While training the model, bias will be trained accordingly. Suppose ideal bias value is +x, if you are subtracting the bias instead of +x bias value will become -x, so overall w -(-x) will become w+x and vice versa.