Neural network having input and hidden layers of equal units

1. An electronic device comprising:

a processor,

at least one input interface, and

an artificial neural network, comprising an input layer, an output layer and at least first and second hidden layers, wherein a number of units of the second hidden layer is equal to a number of units of the input layer;

wherein the processor is configured to generate one or more predicted next items in a sequence of items based on an input sequence item received at the at least one input interface by:

retrieving a context vector corresponding to the input sequence item from an encoding matrix configured to map each of a plurality of input sequence items into a discrete form processable by the artificial neural network;

processing the context vector with the artificial neural network;

generating an output vector at least by transforming the output of the second hidden layer of the artificial neural network using the encoding matrix, wherein the output vector corresponds to a probability distribution for one or more predicted next items in the sequence of items, each of the predicted next items included in the plurality of input sequence items for the encoding matrix; and

outputting the output vector.

2. The electronic device of claim 1 wherein the encoding matrix includes either values of parameters of a character-compositional model which is a predictor configured to compute a location of an item in a multi-dimensional space from individual characters of the item, or the encoding matrix includes item embeddings being locations of items in the multi-dimensional space.

3. The electronic device of claim 2 where the encoding matrix is held in a cache and wherein the processor is configured to compute an item embedding corresponding to the input sequence item from the character-compositional model and to add the item embedding to the cache.

4. The electronic device of claim 1 where the processor is configured to retrieve the context vector corresponding to the input sequence by accessing an item embedding from a cache.

5. The electronic device of claim 1 , wherein the processor is configured to process the context vector with the artificial neural network by:

providing the context vector to the input layer of the artificial neural network;

multiplying the contents of the input layer with a first weight matrix W 0 to generate a first result, and providing the first result to the first hidden layer of the artificial neural network;

processing the input to the first hidden layer with nodes of the first hidden layer to produce an output of the first hidden layer;

multiplying the output of the first hidden layer with a second weight matrix W 1 to generate a second result, and providing the second result to the second hidden layer of the artificial neural network; and

processing the input to the second hidden layer with nodes of the second hidden layer to produce an output of the second hidden layer.

6. The electronic device of claim 5 , wherein the artificial neural network further comprises a recurrent hidden vector.

7. The electronic device of claim 6 , wherein the processor is configured to concatenate the contents of the input layer with the recurrent hidden vector prior to processing the contents of the input layer with the hidden layers of the artificial neural network.

8. The electronic device of claim 6 , wherein the recurrent hidden vector comprises data indicative of a previous state of the artificial neural network.

9. The electronic device of claim 8 , wherein the processor is configured to update the recurrent hidden vector according to the output of the first hidden layer.

10. The electronic device of claim 9 , wherein the processor is configured to update the recurrent hidden vector by replacing the recurrent hidden vector with the output of the first hidden layer.

11. The electronic device of claim 1 , wherein the processor is configured to retrieve the context vector by retrieving a row or column of the encoding matrix that corresponds to the input received at the at least one input interface.

12. The electronic device of claim 11 , wherein the processor is configured to produce a 1-of-N vector corresponding to the input sequence item and to retrieve the context vector by transforming the 1-of-N vector using the encoding matrix.

13. The electronic device of claim 12 , wherein transforming the 1-of-N vector comprises multiplying the 1-of-N vector and the encoding matrix using matrix multiplication.

14. The electronic device of claim 1 , wherein the processor is further configured to retrieve an output class prediction from the output layer of the artificial neural network, wherein the output class prediction defines a group of one or more sequence items.

15. A computer-implemented method for generating one or more predicted next items in a sequence of items based on an input sequence item, the method comprising:

receiving, at an electronic device, the input sequence item;

retrieving, from an encoding matrix configured to map each of a plurality of input sequence items into a discrete form processable by an artificial neural network, a context vector corresponding to the input sequence item;

processing the context vector with the artificial neural network, wherein the artificial neural network comprises an input layer, a first hidden layer, a second hidden layer and an output layer;

generating an output vector at least by transforming the output of the second hidden layer of the artificial neural network using the encoding matrix, wherein the output vector corresponds to a probability distribution for one or more predicted next items in the sequence of items, each of the predicted next items included in the plurality of input sequence items for the encoding matrix; and

outputting the output vector,

wherein a number of units of the second hidden layer of the artificial neural network is equal to a number of units of the input layer of the artificial neural network.