Perhaps quantization with 4 bits is a special case of what I meant. With 4 bits you have 16 different weight values but these values must not be quantized. The

Perhaps quantization with 4 bits is a special case of what I meant. With 4 bits you have 16 different weight values but these values must not be quantized. They could be anything provided gradient descent is possible. Have some combination of binary operations instead of mathematical operations. What these operations do exactly, does not really matter, as long as gradient descent/backpropagation stays possible. In other words, the model will find a way to use the weights correctly.