Seq2seq for Syntactic Parsing (S (NP deep learning (VP is (ADJV very powerful）) Seq2seq! VP NP ADJV deep learning is very powerful
deep learning is very powerful VP S NP ADJV Seq2seq for Syntactic Parsing Seq2seq! (S (NP deep learning ) (VP is (ADJV very powerful ) ) ) 11
Seq2seq for Syntactic Parsing (S (NP deep learning ）(VPis (ADJV very powerful Grammar as a Foreign Language Oriol Vinyals* Lukasz Kaiser* Google Google vinyals@google.com lukaszkaiser@google.com Terry Koo Slav Petrov llya Sutskever Google Google Google terrykoo@google.com slav@google.com ilyasu@google.com Geoffrey Hinton https://arxiv.org Google geoffhinton@google.com /abs/1412.7449 deep earning is very powerful 12
deep learning is very powerful Seq2seq for Syntactic Parsing (S (NP deep learning ) (VP is (ADJV very powerful ) ) ) https://arxiv.org /abs/1412.7449 12
c.f.Multi-class Classification Seq2seg for Multi-label Classification An object can belong to multiple classes. Class 1 Class 1 Class 3 Class 10 Class 3 Class 9 Class 17 Seq2seq Class 9 Class 7 Class 13 https://arxiv.org/abs/1909.03434 https://arxiv.org/abs/1707.05495 13
Seq2seq for Multi-label Classification c.f. Multi-class Classification Seq2seq Class 9 Class 7 Class 13 https://arxiv.org/abs/1909.03434 https://arxiv.org/abs/1707.05495 Class 1 Class 3 Class 1 Class 3 Class 9 Class 10 Class 17 An object can belong to multiple classes. 13
Class Bounding Box Seq2seg for FFN FN Object Detection Decoder https://arxiv.org/abs/2005.12872 Add Norm FFN Encoder Add Norm Add Norm 个 Multi-Head Attention FFN Add Norm Add Norm Multi-Head Self-Attention Multi-Head Self-Attention ▣▣▣回 Image features Spatial positional encoding object queries 14
Seq2seq for Object Detection 14 https://arxiv.org/abs/2005.12872
Output Probabilities ↑ Softmax Seq2seq Linear Add Norm output sequence Feed Forward Add Norm Add Norm Multi-Head Encoder Decoder Feed Attention Forward Nx Nx Add Norm Add Norm Masked Input sequence Multi-Head Multi-Head Attention Attention Positional Positional Encoding ⊕ Encoding Input Output Embedding Embedding <EOS> Inputs Outputs Sequence to Sequence Learning with (shifted right) Neural Networks Transformer https://arxiv.org/abs/1409.3215 https://arxiv.org/abs/1706.03762
Seq2seq Encoder Decoder Input sequence output sequence https://arxiv.org/abs/1706.03762 Sequence to Sequence Learning with Neural Networks https://arxiv.org/abs/1409.3215 Transformer 15