EgoModeling: Real-time Modelling of Brain Connections – This paper proposes a novel method for real-time brain network prediction, via a learning-to-learn paradigm. A supervised learning framework is then developed to perform multi-task, multi-context learning. To handle the need to model long-term dependencies, we propose an iterative update of the neural network, which in turn leverages local recurrent connections to learn to predict long-term connections. The iterative framework, which we call a recurrent-bisterent framework, uses the output of a pair-wise graph to predict the most relevant local connections based on their correlation and is robust to variations in parameters. Moreover, by using the prediction results from this framework, a training set of long-term brain connections is obtained. The proposed method is evaluated on several benchmark data sets, showing that our method has high predictive performance and provides good computational power.

The proposed fast-forward algorithm (FIFTH) is a variant of the L-SAT algorithm that uses binary classification (CAS) instead of explicit classifiability (CAS) for the classification task. The CAS algorithm is based on a fast method for classification based on binary classifiers using the concept that a classifier which can correctly classify the data is a good candidate for CAS (CAS) classification. The main disadvantage of the CAS algorithm is that (1) the CAS algorithm requires many computational resources and (2) an explicit CAS process to operate. Therefore, the CAS algorithm is more suitable for training the CAS system. In this paper, we propose an independent and competitive learning algorithm that combines multiple CAS process and CAS process for CAS classification task. Experimental results on all benchmark datasets show a significant improvement in classification quality over CAS and CAS-SAT algorithms.

Conceptual Constraint-based Neural Networks

Visual Speech Recognition using Deep Learning

EgoModeling: Real-time Modelling of Brain Connections

A Survey of Artificial Neural Network Design with Finite State Counting

The Fast-Forward AlgorithmThe proposed fast-forward algorithm (FIFTH) is a variant of the L-SAT algorithm that uses binary classification (CAS) instead of explicit classifiability (CAS) for the classification task. The CAS algorithm is based on a fast method for classification based on binary classifiers using the concept that a classifier which can correctly classify the data is a good candidate for CAS (CAS) classification. The main disadvantage of the CAS algorithm is that (1) the CAS algorithm requires many computational resources and (2) an explicit CAS process to operate. Therefore, the CAS algorithm is more suitable for training the CAS system. In this paper, we propose an independent and competitive learning algorithm that combines multiple CAS process and CAS process for CAS classification task. Experimental results on all benchmark datasets show a significant improvement in classification quality over CAS and CAS-SAT algorithms.