Federico Raue

• The Symbol Grounding Problem (SGP) is one of the first attempts to proposed a hypothesis about mapping abstract concepts and the real world. For example, the concept "ball" can be represented by an object with a round shape (visual modality) and phonemes /b/ /a/ /l/ (audio modality). This thesis is inspired by the association learning presented in infant development. Newborns can associate visual and audio modalities of the same concept that are presented at the same time for vocabulary acquisition task. The goal of this thesis is to develop a novel framework that combines the constraints of the Symbol Grounding Problem and Neural Networks in a simplified scenario of association learning in infants. The first motivation is that the network output can be considered as numerical symbolic features because the attributes of input samples are already embedded. The second motivation is the association between two samples is predefined before training via the same vectorial representation. This thesis proposes to associate two samples and the vectorial representation during training. Two scenarios are considered: sample pair association and sequence pair association. Three main contributions are presented in this work. The first contribution is a novel Symbolic Association Model based on two parallel MLPs. The association task is defined by learning that two instances that represent one concept. Moreover, a novel training algorithm is defined by matching the output vectors of the MLPs with a statistical distribution for obtaining the relationship between concepts and vectorial representations. The second contribution is a novel Symbolic Association Model based on two parallel LSTM networks that are trained on weakly labeled sequences. The definition of association task is extended to learn that two sequences represent the same series of concepts. This model uses a training algorithm that is similar to MLP-based approach. The last contribution is a Classless Association. The association task is defined by learning based on the relationship of two samples that represents the same unknown concept. In summary, the contributions of this thesis are to extend Artificial Intelligence and Cognitive Computation research with a new constraint that is cognitive motivated. Moreover, two training algorithms with a new constraint are proposed for two cases: single and sequence associations. Besides, a new training rule with no-labels with promising results is proposed.