Neural Network Parameters Calculator

This calculator sums the weight matrices of a fully-connected (dense) feedforward network: Input Neurons times Neurons per Hidden Layer for the first connection, Neurons per Hidden Layer squared for each additional hidden-to-hidden connection, and Neurons per Hidden Layer times Output Neurons for the final layer, then adds one bias term per neuron when Bias Terms is set to "With Bias." Setting Hidden Layers to 0 removes both of those hidden-layer weight blocks entirely and connects Input Neurons directly to Output Neurons with a single weight matrix, modeling a network with no hidden layers at all rather than reusing the one-hidden-layer arithmetic. Because the hidden-to-hidden term scales with the square of Neurons per Hidden Layer while every other term scales linearly with its neuron counts, Neurons per Hidden Layer is the input with the most leverage over Total Weights for a typical multi-hidden-layer network -- doubling it can more than double the parameter count once two or more hidden layers are in play. Hidden Layers adds a full square block of weights (Neurons per Hidden Layer squared) for every additional hidden-to-hidden connection, so parameter count grows roughly linearly with layer count once the layer size is fixed.

Memory Required simply multiplies Total Parameters by 4 bytes, assuming 32-bit floating point storage; it does not account for optimizer state (which can multiply working memory by 2-3x during training for algorithms like Adam), activation memory, or lower-precision formats like float16 or int8 quantization, all of which change real deployed or in-training memory substantially from this raw parameter-count figure. The formula also assumes a plain dense architecture -- convolutional, attention, or embedding layers use entirely different parameter-counting rules and are not represented here.