In our previous works, we presented Logic-Muse as an Intelligent Tutoring System that helps learners improve logical reasoning skills in multiple contexts. Logic-Muse components were validated and argued by experts throughout the designing process (ITS researchers, logicians, and reasoning psychologists). A catalog of reasoning errors (syntactic and semantic) has been established, in addition to an explicit representation of semantic knowledge and the structures and meta-structures underlying conditional reasoning. A Bayesian network with expert validation has been developed and used in a Bayesian Knowledge Tracing (BKT) process that allows the inference of the learner skills. This paper presents an evaluation of the learner-model components in Logic-Muse (a bayesian learner model). We conducted a study and collected data from nearly 300 students who processed 48 reasoning activities. These data were used to develop a psychometric model for initializing the learner's model and validating the structure of the initial Bayesian network. We have also developed a neural architecture on which a model was trained to support a deep knowledge tracing (DKT) process. The proposed neural architecture improves the initial version of DKT by allowing the integration of expert knowledge (through the Bayesian Expert Validation Network) and allowing better generalization of knowledge with few samples. The results show a significant improvement in the predictive power of the learner model. The analysis of the results of the psychometric model also illustrates an excellent potential for improving the Bayesian network's structure and the learner model's initialization process.

This paper presents a simple and intuitive technique to accelerate the convergence of first-order optimization algorithms. The proposed solution modifies the update rule, based on the variation of the direction of the gradient and the previous step taken during training. Results after tests show that the technique has the potential to significantly improve the performance of existing first-order optimization algorithms.

We present a serious game designed to help players/learners develop socio-moral reasoning (SMR) maturity. It is based on an existing computerized task that was converted into a game to improve the motivation of learners. The learner model is computed using a hybrid deep learning architecture, and adaptation rules are provided by both human experts and machine learning techniques. We conducted some experiments with two versions of the game (the initial version and the adaptive version with AI-Based learner modeling). The results show that the adaptive version provides significant better results in terms of learning gain.

Deep Knowledge Tracing (DKT), along with other machine learning approaches, are biased toward data used during the training step. Thus, for problems where we have few amounts of data for training, the generalization power will be low, the models will tend to give good results on classes containing many examples and poor results on those with few examples. Theses problems are frequent in educational data where for example, there are skills that are very difficult (floor) or very easy to master (ceiling). There will be less data on students that correctly answered questions related to difficult knowledge and that incorrectly answered questions related to knowledge easy to master. In that case, the DKT is unable to correctly predict the student's answers to questions associated with those skills. To improve the DKT, we penalize the model using a 'cost-sensitive' technique. To overcome the problem of the few amounts of data, we propose a hybrid model combining the DKT and expert knowledge. Thus, the DKT is combined with a Bayesian Network (built from domain experts) by using the attention mechanism. The resulting model can accurately track knowledge of students in Logic-Muse Intelligent Tutoring System (ITS), compared to the BKT and the original DKT.

Deep learning has been successfully applied to many tasks such as image classification, feature learning, Text classification (sentiments analysis or opinion mining) etc. However, little research has focused on extracting polarity of sentiments expressed in text using a multimodal architecture. In other words, no researches take in consideration the multimodal nature of human behaviors before classifying sentiments. The representation of a person (also call User Modeling in some domains such as Intelligent Tutoring Systems) is an important feature to take in consideration if one wants to extract subjective information such as the polarity of sentiments expressed by the person. To design an effective representation of a user, it is important to consider all sources of data informing about its current state. We present a usersensitive deep multimodal architecture which takes advantage of deep learning and user data to extract a rich latent representation of a user. This rich latent representation mainly helps in text classification tasks. The architecture consists of the combination of a Long Short-Term Memory (LSTM), LSTM-Auto-Encoder, Convolutional Neural Networks and multiple Deep Neural Networks, in order to support the multimodality of data. The resulting model has been tested on a public multimodal dataset and is able to achieve best results compared to state-of-the-art algorithms for a similar task: detection of opinion polarity. The results suggest that the latent representation learnt from multimodal data helps in the discrimination of polarity of opinion.

Prediction of emotions is important for understanding human be-havior and modeling users in learning environments. In this paper,we present a deep multi-modal architecture for emotions predic-tion, which takes advantage of deep learning, user multimodal dataand the hierarchy of human memory. The architecture consists ofthe combination of Long Short-Term memory (LSTMs). One of thenovelty of our approach is that, we enhance the LSTM with anexplicit memory since in brain studies, the memory is often dividedinto two further main types: explicit (or declarative) memory andimplicit (or procedural) memory, the last one being the main pur-pose of LSTMs architectures. The resulting model has been testedon a public multi-modal dataset.

In this paper we present a participatory approach to design Logic-Muse, an Intelligent Tutoring System that helps learners develop reasoning skills in multiple contexts (situations). The study was conducted jointly with the active participation of experts in the field of logic and the psychology of reasoning. An explicit catalogue of systematic errors in classical logic is built, followed by an explicit representation and encoding of the semantic knowledge behind reasoning as well as reasoning procedural structures and meta-structures. Logic-Muse innovates through its design rationale, which leads to strong structures on which learning is based. It also innovates with the projection of reasoning skills in a variety of well-defined classes of situations to ensure an absolute mastery of reasoning skills regardless of the content effect.