Brain-inspired learning guidelines have been successfully applied to improve pulsed neural networks.

Recently, Zeng Yi, a researcher at the Center for Brain Intelligence Research at the Institute of Automation of the Chinese Academy of Sciences, and team members summarized seven brain-inspired learning guidelines in a study that was successfully applied to improve pulsed neural networks.
by combining different brain-inspired rules, experimental studies have confirmed that deep pulsed neural networks are getting better and better at classifying as more and more carefully selected, brain-inspired rules are introduced.
Although pulse neurons have a more solid biological reality than traditional artificial neurons, the traditional pulsed neural network model only captures the initial local learning and training rules in the processing of brain information.
because the biological brain in the learning process of several laws, it can not be expected that only the use of fewer laws designed brain pulse neural network model can achieve or even beyond the biological brain of various learning abilities. The seven learning guidelines proposed by the
research team are derived from the experimental study of the biological brain and each reflects the learning characteristics of the biological network from different sides, such as the dynamic distribution of neurons, the adaptive growth and extinction mechanism of synapses, different synoptic plastic learning mechanisms (such as different types of timing-dependent synoptic plasticity), the regulation mechanism of network background noise to learning, the mechanism of excitability and the ratio of inhibitory neurons to learning.
these learning criteria, the dynamic distribution of neurons, the generation and extinction of synapses, plasticity models and so on are considered to be important characteristics of brain neural network processing information.
team introduced the brain-inspired rules into pulsed neural network models in the hope of improving the efficiency of traditional pulsed neural networks.
the pulse neural network model proposed by the research team mainly consists of three parts: pulse generation layer, hidden layer, output layer.
in the pulse generation layer, the static image input is converted into a pulse sequence.
In the cryptosphere, dynamic distribution of neurons (R1, R2), synoptic growth and extinction (R3, R4), different types of background noise (R5), different types of pulse timing dependent plasticity models (R6), excitable and inhibitory neurons (R7) were introduced as important mechanisms of the brain.
the output layer, excitatory neurons are responsible for classifying, inhibiting neurons to achieve the winner-take-all mechanism (WTA).
experimental verification, the study used the handwriting number data set MNIST.
experiments show that the correct rate of the model will gradually improve when the carefully selected brain-like mechanism is introduced.
researchers say optimal model performance is not the ultimate goal of the study. The model proposed by the
team is characterized by a more biologically interpretable computer system than traditional neural network models, in which case understanding the meaning of brain information processing mechanisms and processes through computational modeling, and the resulting inspiration, is far greater than the pursuit of performance.
the main conclusion of this study is that each of the introduced mechanisms has its own unique contribution to pulsed neural network processing pattern recognition class problems and is irreplaceable to each other (Figure 1).
In addition, the proportion of excitable and inhibitory neurons that achieved the best pattern recognition efficiency on the MNIST dataset was 15% inhibitory neurons and 85% excitable neurons (as shown in Figure 2), which is closely consistent with the ratio of excitability and inhibitory neurons in the biopsy cerebral cortical region, which initially supports the evolution of the brain nervous system in the direction of cognitive task optimization from a computational perspective.
the study was published in Science China Information Sciences.
.