能否构建输入含不同单位的Neural Networks？拆分网络合并结果可行吗？

Great question! Let’s break this down clearly and practically:

1. Yes, You Can Use Mixed-Unit Inputs in a Single Neural Network #

You absolutely don’t need to split your network just because your inputs have different units (like temperature in °C and a chemical product’s percentage). The only critical step here is feature scaling—neural networks are sensitive to the magnitude of input values. For example, if temperature ranges from 0–100 and your percentage spans 0–1, the larger-scale temperature feature will dominate the model’s training and throw off its ability to learn patterns from the percentage data.

Fixing this is straightforward with two common methods:

• Standardization (Z-score scaling): Convert each feature to have a mean of 0 and standard deviation of 1. Works well if your features follow a roughly normal distribution.
• Min-Max Normalization: Scale each feature to a fixed range (usually 0–1). Ideal if you know the exact min/max values of your features and want bounded outputs.

Here’s a quick Python example using scikit-learn to scale mixed-unit inputs:

from sklearn.preprocessing import StandardScaler
import numpy as np

# Sample input data: [temperature (°C), chemical percentage]
raw_inputs = np.array([[22, 0.15], [35, 0.3], [18, 0.08]])

# Fit scaler to the data and transform inputs
scaler = StandardScaler()
scaled_inputs = scaler.fit_transform(raw_inputs)

# Now both features are on a comparable scale for the neural network
print(scaled_inputs)

Once scaled, your neural network can learn patterns across all input features just like any other dataset.

2. Splitting into Subnetworks (and Merging Results) is Also a Valid Option #

If you want to modularize your model—for example, if temperature data has complex temporal patterns (like hourly readings) that need a dedicated recurrent layer, while percentage data is static and works better with dense layers—splitting into subnetworks is a smart approach.

Here’s a high-level structure for this setup:

• Subnetwork A: Takes temperature data as input, processes it with layers tailored to its structure (e.g., LSTMs for time-series, dense layers for static values), and outputs a compressed feature vector.
• Subnetwork B: Takes chemical percentage data (and any related features) as input, processes it with appropriate layers, and outputs another feature vector.
• Merge Step: Concatenate the two feature vectors from the subnetworks, then pass them through a final set of dense layers to produce the final prediction.

This approach lets each subnetwork specialize in learning patterns from its specific feature type, which can be useful for complex, multi-modal datasets. That said, it’s overkill for simple cases where scaling alone works perfectly.

Final Takeaway #

Start with a single neural network plus proper feature scaling—it’s simpler, easier to train, and works for most cases. Reserve the subnetwork approach for scenarios where different input types require distinct processing logic (e.g., time-series vs. tabular data).

内容的提问来源于stack exchange，提问作者Aceconhielo