https://www.toutiao.com/a6701844289518830091/
The topology of the neural network focuses pulses, pulse sequence information encoding method, the pulse neural network learning algorithm and evolutionary methods.
A pulse neural network topology
With traditional artificial neural networks, spiking neural network is also divided into three topologies. They are a feedforward neural network type pulse (feed-forward spiking neural network), the pulse recursive neural network (recurrent spiking neural network) and a hybrid type pulse neural network (hybird spiking neural network).
1. Pulse feedforward neural network
In the structure of the front feed pulse multilayer neural network, the neurons in the network is layered, the pulse sequence of each input layer neuron represents the input encoded data on a specific issue, and the next layer of the neural network input pulse . Last layer is the output layer which pulse sequence of each neuron output constitutes the output of the network. There may be one or more hidden layers between the input layer and the output layer.
Further, in the conventional feed-forward artificial neural network, only one synaptic connection between two neurons, and the network structure of the neural network may employ a multi-pulse of synaptic connections, a plurality of projections may be between two neurons contact connections, each having different delays and synapses modifiable connection weights. Polysynaptic different pulse delay allows the presynaptic neuron input pulse can be postsynaptic neuron issuance of an impact on a longer time scale. Presynaptic neuronal transmission of the plurality of pulses then have different postsynaptic potentials depending on the magnitude of the synaptic weights.
2. recursive spiking neural networks
Recursive neural network is different from a single layer feed forward neural networks and multi-layer neural network, the network having a feedback loop structure, i.e., the network output neuron is the previous time step of the recursive functions of neuron output. Recurrent neural network can be simulated time series, to complete the control, task prediction, the feedback mechanism which on the one hand so that they can exhibit more complex time-varying system; on the other hand makes the design and Convergence Analysis of a more effective learning algorithm difficult. Backpropagation learning algorithm conventional two classic recurrent artificial neural network learning are real recursive (real-time recurrent learning) algorithm and the evolution over time (backpropagation through time) algorithm, both algorithms are calculated recursively gradient learning algorithm.
Pulse Recurrent Neural Network refers to a network having a pulse neural network feedback loop, because of its information feedback mechanism different from the conventional encoding and recurrent neural network, whereby the network learning algorithm and Construction kinetic analysis difficult. Pulse Recurrent Neural Network may be applied to many complex problem solving, language such as modeling, digital handwriting recognition and speech recognition. Pulse Recurrent Neural Network can be divided into two broad categories: Global Pulse Recurrent Neural Network (fully recurrent spiking neural network); another is a partial pulse neural network (locally recurrent spiking neural network).
3. Impulsive Neural Network
Impulsive i.e. neural network comprises a feedforward structure, and comprising a recursive structure.
Second, the pulse sequence information encoding method
From the viewpoint of neuroscience, second-generation artificial neural network is a neuron calculation "firing frequency" based. With further research, the scientists pointed out that the nerve pulse timing of biological nervous system neurons to encode information, and not just with the pulse of neurons "firing frequency" to encode information. In fact, the neuron firing frequency pulse not fully capture the information contained in the pulse sequence. For example, it has been found that the primary auditory cortex neuron population can be coordinated in a short time by the packet relative time adjacent pulses of action potentials, and does not change the number of pulses per second issue such that even neurons are no changes in the average firing frequency given specific case where the stimulation signal.
More biologically interpretable pulse neural network, using precisely timed pulse sequence to encode information nerve. Neural network internal information transmission is accomplished by a pulse sequence, the pulse sequence is a pulse time sequence of discrete time point of the composition, and therefore, during the simulated neural network to calculate the pulse, comprising the steps of: ① When the input data or neurological epoch external stimuli, through a particular pulse sequences encoding method, the data may be encoded into a specific external stimulus or pulse sequence; ② and pulse sequence is transmitted after a certain processing between neurons, output the pulse sequence after processing by the the specific decoding method decodes and give a specific response.
For information pulse sequences encoding neurological problems, drawing on information coding mechanism of biological neurons, researchers have proposed a number of pulse sequences encoding method spiking neural networks. For example, the encoding method of the first pulse trigger time, delay phase encoding method, the encoding method and the like groups.
Third, the pulse neural network learning algorithm
Learning is the core issue in the field of artificial intelligence, for SNN, the study pulse time level learning method based on a theoretical model to be verified by the information processing and learning mechanism of biological nervous system is a must. Artificial neural systems through bio-interpretable way, scientists hope to achieve the intended purpose can experiment by neuroscience and behavior. Learning in the brain can be understood as the process of synaptic connection strength change with time, this ability is called synaptic plasticity (synaptic plasticity). Learning spiking neural networks include unsupervised learning (unsupervised learning), supervised learning (supervised learning) and reinforcement learning (reinforcement learning) and so on.
1. unsupervised learning algorithm
Unsupervised learning algorithm to dominate humans and animals learn, people can discover by observing the internal structure of the world, rather than being told the name of each objective things. Artificial neural networks unsupervised learning algorithm is designed primarily for unlabeled training data set, it requires the application of the rules of unsupervised learning connection weights or structure of the neural network adaptive adjustments. In other words, under the supervision of no "teacher" signal, the neural network must find their own laws (such as statistical characteristics, correlation or other categories) from the input data, and to classify or decision by output. In general, only when there is redundancy in the input dataset, unsupervised learning makes sense, otherwise unsupervised learning can not properly detect any pattern or signature input data that provides redundancy knowledge.
Most of unsupervised learning neural network algorithm impulse is to learn from the traditional artificial neural network unsupervised learning algorithm is proposed based on different variants of Hebb learning rules. Neuroscience research shows that the pulse sequence of biological nervous system not only can cause persistent changes in neuronal synapses, and that satisfies the timing dependent plasticity (spike timing-dependent plasticity, STDP) mechanism. In the decisive time window, depending on the relative timing relationship between the presynaptic neurons and postsynaptic pulse sequence of neural firing, the application of STDP learning rule can be adjusted unsupervised way of synaptic weights.
2. Pulse supervised learning neural network
Pulse supervised learning neural network means for a given input pulse train and a plurality of multiple target pulse sequence to find a suitable pulse synaptic neural network weight matrix, the target pulse sequence and the corresponding sequence of output pulses do neurons It may be close to, i.e., both the minimum error evaluation function. For spiking neural networks, neural information is represented in the form of a pulse sequence, neuronal function of internal state variables and the error no longer satisfies the properties continuously differentiable, the effective pulse neural network supervised learning algorithm is very difficult to construct, but also in the field an important research direction.
Depending on the basic idea of supervised learning adopted, existing supervised learning algorithm can be divided into three categories:
- Based on the principle of gradient descent supervised learning algorithm using the error between the neurons and the target output and the actual output error back propagation process, the reference amount to obtain the calculation result as the gradient descent to adjust the synaptic weights, and ultimately reduce this error. In supervised learning based on gradient descent algorithm is a mathematical method for analyzing, learning rule in the derivation of the required state variables must neuron model is analytic expression, mainly linear neuron model fixed threshold, such as the impulse response model (spike response model) and Integrate-and-Fire neuron model.
- Based on the basic idea of synaptic plasticity supervised learning algorithm is to use the mechanisms of synaptic plasticity neuronal firing pulse time sequence of correlation caused by design neurons to learn the rules of the synaptic weights adjustment, which is a biologically interpretable supervision of learning.
- Supervised Learning Algorithm pulse sequence supervised learning algorithm convolutional inner product of the difference pulse sequence pulse neural network configured to adjust the synaptic weights depend on the specific convolution kernel is calculated, the pulse sequence may be implemented to learn the spatial and temporal patterns.
3. Pulse neural network reinforcement learning
Jackpot value reinforcement learning learning environment is a mapping from state to act, so that the agent acts to get the maximum from the environment. Learning mechanism based bio-inspired artificial neural network reinforcement learning research focused on exploring adaptive agent optimization strategy, in recent years, it is one of the neural networks and intelligent control of the main methods. Reinforcement learning concern is how to take a series of actions the agent in the environment, through reinforcement learning, an intelligent body should know in what state should take what actions. It can be seen to strengthen the difference between learning and supervised learning lies mainly in the following two points:
- Reinforcement learning is trial and error learning, because there is no direct "teachers' guide information, the agent must continue to interact with the environment to get the best strategy through trial-and-error;
- Delayed return instructions, reinforcement learning are few and often after the fact (the last state) was given, which leads to a problem that the positive returns or after negative returns, how will report assigned to the previous state .
Fourth, the evolutionary approach of impulsive neural networks
Evolutionary algorithms (evolutionary algorithm) is a simulation model of biological evolution, is a class based on biological evolution mechanism of natural selection and genetic mutation of global probabilistic search algorithms, including genetic algorithms (genetic algorithm), evolutionary programming (evolutionary programming) and evolutionary strategy (evolutionary strategy) and so on. Although these algorithms have some differences in terms of realization, but they have one thing in common, that is to solve practical problems by means of the concepts and principles of biological evolution.
The evolutionary algorithms and neural networks combine the pulse, the researchers opened up the field of evolutionary research spiking neural network (evolutionary spiking neural network) to improve the ability to solve complex problems. Evolutionary pulse neural networks can be used as a generic framework for adaptive systems, adaptive adjustment without human intervention neuron system parameters, connection weights, and the network structure learning rules.