Most neural networks are constructed by a large number of neurons with single activation function, in this way, huge number of neurons are used to fit the training data. It is very easy to overfit the training samples, and consequently resulting in poor generalization ability. Moreover, the more neurons are included in the network, the greater the computing power and storage are needed. In this paper, a Self-Adaptive and Multiple Activation Function Neural Network with a small number of neurons is proposed to recognize facial expression images. The experiment demostrates that the accuracy of our network is competitive with the existing networks with the accurancy is improved by 9.85% while the VRAM (Video RAM) and the GPU usage is reduced by 61.74% and 38.5% in average. When the noise perturbation is introduced into the neuron activation function group to enhance the generalization ability of network, the VRAM is kept the same, the average GPU usage is increased by 2.5% and the average accuracy is improved by 3.83% further comparing with the results of the proposed network without noise perturbation.

This paper proposes and demonstrates a single-line discontinuous track recognition system by associating the track recognition problem of a humanoid robot with the lane detection problem. The proposal enables the robot to achieve stable running on the single-line discontinuous track. The system consists of two parts: the robot end and the graphics computing end. The robot end is responsible for collecting track information and the graphics computing end is responsible for high-performance computing. These two parts use the TCP for communication. The graphics computing side uses PolyLaneNet lane detection algorithm to train the track image captured from the first perspective of the darwin-op2 robot as the data set. In the inference, the robot end sends the collected tracking images to the graphics calculation end and uses the graphics processor to accelerate the calculation. After obtaining the motion vector, it is transmitted back to the robot end. The robot end parses the motion vector to obtain the motion information of the robot so that the robot can achieve stable running on the single-line discontinuous track. The proposed system realizes the direct recognition of the first perspective image of the robot and avoids the problems of poor stability, inability of identifying curves and discontinuous lines, and other problems in the traditional line detection method. At the same time, this system adopts the method of cooperative work between the PC side and the robot by deploying the algorithm with high computational requirements on the PC side. The data transmission is carried out by stable TCP communication, which makes it possible for the robot equipped with weak computational controllers to use deep-learning-related algorithms. It also provides ideas and solutions for deploying deep-learning-related algorithms on similar low computational robots.

This paper proposes a meta-learning and reinforcement-learning based control system ”Cerebellum” for robotic arms to address the issue of adapting to hardware parameter changes. By leveraging the con- cept of meta-learning, the system achieves adaptive control by learn- ing from few samples. Experimental results show our ”Cerebellum” reach 90.09% accuracy in a simulation environment with control- ling newly unseen robotic arm. However, challenges remain in hard- ware experiments due to image distortions and Mechanical collide issues. Further research is needed to overcome these challenges.

Artificial neural networks, particularly convolutional neural networks (CNNs), have shown remarkable performance in tasks like image classi- fication. However, their robustness under dendritic degradation remains a challenge. Inspired by neurophysiology, we propose a Probabilis- tic Dendritic Activation (PDA) method to improve CNNs’ resilience under dendritic degradation. Our method enhances memory retention in the visual cortex, even when a significant proportion of dendrites are randomly deactivated. We trained a CNN with PDA on the Ima- geNet dataset and observed a 7.5% and 1.6% improvement in Top-1 and Top-5 classification accuracies, respectively, under 50% random dendritic degradation. This study highlights the bidirectional contribu- tion between computational techniques and neurobiology, demonstrating the potential of PDA in enhancing neural network robustness and fostering a deeper understanding of biological neural systems’ resilience.

The existing computer neural network has relatively poor fault tolerance under the condition of internal dendritic inactivation, that is, when some neurons die or neuronal dendrites are inactivated, the function of the neural network will change greatly. We consider that the reason is the design of computer neural networks is too simple to recognize neuronal dendritic activation patterns. Therefore, we propose a Probabilistic activation pattern hypothesis of biological neuron dendrites. The hypothesis holds that if the dendrites of the neurons constituting the neural network are inactivated under a certain probability, the fault tolerance of the neural network to its internal dendrite inactivation can be significantly increased. Under this assumption, this paper designs typical computer neural network (such as classification network) experiments to verify the proposed hypothesis. In the control experiment, it is found that in the robustness test which contains a half of parameters' perturbation, the accuracy of Top1 and Top5 of the proposed method is 30.7% and 43.0% higher than that of the widely used Dropout method of half neurons death, and 7.5% and 1.6% higher than that of the method with full neurons activation. From the above experimental results, we can speculate that the brain dendrites of some animals may conform to the Probabilistic working mode.