Gone are the days when Excel was just for accounting. By leveraging the function—which makes Excel Turing-complete—you can now define complex recursive logic like backpropagation and weight updates right in your formula bar. 1. Architecture: The Grid Layout
Because native Excel cells do not self-update iteratively without circular references, we have two modern ways to execute training loop iterations:
For a simple single-hidden layer feed-forward network, follow these steps to set up the architecture manually without VBA. 1. Input and Weight Initialization
For a simple network with one input column, weights in another, and a bias stored in a separate cell, you can compute the weighted sum with =SUMPRODUCT(input_range, weight_range) + bias and then apply the activation. For larger networks, MMULT and array formulas allow you to scale up efficiently.
=LET( Z2, MMULT(Hidden_Active#, Weights!B6#) + Weights!E6#, A2, SIGMOID(Z2), A2 ) Use code with caution. Name this spilled result Predictions# . 6. The Backward Pass (Gradient Descent)
Tip: Use RAND() to initialize weights between 0 and 1 initially. 4. Implementing Forward Propagation (The Math)
Modern Excel allows you to build custom, reusable functions without VBA using LAMBDA . We need two activation functions: ReLU and Sigmoid.
output = 1 / (1 + exp(-(weight1 * neuron1_output + weight2 * neuron2_output + bias)))
No environment setup or code libraries (like TensorFlow or PyTorch) required.