Adaptive attention model for image captioning
The technology disclosed presents a novel spatial attention model that uses current hidden state information of a decoder long short-term memory (LSTM) to guide attention and to extract spatial image features for use in image captioning. The technology disclosed also presents a novel adaptive attention model for image captioning that mixes visual information from a convolutional neural network (CNN) and linguistic information from an LSTM. At each timestep, the adaptive attention model automatically decides how heavily to rely on the image, as opposed to the linguistic model, to emit the next caption word. The technology disclosed further adds a new auxiliary sentinel gate to an LSTM architecture and produces a sentinel LSTM (Sn-LSTM). The sentinel gate produces a visual sentinel at each timestep, which is an additional representation, derived from the LSTM's memory, of long and short term visual and linguistic information.
1. A system comprising:
an encoder for processing an input image to generate encoded image features;
a decoder for processing a previously emitted caption word combined with the encoded image features to produce, at each decoder iteration, a current hidden state of the decoder and a visual sentinel;
an adaptive attender for:
attending to the encoded image features at each decoder iteration to produce an image context conditioned on the current hidden state of the decoder; and
mixing the image context and the visual sentinel to produce an adaptive context at each decoder iteration; and
an emitter for generating a natural language caption for the input image based on the adaptive contexts produced over successive decoder iterations.
2. The system of claim 1 , wherein the encoder comprises a convolutional neural network (CNN).
3. The system of claim 1 , wherein the decoder comprises a long short-term memory network (LSTM).
4. The system of claim 1 , wherein the adaptive attender enhances attention directed to the image context when a next caption word is a visual word.
5. The system of claim 1 , wherein the adaptive attender enhances attention directed to the visual sentinel when a next caption word is a non-visual word or linguistically correlated to the previously emitted caption word.
6. The system of claim 1 , further configured to prevent, during training, backpropagation of gradients from the decoder to the encoder when a next caption word is a non-visual word or linguistically correlated to the previously emitted caption word.
7. The system of claim 1 , wherein the visual sentinel includes visual context determined from previously processed image features and textual context determined from previously emitted caption words.
8. A method comprising:
processing an input image with an encoder to generate encoded image features;
processing with a decoder a previously emitted caption word combined with the encoded image features to produce, at each decoder iteration, a current hidden state of the decoder and a visual sentinel;
attending to the encoded image features at each decoder iteration to produce an image context conditioned on the current hidden state of the decoder; and
mixing the image context and the visual sentinel to produce an adaptive context at each decoder iteration; and
generating a natural language caption for the input image with an emitter based on the adaptive contexts produced over successive decoder iterations.
9. The method of claim 8 , wherein the encoder comprises a convolutional neural network (CNN).
10. The system of claim 8 , wherein the decoder comprises a long short-term memory network (LSTM).
11. The method of claim 8 , comprising enhancing attention directed to the image context when a next caption word is a visual word.
12. The method of claim 8 , comprising enhancing attention directed to the visual sentinel when a next caption word is a non-visual word or linguistically correlated to the previously emitted caption word.
13. The method of claim 8 , comprising preventing, during training, backpropagation of gradients from the decoder to the encoder when a next caption word is a non-visual word or linguistically correlated to the previously emitted caption word.
14. The method of claim 8 , wherein the visual sentinel includes visual context determined from previously processed image features and textual context determined from previously emitted caption words.
15. A non-transitory machine-readable medium comprising a plurality of machine-readable instructions which, when executed by one or more processors, are adapted to cause the one or more processors to perform a method comprising:
processing an input image with an encoder to generate encoded image features;
processing with a decoder a previously emitted caption word combined with the encoded image features to produce, at each decoder iteration, a current hidden state of the decoder and a visual sentinel;
attending to the encoded image features at each decoder iteration to produce an image context conditioned on the current hidden state of the decoder; and
mixing the image context and the visual sentinel to produce an adaptive context at each decoder iteration; and
generating a natural language caption for the input image with an emitter based on the adaptive contexts produced over successive decoder iterations.
16. The non-transitory machine-readable medium of claim 15 , wherein the encoder comprises a convolutional neural network (CNN).
17. The non-transitory machine-readable medium of claim 15 , wherein the decoder comprises a long short-term memory network (LSTM).
18. The non-transitory machine-readable medium of claim 15 , comprising enhancing attention directed to the image context when a next caption word is a visual word.
19. The non-transitory machine-readable medium of claim 15 , comprising enhancing attention directed to the visual sentinel when a next caption word is a non-visual word or linguistically correlated to the previously emitted caption word.
20. The non-transitory machine-readable medium of claim 15 , comprising preventing, during training, backpropagation of gradients from the decoder to the encoder when a next caption word is a non-visual word or linguistically correlated to the previously emitted caption word.
21. The non-transitory machine-readable medium of claim 15 , wherein the visual sentinel includes visual context determined from previously processed image features and textual context determined from previously emitted caption words.