I knew the formula but fumbled explaining it cleanly under pressure.
Start by defining the variables and explaining how each affects the parameter count. Then derive the formula step-by-step, clearly separating the weight and bias contributions, and finally present the total with and without bias.
Pro tip: Mention that the number of parameters is independent of the input spatial dimensions and stride/padding, which only affect the output size. This shows you understand the distinction between parameter count and computational cost.
Clearly state the input channels (C_in), output channels (C_out), kernel height (K_h), kernel width (K_w), and whether bias is used.
Each output channel has a kernel of size K_h x K_w for each input channel, so weights = C_out * C_in * K_h * K_w.
If bias is used, add one bias parameter per output channel, so bias parameters = C_out.
Total with bias = weights + bias; without bias = weights only. Present both formulas.
Briefly mention how parameter count scales with channels and kernel size, and note that stride/padding do not affect it.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.