You don’t need Python, TensorFlow, or expensive hardware to understand how deep learning works. In fact, you can build a fully functional neural network using only Microsoft Excel. This exercise is one of the most powerful ways to demystify the mathematics behind AI—forward propagation, backpropagation, and gradient descent—because you can see every calculation, every weight update, and every error change in real-time.
Formula for Z1=(X1⋅W11(1))+(X2⋅W21(1))+B1(1)Formula for cap Z sub 1 equals open paren cap X sub 1 center dot cap W sub 11 raised to the open paren 1 close paren power close paren plus open paren cap X sub 2 center dot cap W sub 21 raised to the open paren 1 close paren power close paren plus cap B sub 1 raised to the open paren 1 close paren power
An XOR gate takes two binary inputs and returns 1 if the inputs are different, and 0 if they are the same. This problem is famous in machine learning history because it is non-linearly separable, meaning a single-layer perceptron cannot solve it. It requires at least one hidden layer. Our Excel network will feature the following architecture: 2 nodes ( Hidden Layer: 2 nodes ( ) with a Sigmoid activation function Output Layer: 1 node ( ) with a Sigmoid activation function The Training Data
For a simple demonstration, we will build a network that can learn basic logic (like an XOR gate) or simple regression. : 2 features (e.g., and ). Hidden Layer : 2 neurons ( ). Output Layer : 1 neuron ( ). Activation Function : Sigmoid ( ). 2. Forward Propagation Formulas build neural network with ms excel full
We will build a :
He took a deep breath. He entered the four scenarios of XOR into rows 2 through 5:
Weight_Input1_Hidden1 = Weight_Input1_Hidden1 - Learning Rate * dE/dWeight_Input1_Hidden1 You don’t need Python, TensorFlow, or expensive hardware
In professional settings, you won't use Excel to train large-scale networks. Its strength lies in its ability to . Because every calculation is laid out in plain view, you can:
, to squash our values between 0 and 1. In Excel, use the EXP function: Formula for ah1a sub h 1 end-sub : =1 / (1 + EXP(-z_h1)) Formula for ah2a sub h 2 end-sub : =1 / (1 + EXP(-z_h2)) Step 3.3: Calculate Output Layer Input and Prediction Now, treat the hidden layer activations ( ) as inputs for the final output node: Formula for : =($a_h1*W_o1) + ($a_h2*W_o2) + b_o Formula for Final Prediction ( ): =1 / (1 + EXP(-z_o)) Step 3.4: Calculate Total Error
We need 4 weight matrices and 2 bias vectors. Our Excel network will feature the following architecture:
Now came the magic. Arthur stared at the grid, the cursor blinking in cell H2, his first Hidden Neuron.
After Solver finishes, the "Total Error" should be very low (e.g.,
– in N14 : =J14*$J$4 + K14*$J$5 + L14*$J$6 + M14*$J$7 + $J$10