Attention neural networks with linear units
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for performing a machine learning task on a network input to generate a network output. In one aspect, one of the systems includes an attention neural network configured to perform the machine learning task, the attention neural network including one or more attention layers, each attention layer comprising an attention sub-layer and a feed-forward sub-layer that applies an element-wise multiplication between two vectors generated as a result of two different linear transformations performed on the same attended layer input.
1 . A system comprising one or more computers and one or more storage devices storing instructions that, when executed by the one or more computers, cause the one or more computers to implement:
an attention neural network configured to process a network input to generate a network output for a machine learning task, the attention neural network comprising a plurality of attention layers, each attention layer comprising an attention sub-layer and a feed-forward sub-layer, the attention sub-layer configured to:
receive an input sequence for the attention layer comprising a respective layer input at each of one or more positions; and
generate an attended input sequence at least in part by applying an attention mechanism to the input sequence for the attention layer, the attended input sequence comprising a respective attended layer input at each of the one or more positions, and the feed-forward sub-layer configured to:
receive the attended input sequence; and
generate an output sequence for the attention layer from the attended input sequence, the output sequence comprising a respective layer output at each of the one or more positions, and the generating comprising, for each position of the one or more positions:
generating a first linear transformation output by applying a first linear transformation to the attended layer input at the position;
generating a first transformed input by applying a non-linear activation function to the first linear transformation output;
generating a second transformed input by applying a second linear transformation to the attended layer input at the position;
generating a third transformed input by performing an element-wise multiplication between the first transformed input and the second transformed input; and
generating the layer output at the position based on the third transformed input,
wherein
the first and second linear transformations have been learned during training of the attention neural network to perform the machine learning task,
the training of the attention neural network comprises training the attention neural network on an unsupervised data set through unsupervised learning,
the same first linear transformation is applied to the attended layer inputs at each of the positions in the input sequence,
the same second linear transformation is applied to the attended layer inputs at each of the positions in the input sequence, and
the attention neural network further comprises one or more output layers that are configured to process at least one of the layer outputs in the output sequence generated by one of the plurality of attention layers to generate at least a portion of the network output for the machine learning task.
2 . The system of claim 1 , wherein the non-linear activation function is a Swish activation function.
3 . The system of claim 1 , wherein the non-linear activation function is a ReLu activation function.
4 . The system of claim 1 , wherein the non-linear activation function is a sigmoid activation function.
5 . The system of claim 1 , wherein the non-linear activation function is a Gaussian Error Linear Units activation function.
6 . The system of claim 1 , wherein applying the first linear transformation to the attended layer input at the position comprises:
multiplying the attended layer input by a first weight matrix to generate a first product.
7 . The system of claim 1 , wherein applying the second linear transformation to the attended layer input at the position comprises:
multiplying the attended layer input by a second weight matrix to generate a second product.
8 . The system of claim 7 , wherein weights in the first weight matrix have different values than weights in the second weight matrix.
9 . The system of claim 1 , wherein the attention mechanism comprises a multi-head self-attention mechanism.
10 . The system of claim 1 , wherein the machine learning task comprises one of:
a text generation task where the network output comprises a sequence of text; or
an image generation task where the network output comprises a sequence of intensity values for pixels of an image.
11 . A method performed by one or more computers, the method comprising:
receiving a network input; and
processing the network input using an attention neural network to generate a network output for a machine learning task, the attention neural network comprising a plurality of layers, each layer comprising an attention sub-layer and a feed-forward sub-layer, the attention sub-layer configured to:
receive an input sequence for the attention layer comprising a respective layer input at each of one or more positions; and
generate an attended input sequence at least in part by applying an attention mechanism to the input sequence for the attention layer, the attended input sequence comprising a respective attended layer input at each of the one or more positions, and the feed-forward sub-layer configured to:
receive the attended input sequence; and
generate an output sequence for the attention layer from the attended input sequence, the output sequence comprising a respective layer output at each of the one or more positions, and the generating comprising, for each position of the one or more positions:
generating a first linear transformation output by applying a first linear transformation to the attended layer input at the position;
generating a first transformed input by applying a non-linear activation function to the first linear transformation output;
generating a second transformed input by applying a second linear transformation to the attended layer input at the position;
generating a third transformed input by performing an element-wise multiplication between the first transformed input and the second transformed input; and
generating the layer output at the position based on the third transformed input,
wherein
the first and second linear transformations have been learned during training of the attention neural network to perform the machine learning task,
the training of the attention neural network comprises training the attention neural network on an unsupervised data set through unsupervised learning,
the same first linear transformation is applied to the attended layer inputs at each of the positions in the input sequence,
the same second linear transformation is applied to the attended layer inputs at each of the positions in the input sequence, and
the attention neural network further comprises one or more output layers that are configured to process at least one of the layer outputs in the output sequence generated by one of the plurality of attention layers to generate at least a portion of the network output for the machine learning task.
12 . The method of claim 11 , wherein the non-linear activation function is a Swish activation function.
13 . The method of claim 11 , wherein the non-linear activation function is a ReLu activation function.
14 . The method of claim 11 , wherein the non-linear activation function is a sigmoid activation function.
15 . The method of claim 11 , wherein the non-linear activation function is a Gaussian Error Linear Units activation function.
16 . The method of claim 11 , wherein applying the first linear transformation to the attended layer input at the position comprises:
multiplying the attended layer input by a first weight matrix to generate a first product.
17 . The method of claim 11 , wherein applying the second linear transformation to the attended layer input at the position comprises:
multiplying the attended layer input by a second weight matrix to generate a second product.
18 . The method of claim 17 , wherein weights in the first weight matrix have different values than weights in the second weight matrix.
19 . The method of claim 11 , wherein the attention mechanism comprises a multi-head self-attention mechanism.
20 . The method of claim 11 , wherein the machine learning task comprises one of:
a text generation task where the network output comprises a sequence of text; or
an image generation task where the network output comprises a sequence of intensity values for pixels of an image.
21 . One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to implement:
an attention neural network configured to process a network input to generate a network output for a machine learning task, the attention neural network comprising a plurality of attention layers, each attention layer comprising an attention sub-layer and a feed-forward sub-layer, the attention sub-layer configured to:
receive an input sequence for the attention layer comprising a respective layer input at each of one or more positions; and
generate an attended input sequence at least in part by applying an attention mechanism to the input sequence for the attention layer, the attended input sequence comprising a respective attended layer input at each of the one or more positions, and the feed-forward sub-layer configured to:
receive the attended input sequence; and
generate an output sequence for the attention layer from the attended input sequence, the output sequence comprising a respective layer output at each of the one or more positions, and the generating comprising, for each position of the one or more positions:
generating a first linear transformation output by applying a first linear transformation to the attended layer input at the position;
generating a first transformed input by applying a non-linear activation function to the first linear transformation output;
generating a second transformed input by applying a second linear transformation to the attended layer input at the position;
generating a third transformed input by performing an element-wise multiplication between the first transformed input and the second transformed input; and
generating the layer output at the position based on the third transformed input,
wherein
the first and second linear transformations have been learned during training of the attention neural network to perform the machine learning task,
the training of the attention neural network comprises training the attention neural network on an unsupervised data set through unsupervised learning,
the same first linear transformation is applied to the attended layer inputs at each of the positions in the input sequence,
the same second linear transformation is applied to the attended layer inputs at each of the positions in the input sequence, and
the attention neural network further comprises one or more output layers that are configured to process at least one of the layer outputs in the output sequence generated by one of the plurality of attention layers to generate at least a portion of the network output for the machine learning task.