Baidu AI participated in two days of training camp, I learned something about the simple record
A knowledge map
Knowledge Mapping is to make the machine has the ability to accumulate knowledge, application of knowledge.
Due to the current large amount of knowledge, so artificial mark can not be met, you can use data-driven, bottom-up way to automatically build a common knowledge map.
Industry knowledge map can not be constructed using the method described above, we also need to build human-computer expert.
Second, the voice
Near-field identification (one meter) has been very accurate, manufacturers often claim 98.7%, can be improved, but not necessary.
Far-field support equipment required, limit 10 meters.
TTS costs by Deep learning ways to build sound library.
Voiceprint exists drift, young and old points more reliable.
English proficiency test currently available machines discrimination
Three, NLP
Semantic representation: vector, CBOW simple and effective method, is represented by the surrounding words
Baidu unique chapter in knowledge acquisition
Chapter said: multimodal integration of knowledge to understand, for example, video tagging who said what at what time content, video is automatically label