Broad Learning System (BLS) 宽度学习系统

        宽度学习（Broad Learning System, BLS）是一种有效的神经网络学习框架，旨在通过扩展网络的宽度而不是深度来提高学习能力和效率。与传统的深度学习相比，宽度学习通过堆叠多层特征节点和增强节点来构建网络，从而避免了深度学习中常见的梯度消失和复杂的训练过程。

BLS结构以及增量算法 

        宽度学习系统在 RVFLNN 基础上做出了改进。首先，宽度学习可以利用别的模型提取到的特征来训练，即可以和别的机器学习算法灵活地结合。其次，宽度学习中加入了增量学习算法，它允许在网络结构中加入新的结点时，以很小的计算开销来更新网络权重。这一特性使 BLS 在面对大规模的数据时，相对于深度结构具有巨大的优势。

通过 (a) 图基于 Pytorch 简单实现的 BLS 模型 ：

class BLS(nn.Module):
    def __init__(self, input_dim, num_feature_nodes, num_enhancement_nodes, output_dim):
        super(BLS, self).__init__()
        # Step 1: Define feature mapping layers
        self.feature_layers = nn.ModuleList([nn.Sequential(
            nn.Linear(input_dim, num_feature_nodes),
            nn.ReLU()) for _ in range(num_feature_nodes)])

        # Step 2: Define enhancement nodes
        self.enhancement_layers = nn.ModuleList([nn.Sequential(
            nn.Linear(num_feature_nodes * num_feature_nodes, num_enhancement_nodes),  # 100,400
            nn.ReLU()) for _ in range(1)])

        # Step 3: Output layer
        self.output_layer = nn.Linear(
            num_feature_nodes * num_feature_nodes + num_enhancement_nodes, output_dim)

    def forward(self, x):
        print(x.shape) # torch.Size([32, 224])
        # Generate feature nodes Z^n
        Z = torch.cat([layer(x) for layer in self.feature_layers], dim = 1)
        print(Z.shape) # torch.Size([32, 100])
        # Generate enhancement nodes H^m
        H = torch.cat([layer(Z) for layer in self.enhancement_layers], dim = 1)
        print(H.shape) # torch.Size([32, 25])
        # Concatenate and predict
        combined = torch.cat((Z, H), dim = 1)
        print(combined.shape) # torch.Size([32, 125])
        output = self.output_layer(combined)
        print(output.shape) # torch.Size([32, 1])
        return output

通过 (b) 图基于 Pytorch 简单实现的 BLS 模型 ：

class BLSv2(nn.Module):
    def __init__(self, input_dim, num_feature_nodes, num_enhancement_nodes, output_dim):
        super(BLSv2, self).__init__()
        # Step 1: Define feature mapping layers
        self.feature_layers = nn.ModuleList([nn.Sequential(
            nn.Linear(input_dim, num_feature_nodes),
            nn.ReLU()) for _ in range(num_feature_nodes)])

        # Step 2: Define enhancement nodes
        self.enhancement_layers = nn.ModuleList([nn.Sequential(
            nn.Linear(num_feature_nodes * num_feature_nodes, num_enhancement_nodes), # 100,20
            nn.ReLU()) for _ in range(num_enhancement_nodes)])

        # Step 3: Output layer
        self.output_layer = nn.Linear(
            num_feature_nodes * num_feature_nodes + num_enhancement_nodes * num_enhancement_nodes, output_dim)

    def forward(self, x):
        print(x.shape) # torch.Size([32, 224])
        # Generate feature nodes Z^n
        Z = torch.cat([layer(x) for layer in self.feature_layers], dim = 1)
        print(Z.shape) # torch.Size([32, 100])
        # Generate enhancement nodes H^m
        H = torch.cat([layer(Z) for layer in self.enhancement_layers], dim = 1)
        print(H.shape) # torch.Size([32, 25])
        # Concatenate and predict
        combined = torch.cat((Z, H), dim = 1)
        print(combined.shape) # torch.Size([32, 125])
        output = self.output_layer(combined)
        print(output.shape) # torch.Size([32, 1])
        return output

更多资料： 

宽度学习系统（BLS）的原理、变体形式及当前应用（随时更新……）「建议收藏」-腾讯云开发者社区-腾讯云 (tencent.com)

DeepLearning | Broad Learning System 宽度学习系统 : 高效增量式浅层神经网络-CSDN博客