4 Conclusions

This study demonstrates the ecacy of neural implementations of multistage and tree-search quantization. For xed branching ratio we have seen that the tree-search quantizer consistently outperforms the multi-stage structure, though at considerable resource cost. For networks with equal neural resource, the multistage architecture returns signicantly lower mse than the at and tree-search architectures. Finally, experiments show that pattern rescaling oers a degree of noise immunity. It is intriguing to us that a model based on physiology is so closely related to an important signal processing technique and that the biological system has apparently developed an architecture of great eciency.