Multi-attention Network for one Shot learning

Multi-attention Network for one Shot learning

2018-05-15 22:35:50

The contribution of this article is:

1. Indicates that the category label information can be helpful to one shot learning and design a method to excavate the information;

2. Propose a attention network to produce attention maps for creating the image representation of a exemplar image in novel class Ba SED on its class tag.

3. Further propose a multi-attention scheme to enhance the performance of the model;

4. Two new datasets were collected and an evaluation criterion was built.

The flowchart of this article:

Attention Map Generation:

The calculation of attention value in this article is also dependent on the response between Visual feature and Language feature. The approximate process is as follows:

1. First Use Word embedding method, get the category label of the feature C, and then the feature further learning, you can use the LSTM or FC layer, namely:

Of these, both W and B are model parameters that can be learned.

2. After obtaining the hidden state, we multiply it with visual feature and get a response:

3. Normalization of attention value:

4. Multiply attention values and features to get the weighted feature:

Multi-attention mechanism:

The multi-attention mechanism here is an extension of just that mechanism, with different parameters to get attention value at different angles.

---done!

Multi-attention Network for one Shot learning