如何用卷积层替代全连接层实现25×375到30×375矩阵投影（Matlab）

Hey there! Let's break down how to work around MATLAB's limitation with per-column fully connected layers, using an equivalent convolutional layer that does exactly what your paper describes. This trick is perfect for structured time-series data like EEG recordings.

Core Idea #

Your original fully connected layer takes a 25×375 matrix (25 features × 375 time steps) and applies a 30-node FC layer to each column individually—turning every 25-dimensional time step into a 30-dimensional vector, resulting in a 30×375 output.

This is exactly what a 1×1 convolutional layer does! A 1×1 convolution operates on each spatial position (here, each time step) independently, combining all input channels (your 25 features) into new output channels (the 30 you need). It’s a direct, drop-in equivalent for your per-column FC layer.

Parameter Design (MATLAB-Friendly) #

MATLAB’s deep learning toolbox expects inputs in the format height × width × channels × batch size, so we’ll adjust our setup to fit that:

1. Reshape Your Input

   • For a single 25×375 EEG sample, reshape it to 375 × 1 × 25 × 1:
     • Height = number of time steps (375)
     • Width = 1 (we’re treating each time step as a single "pixel" in width)
     • Channels = original feature count (25)
     • Batch size = 1 (for a single sample)
   • For batches, use 375 × 1 × 25 × N where N is your batch size.

2. Configure the Convolutional Layer

   • Kernel size: [1 1] (we don’t want to mix time steps—just process each one alone)
   • Number of output channels: 30 (matches the FC layer’s 30 nodes)
   • Padding: 'valid' (no padding needed, since 1×1 kernels fit perfectly)
   • Stride: [1 1] (process every time step)
   • Weight/bias mapping:
     • Your original FC layer has a 30×25 weight matrix W. MATLAB’s conv layer expects weights in kernel height × kernel width × input channels × output channels format, so reshape W' to [1 1 25 30].
     • The bias vector (30×1) from the FC layer can be used directly for the conv layer.

MATLAB Code Example #

% Assume you have the original FC layer weights (W: 30×25) and bias (b: 30×1)
inputDims = [375 1 25]; % Height × Width × Input Channels
numOutputs = 30;

% Create the equivalent 1x1 convolution layer
convLayer = conv2dLayer([1 1], numOutputs, ...
    'Padding', 'valid', ...
    'Stride', [1 1]);

% Transfer weights and bias from the original FC layer
convLayer.Weights = reshape(W', [1 1 25 30]); % Transpose W to match conv dimensions
convLayer.Bias = b;

% Test with a sample EEG input
sampleEEG = randn(25, 375); % 25 features × 375 time steps
% Reshape to fit MATLAB's conv layer input format
reshapedInput = reshape(sampleEEG', [375 1 25 1]);
% Run forward pass
convOutput = predict(convLayer, reshapedInput);
% Convert back to your original 30×375 format
finalOutput = reshape(convOutput, [375 30])';

% Verify equivalence with original FC layer output
fcOutput = W * sampleEEG + b;
disp(max(abs(finalOutput - fcOutput))); % Should be ~0, confirming they're identical

Why This Works #

• A 1×1 convolution is mathematically identical to applying a fully connected layer to each spatial position (time step) independently.
• MATLAB’s conv layers handle this natively, no workarounds needed for per-column processing.
• You can directly reuse the weights and biases from the paper’s FC layer, so your model will behave exactly as described.

内容的提问来源于stack exchange，提问作者S.MC.