《人工智能（智能系统指南，第二版）》读书笔记——7、第六章

1. introduction to knowledge-based intelligent systems(summary / questions for review / references)

2. rule-based expert systems

3. uncertainty management in rule-based expert systems

4. fuzzy expert systems

5. frame-based expert systems

6. artificial neural networks

7. evolutionary computation

8. hybrid intelligent systems

9. knowledge engineering and data mining

6. artificial neural networks

Machine learning involves adaptive mechanisms(自适应机制) that enable computers to learn from experience, learn by example and learn by analogy(类比). Learning capabilities can improve the performance of an intelligent system over time. The most popular approaches
to machine learning are artificial neural networks(人工神经网络) and genetic algorithms(遗传算法). This chapter is dedicated to neural networks.

An artificial neural network consists of a number of very simple and highly interconnected processors, called neurons(神经元), which are analogous to the biological neurons in the brain. The neurons are connected by weighted links that pass signals from one
neuron to another. Each link has a numerical weight associated with it. Weights are the basic means of long-term memory in ANNs. They express the strength, or importance, of each neuron input. A neural network "learn" through repeated adjustments of these
weights.

A neuron is a simplest computing element. Figure 6.3 is "diagram of a neuron". The input signal can be raw data or outputs of other neurons, the output signal can be either a final solution to the problem or an input to other neurons. The outputs is determined
by inputs transfer or activation function(激活函数). The most common choices for activation function are:the step, sign, linear and sigmoid functions(阶跃、符号、线性和S形函数), which are illustrated in figure 6.4. The step and sign activation functions, also called hard
limit functions(硬限幅函数), are often used in decision-making neurons for classification and pattern recognition tasks. Neurons with the linear function are often used for linear approximation(线性近似). Neurons with the sigmoid function are often used in the back-propagation
networks(后向传递网络).





In the 1940s, Warren McCulloch and Walter Pitts proposed a simple neuron model that is still the basis for most artificial neural networks. The neuron computes the weighted sum of the input signals and compares the result with a threshold value using the
linear function.

Frank Rosenblatt suggested the simplest form whof a neural network, ich he called a perceptron(感知器). The operation of the perceptron is based on the McCulloch and Pitts neuron model. It consists of a single neuron with adjustable synaptic weights(突触权重) and
a hard limiter. The perceptron learns its task by making small adjustments in the weights to reduce the difference between the actual and desired outputs. The inital weights are randomly assigned and then updated to obtain the output consistent with(和..一致)
the training examples.【initialisation, activation, weight training, iteration】

A perceptron can learn only linearly separable functions(线性分割函数,such as operations AND/OR but not Exclusive-OR,regarldless of whether we use a hard-limit or soft-limit activation function) and cannot make global generalisations
on the basis of examples learned locally. The limitations of Rosenblatt's perceptron can be overcome by advanced forms of neural networks, such as multilayer perceptrons trained with the back-propagation algorithm.

A multilayer perceptron is a feedforward neural network with an input layer of source neurons, at least one middle or hidden layer of computational neurons(with one hidden layer, we can represent any continuous function
of the input signals, and with two hidden layers even discontinuous functions can be represented), and an an output layer of computational neurons. The input layer accepts input signals from the outside world and redistributes these signals to all neurons
in the hidden layer. The hidden layer detects the feature. The weights of the neurons in the hidden layer represent the features in the input patterns. The output layer establishes the output pattern of the entire network.



Learning in a multilayer network proceeds in the same way as in a perceptron. The learning algorithm has two phases. First, a training input pattern is presented to the network input layer. The network propagates the input
pattern from layer to layer until the output pattern is generated by the output layer. If it is different from the desired output, an error is calculated and 【then propagated backwards through the network from the output layer to the input layer】. The weights
are modified as the error is propagated.【initialisation, activation, weight training, iteration】



Although widely used, back-propagation learning is not immune from problems. 1)Biological neurons do not work backward to adjust the strengths of their interconnections, and thus back-propagation learning cannot be viewed as a process that emulates brain-like
learning. 2)Because the calculations are extensive, and thus training is slow. The are several possible ways to improve computational efficiency. A multilayer network learns much faster when the sigmoidal activation function is represented by a hyperbolic
tangent(双曲正切函数). The use of momentum term(要素项) and adaptive learning rate(自适应学习速度) also significantly improves the performance of a multilayer back-propagation neural network.

While multilayer back-propagation neural networks are used for pattern recognition problems, the associative memory(关联记忆) of humans is emulated by a different type fo network called recurrent:a recurrent network(循环网络),
【which has feedback loops from its outputs to its inputs】.John Hopfield formulated the physical principle of the storing information in a dynamically stable network, and also proposed a single-layer recurrent network using McCulloch and Pitts neurons with
the sign activation function.

The Hopfield network training algorihm has two basic phases: storage and retrieval(检索). In the first phase, the network is required to stroe a set of states, or fundamental memories, determined by the current outputs of
all neurons. This is achieved by calculating the network's weight matrix. Once the weights are calculated, they remain fixed. In the second phase, an unknown corrupted(含糊的) or incomplete version of the fundamental memory is presented to the network. The network
output is calculated and fed back to adjust the input. This process is repeated until the output becomes constant. For the fundamental memories to be retrievable, the storage capacity of the Hopfield network has to be kept small.【storage, testing, retrieval】



The Hopfield network represents an autoassociative type of memory. It can retrieve a corrupted or incomplete memory but cannot associate one memory with another. To overcome this limitation, Bart Kosko proposed thebidirectional
associative memory(BAM,双向关联记忆). BAM is a heteroassociative network(异质相关网络). It associates patterns from one set to patterns from another set and vice versa(反之亦然). As with a Hopfield network, the BAM can generalise and produce correct outputs despite
corrupted or incomplete inputs. The basic BAM architecture consists of two connected layers - an input layer and an output layer.



The idea behind the BAM is to store pattern pairs so that when n-dimensional vector X from set A is presented as input, the BAM recalls m-dimensional vector Y from set B, and when Y is presented as input, the BAM recalls X. But there are some constraints:
the number of associations to be stored in the BAM should not exceed the number of neurons in the smaller layer and the BAM may not always produce the closest association(incorrect convergence).

In contrast to supervised learning, or learning with an external "teacher" who presents a training set to the network, unsupervised or self-organised learning does not require a teacher. During a training session, the neural network receives a number of
different input patterns, discovers significant features in these patterns and learns how to classify input.

(Hebbin learning)Hebb's Law states that iff neuron i is near enough to excite neuron j and repeatedly participates in its activation, the synaptic(突触) connection between these two neurons is strengthened and neuron j becomes
more sensitive to stimuli(刺激) from neuron i. This law provides the basis for learning without a teacher. Learning here is a local phenomenon occurring without feedback from the environment.



Another popular type of unsupervised learning is competitive learning. In competitive learning, neurons compete among themselves to become active. The output neuron that wins the "competition" is called the winner-takes-all
neuron. Although competitive learning was proposed in the early 1970s, it was rediscovered when Teuvo Kohonen introduced a special class of artificial neural networks called self-organising feature maps(自组织特征映射). He also formulated the principle of topographic
map(拓扑映射) formation which states that the spatial location(空间位置) of an output neuron in the topographic map corresponds to a particular feature of the input pattern.

The Kohonen network consists of a single layer of computation neurons, but it has two different types of connections. There are forward connections from the neurons in the input layer to the neurons in the output layer,
and lateral connections(横向链接) between neurons in the output layer. The lateral connections are used to create a competition between neurons. In the Kohonen network, a neuron learns by shifting its weights from inactive connections to active ones. Only the
winning neuron and its neighbourhood are allowed to learn. If a neuron does not respond to a given input pattern, then learning does not occur in that neuron.