CS224N Lecture3 筆記
02-12
主要內容：
Why do we need approximations of the original skip-gram formulation?
How is the problem mitigated by using negative sampling?
How is CBOW different from skip-gram?
What are the limitations of using SVD on the co-occurrence matrix to
get word vectors?How does GloVe combine the advantages of count-based models and
predictive models?How to evaluate word vectors? What are the most commonly used tasks
for intrinsic evaluation?Which factors could affect the quality of learned vectors?
推薦閱讀：
※【Neural Networks and Deep Learning】3.改進神經網路的學習方法（過擬合）
※使用py-faster-rcnn進行目標檢測（object detect）
※【Neural Networks and Deep Learning】4.神經網路可以計算任何函數的可視化證明(1)
※Inception-v2/v3結構解析（原創）