in compliance with the work plan approved for COIN, the research activities have been integrated with teaching and supervising activities, as described below.
Description of the work performed since the beginning of the project and main results
Compositionality is a possibly human-specific computational ability that gives rise to the expressive power of human languages and human thoughts (Fodor, 2002; Hauser, Chomsky & Fitch, 2002). It consists in the ability to combine known concepts together to create novel concepts. For example, by composing the concept of NOT and the concept of BROWN, we obtain the concept of NOT BROWN, the complement of the set of brown entities in the world. However, it is not clear whether compositionality is innate or, if it is not, by what mechanisms it could be acquired. In this project, we propose to study the ability of young infants to combine mental representations. Moreover, considering the importance of compositionality in the language faculty (Fodor, 1975; Hauser et al., 2002), we ask whether the acquisition of linguistic symbols is a precondition for the emergence of compositionality.
The best case studies of compositionality involve logical operators, for these most likely involve
abstract symbols. There has been no study of logical operators in prelinguistic infants. Here, we begin to explore whether prelinguistic infants have abstract symbols that express the concepts SAME and DIFFERENT and whether they represent DIFFERENT as NOT SAME; i.e. as the combination of the representations of NOT and SAME.
SAME can be studied in experiments where participants need to recognize identity relations between stimuli to generalize structures defined by the repetition of elements (e.g. AAB, ABA). DIFFERENT, in contrast, is crucial to generalize structures like ABC or AB where all elements differ from each other.
Our initial experimental work built on previous results suggesting that infants could learn that the SAME relation between to elements (e.g. syllables, vowels or visual shapes) is predictive of an event (e.g. that a puppet wi appear in a given location on a computer screen). For instance, Hochmann et al. (2011) showed that infants could learn that the SAME relation between vowels predicts that a puppet will appear in one location on the screen, but they were unabe to learn at the same time that the DIFFERENT relation between vowels predicts the puppet’s appearance in another location. This failure could be due either to an unability to learn two predictions in parallel, or to an unability to represent the relation DIFFERENT.
We thus conducted a series of experiment to understand at what age infants can learn two predictions in parallel. Precisely, we taught infants that pictures of dogs predict the appearance of a puppet on the right of the screen, and pictures of cars predict the appearance on the left of the screen. We conducted this work with groups of 14-, 18- and 24-months. Cars and dogs were chosen because those are categories that infants do know and represent well at least by 12-months. Furthermore, even younger infants should be able to distinguish the two categories as objects vs. agents. Our prediction was that, if infants can learn only one prediction, they should learn the prediction made by dog pictures, as agents usually attracts more attention than inanimate objects. This is precisely what we observed with 14- and 18-month-olds. 24-month-olds also tended to learn only one prediction, though the general preference for dogs disappeared and some infants learned the prediction made by cars. Importantly, very few infants learned the predictions made by both cars and dogs.
These results suggest that infants’ failure to learn a rule such as “DIFFERENT predicts left” in parallel to learning “SAME predicts right” may be due to a general inability to learn two predictions in parallel. Therefore, we cannot conclude about infants’ ability or inability to represent the relation DIFFERENT.
We therefore developped a novel paradigm to directly study infants representation of SAME and DIFFERENT in a paradigm that requires learning only one predictive rule. We adapted the Match-to-Sample (MTS) and Mismatch-to-Sample (mMTS) classically usedin the animal literature to study the concepts SAME and DIFFERENT (e.g. Giurfa et al., 2001). In both paradigms, participants are presented with a sample (e.g. A) and two alternatives (A or B). The correct alternative is the one that is the same as the sample in MTS, and the one that is different from the sample in mMTS. Success in MTS is traditionally thought to indicate a representation of the concept SAME, and success in mMTS is traditionally thought to indicate a repesentaiton of the concept DIFFERENT. We developped an anticipatory looking version of these tasks to test infants with an eye-tracker. Results suggest that 14-month-olds can learn to scucceed both in the MTS and in the mMTS.
Furthermore, we observed that infants could succeed in mMTS in two different ways. They could either learn to choose the alternative that is different from the sample (thus representing a rule based on the concept DIFFERENT), or learn to avoid the alternative that is the same as the sample (thus representing a rule based on the concept SAME).
We thus ran a modified version of the task, to invetigate which of these two rules infants actually learned. We showed that infants in mMTS learned to avoid the alternative that is the same as the sample, and infants in MTS learned to choose that alternative. In sum, we found convincing evidence that 14-month-old infants represent the abstract relational concept SAME, but we found no evidence that represent (or use) the concepts DIFFERENT (Hochmann, Mody & Carey, under review).
In parallel, we ran studies with children to investigate the development of the understanding of the words same and different, and the relation between the acquisition of these words and the manipulation of the corresponding concepts in a non-verbal task. Results suggest that children understand the words same and different betwee 3- and 4-years, but cannot flexibly use the concepts SAME and DIFFERENT before they turn 5. Interestingly, we also found that the relative simplicity of SAME with respect to DIFFERENT has disappeared when children learned the associated words: either chidlren understand both words, or they do not understand either.
Expected final results and their potential impact and use
Many authors have noted that a unique characteristic of the human mind is its productivity. No other animal can invent novel concepts as we routinely do. Compositionality is the ability at the core of human productivity. By understanding better how it develops in human infants, we will better understand human nature. Using the concepts SAME and DIFFERENT as a case study, we repeatedly found no evidence for compositionality in infants before they acquire the relevant words. We are pursuing our work to better understand how infants’ conceptual repesentations change when words are acquired, and how compositionality emerges.
The COIN project aims at investigating compositionality in infants, using the concepts SAME and DIFFERENT as a case study. Precisely, we aim at understanding (1) whether infants have abstract representations of the relational concepts SAME and DIFFERENT prior to their understanding of the corresponding words same and different, (2) whether DIFFERENT is represented as the negation of the concept SAME (i.e. a case of compositionality: SAME = NOT DIFFERENT), and (3) whether language plays a role in the emergence of compositionlity. As part of the training, and in compliance with the work plan approved for COIN, the research activities have been integrated with teaching and supervising activities, as described below.
Description of the work performed since the beginning of the project and main results
Compositionality is a possibly human-specific computational ability that gives rise to the expressive power of human languages and human thoughts (Fodor, 2002; Hauser, Chomsky & Fitch, 2002). It consists in the ability to combine known concepts together to create novel concepts. For example, by composing the concept of NOT and the concept of BROWN, we obtain the concept of NOT BROWN, the complement of the set of brown entities in the world. However, it is not clear whether compositionality is innate or, if it is not, by what mechanisms it could be acquired. In this project, we propose to study the ability of young infants to combine mental representations. Moreover, considering the importance of compositionality in the language faculty (Fodor, 1975; Hauser et al., 2002), we ask whether the acquisition of linguistic symbols is a precondition for the emergence of compositionality.
The best case studies of compositionality involve logical operators, for these most likely involve
abstract symbols. There has been no study of logical operators in prelinguistic infants. Here, we begin to explore whether prelinguistic infants have abstract symbols that express the concepts SAME and DIFFERENT and whether they represent DIFFERENT as NOT SAME; i.e. as the combination of the representations of NOT and SAME.
SAME can be studied in experiments where participants need to recognize identity relations between stimuli to generalize structures defined by the repetition of elements (e.g. AAB, ABA). DIFFERENT, in contrast, is crucial to generalize structures like ABC or AB where all elements differ from each other.
Our initial experimental work built on previous results suggesting that infants could learn that the SAME relation between to elements (e.g. syllables, vowels or visual shapes) is predictive of an event (e.g. that a puppet wi appear in a given location on a computer screen). For instance, Hochmann et al. (2011) showed that infants could learn that the SAME relation between vowels predicts that a puppet will appear in one location on the screen, but they were unabe to learn at the same time that the DIFFERENT relation between vowels predicts the puppet’s appearance in another location. This failure could be due either to an unability to learn two predictions in parallel, or to an unability to represent the relation DIFFERENT.
We thus conducted a series of experiment to understand at what age infants can learn two predictions in parallel. Precisely, we taught infants that pictures of dogs predict the appearance of a puppet on the right of the screen, and pictures of cars predict the appearance on the left of the screen. We conducted this work with groups of 14-, 18- and 24-months. Cars and dogs were chosen because those are categories that infants do know and represent well at least by 12-months. Furthermore, even younger infants should be able to distinguish the two categories as objects vs. agents. Our prediction was that, if infants can learn only one prediction, they should learn the prediction made by dog pictures, as agents usually attracts more attention than inanimate objects. This is precisely what we observed with 14- and 18-month-olds. 24-month-olds also tended to learn only one prediction, though the general preference for dogs disappeared and some infants learned the prediction made by cars. Importantly, very few infants learned the predictions made by both cars and dogs.
These results suggest that infants’ failure to learn a rule such as “DIFFERENT predicts left” in parallel to learning “SAME predicts right” may be due to a general inability to learn two predictions in parallel. Therefore, we cannot conclude about infants’ ability or inability to represent the relation DIFFERENT.
We therefore developped a novel paradigm to directly study infants representation of SAME and DIFFERENT in a paradigm that requires learning only one predictive rule. We adapted the Match-to-Sample (MTS) and Mismatch-to-Sample (mMTS) classically usedin the animal literature to study the concepts SAME and DIFFERENT (e.g. Giurfa et al., 2001). In both paradigms, participants are presented with a sample (e.g. A) and two alternatives (A or B). The correct alternative is the one that is the same as the sample in MTS, and the one that is different from the sample in mMTS. Success in MTS is traditionally thought to indicate a representation of the concept SAME, and success in mMTS is traditionally thought to indicate a repesentaiton of the concept DIFFERENT. We developped an anticipatory looking version of these tasks to test infants with an eye-tracker. Results suggest that 14-month-olds can learn to scucceed both in the MTS and in the mMTS.
Furthermore, we observed that infants could succeed in mMTS in two different ways. They could either learn to choose the alternative that is different from the sample (thus representing a rule based on the concept DIFFERENT), or learn to avoid the alternative that is the same as the sample (thus representing a rule based on the concept SAME).
We thus ran a modified version of the task, to invetigate which of these two rules infants actually learned. We showed that infants in mMTS learned to avoid the alternative that is the same as the sample, and infants in MTS learned to choose that alternative. In sum, we found convincing evidence that 14-month-old infants represent the abstract relational concept SAME, but we found no evidence that represent (or use) the concepts DIFFERENT (Hochmann, Mody & Carey, under review).
In parallel, we ran studies with children to investigate the development of the understanding of the words same and different, and the relation between the acquisition of these words and the manipulation of the corresponding concepts in a non-verbal task. Results suggest that children understand the words same and different betwee 3- and 4-years, but cannot flexibly use the concepts SAME and DIFFERENT before they turn 5. Interestingly, we also found that the relative simplicity of SAME with respect to DIFFERENT has disappeared when children learned the associated words: either chidlren understand both words, or they do not understand either.
Expected final results and their potential impact and use
Many authors have noted that a unique characteristic of the human mind is its productivity. No other animal can invent novel concepts as we routinely do. Compositionality is the ability at the core of human productivity. By understanding better how it develops in human infants, we will better understand human nature. Using the concepts SAME and DIFFERENT as a case study, we repeatedly found no evidence for compositionality in infants before they acquire the relevant words. We are pursuing our work to better understand how infants’ conceptual repesentations change when words are acquired, and how compositionality emerges.