Dendritic Neural Mode

  • Abstract: DNM is employed to solve classification problems. Since the DNM is a feed-forward model, and all the excitation functions are continuous and differentiable, the Error Back-propagation (EBP) algorithm is used to train it, whose learning rule is derived from the Least Squared Error between the actual outputs and the desired targets. Further, the DNM will be simplified and then be transformed by a Logic Circuit Classifier (LCC), which consists of the comparators and logic AND, OR and NOT gates. Although the EBP algorithm may trap into local minima occasionally, it has very fast convergence speed and satisfactory optimization performances.

  • Source code: Link Latest Update Date: 2021.06.07

  • Citation: Junkai Ji, Shangce Gao, Jiujun Cheng, Zheng Tang, and Yuki Todo. “An approximate logic neuron model with a dendritic structure,” Neurocomputing 173 (2016): 1775-1783. Link

Dendritic Neural Regression

  • Abstract: The dendrite neural regression (DNR) employs a novel weight to describe the dendrite strength, which can enhance the regression ability of the model significantly. A recently proposed optimization algorithm named AMSGrad is utilized in DNR training. AMSGrad is a variant of the Adam algorithm, which can speed up the convergence of DNR during the optimization procedure. The DNR trained by the AMSGrad algorithm (ADNR) has demonstrated excellent and stable performance in the real regression problem.

  • Source code: Link Latest Update Date: 2022.06.07.

  • Citation: Junkai Ji, Minhui Dong, Qiuzhen Lin, and Kay Chen Tan. "Noninvasive Cuffless Blood Pressure Estimation With Dendritic Neural Regression."IEEE Transactions on Cybernetics 2022. Accepted.

DNM Trained by the States of Matter Search Algorithm

  • Abstract: DNM is also applied to solve classification problems. It is trained by a metaheuristic algorithm, named the states of matter search (SMS) algorithm. The evolutionary operations of SMS are based on the physical principle of the thermal-energy motion ratio, and the whole optimization process is divided into the following three phases: the gas state, the liquid state and the solid state. Each state has its own operations with different exploration–exploitation ratios. The SMS algorithm can be regarded as a more global search approach. Empirical evidences have verified that it can provide better training performance than several state-of-the-art evolutionary algorithms, such as: Genetic Algorithm (GA), Particle Swarm Optimization (PSO) and adaptive differential evolution with optional external archive (JADE).

  • Source code: Link Latest Update Date: 2021.06.07

  • Citation: Junkai Ji, Shuangbao Song, Yajiao Tang, Shangce Gao, Zheng Tang, and Yuki Todo. "Approximate logic neuron model trained by states of matter search algorithm." Knowledge-Based Systems  163 (2019): 120-130. Link

DNM Trained by the Multi-objective Differential Evolution Algorithm

  • Abstract: The architecture of DNM affects the learning capacity, generalization capability, computing time and approximation of LCC. Thus, a Pareto-based multiobjective differential evolution (MODE) algorithm is proposed to simultaneously optimize DNM’s topology and weights. The mean squared error of the training dataset and the model complexity are selected as the two objectives of MODE. MODE can generate a concise and accurate LCC for every specific task from DNM.

  • Source code: Link Latest Update Date: 2022.06.07

  • Citation: Junkai Ji, Yajiao Tang, Lijia Ma, Jianqiang Li, Qiuzhen Lin, Zheng Tang, and Yuki Todo. "Accuracy versus simplification in an approximate logic neural model." IEEE Transactions on Neural Networks and Learning Systems 32.11 (2020): 5194-5207. Link