# 受限玻茲曼機

## 用途

• 喺向前傳遞途中，個網絡可以俾有關 ${\displaystyle \Pr(\mathbf {h} _{0}|\mathbf {v} _{0})}$ 嘅資訊（由 ${\displaystyle \mathbf {v} _{0}}$${\displaystyle \mathbf {h} _{0}}$概率分佈）；而
• 喺重構途中，個網絡就可以俾有關 ${\displaystyle \Pr(\mathbf {v} _{1}|\mathbf {h} _{0})}$ 嘅資訊（由 ${\displaystyle \mathbf {h} _{0}}$${\displaystyle \mathbf {v} _{1}}$ 嘅概率分佈）。

• 準確噉計 ${\displaystyle \Pr(\mathbf {h} _{0}|\mathbf {v} _{0})}$ 等同「見到呢幅圖，估幅圖入面有邊啲身體部位」；
• 準確噉計 ${\displaystyle \Pr(\mathbf {v} _{1}|\mathbf {h} _{0})}$ 等同「諗到呢啲身體部位，幅圖大致會係點嘅樣」；
• 準確噉計 ${\displaystyle \Pr(\mathbf {\text{new h}} _{0}|\mathbf {h} _{0})}$ 等同「按照幅圖入面有嘅身體部位，估計幅圖係乜嘢動物」（例：如果幅圖有四隻腳，噉嗰隻嘢應該唔會係昆蟲）；
• 準確噉計 ${\displaystyle \Pr(\mathbf {h} _{0}|\mathbf {\text{new h}} _{0})}$ 等同「已知手上有呢種動物，呢種動物有乜身體部位」

－呢一個網絡成功做到分層嘅知識表示嘅效果[3][4]

## 攷

1. Larochelle, H.; Bengio, Y. (2008). Classification using discriminative restricted Boltzmann machines (PDF). Proceedings of the 25th international conference on Machine learning - ICML '08.
2. Restricted Boltzmann Machines - Simplified. Towards Data Science.
3. Bengio, Y. (2009). "Learning Deep Architectures for AI". Foundations and Trends in Machine Learning. 2 (1): 1–127.