Answering cognitive queries from sensor input signals
View Patent ↗A computer-implemented method for answering a cognitive query from sensor input signals may be provided. The method comprises feeding sensor input signals to an input layer of an artificial neural network comprising a plurality of hidden neuron layers and an output neural layer, determining hidden layer output signals from each of the plurality of hidden neuron layers and output signals from the output neural layer, and generating a set of pseudo-random bit sequences by applying a set of mapping functions using the output signals of the output layer and the hidden layer output signals of one of the hidden neuron layers as input data for one mapping function. Furthermore, the method comprises determining a hyper-vector using the set of pseudo-random bit sequences, and storing the hyper-vector in an associative memory, in which a distance between different hyper-vectors is determinable.
1. A method for answering a cognitive query from sensor input signals, said method comprising:
feeding sensor input signals to an input layer of an artificial neural network comprising a plurality of hidden neuron layers and an output neural layer;
determining hidden layer output signals from each of said plurality of hidden neuron layers and output signals from said output neural layer;
generating a set of pseudo-random bit sequences by applying a plurality of mapping functions by at least feeding said output signals of said output layer and feeding said hidden layer output signals of at least one of said hidden neuron layers directly from said at least one of said hidden neuron layers as input data to one mapping function, the one mapping function receiving the output signals of the output layer and the hidden layer output signals of at least one of the hidden neuron layers additional to receiving the output signals of the output layer, the at least one of the hidden neuron layers including at least one hidden neuron layer other than a hidden neuron layer immediately preceding the output layer, wherein the one mapping function generates at least one pseudo-random bit sequence of the set of pseudo-random bit sequences based on the received output signals of said output layer and intermediate information of the artificial neural network represented by the hidden layer output signals received additionally to the output signals of said output layer, each of the plurality of mapping functions being fed as input, at least one selected from the group consisting of: signals received directly from at least two of the plurality of hidden neuron, and signals received directly from the output neural layer and the at least one of the plurality of hidden neuron layers, wherein each one of the plurality of hidden neuron layers is used in direct feeding in generating the set of pseudo-random bit sequences;
determining a hyper-vector using said set of pseudo-random bit sequences by concatenating the set of pseudo-random bit sequences generated by the plurality of mapping functions; and
storing said hyper-vector in an associative memory, in which a distance between different hyper-vectors is determinable.
2. The method according to claim 1 , also comprising
combining different hyper-vectors stored in said associative memory for deriving said cognitive query and candidate answers, each of which is obtained from said sensor input signals.
3. The method according to claim 2 , also comprising
measuring a distance between said hyper-vector representing said cognitive query and said hyper-vectors representing said candidate answers.
4. The method according to claim 3 , also comprising
selecting said hyper-vector relating to said candidate answers having a minimum distance from said hyper-vector representing said cognitive query.
5. The method according to claim 2 , wherein said feeding sensor input data to an input layer of an artificial neural network also comprises
feeding said sensor input data to input layers of a plurality of artificial neural networks.
6. The method according to claim 1 , wherein said generating a set of pseudo-random bit sequences also comprises
using said output signals of said output layer and said hidden layer output signals of a plurality of said hidden neuron layers as input data for one mapping function.
7. The method according to claim 1 , wherein said neural network is a convolutional neural network.
8. The method according to claim 1 , also comprising
generating a plurality of random hyper-vectors as role vectors at an end of a training of said neural network to be stored in said associative memory.
9. The method according to claim 5 , also comprising
generating continuously a plurality of pseudo-random hyper-vectors as filler vectors based on output data of said plurality of neural networks for each new set of sensor input data, said pseudo-random hyper-vectors to be stored in said associative memory.
10. The method according to claim 2 , wherein said combining comprises
binding different hyper-vectors by a vector-element-wise binary XOR operation.
11. The method according to claim 2 , wherein said combining comprises
bundling different hyper-vectors by a vector-element-wise binary average.
12. The method according to claim 8 , wherein said role vectors are stored in said associative memory at said end of a training of said artificial neural network if a stochastic gradient descent value determined between two sets of input vectors of said sensor input signals to said neural network remain below a predefined threshold value.
13. A machine-learning system for answering a cognitive query from sensor input signals, said machine-learning system comprising:
a sensor adapted for feeding sensor input signals to an input layer of an artificial neural network comprising a plurality of hidden neuron layers and an output neural layer;
at least one computer processor adapted for determining hidden layer output signals from each of said plurality of hidden neuron layers and output signals from said output neural layer;
the at least one computer processor further for generating a set of pseudo-random bit sequences by applying a plurality of mapping functions by at least feeding said output signals of said output layer and feeding said hidden layer output signals of at least one of said hidden neuron layers directly from said at least one of said hidden neuron layers as input data to one mapping function, the one mapping function receiving the output signals of the output layer and the hidden layer output signals of at least one of the hidden neuron layers additional to receiving the output signals of the output layer, the at least one of the hidden neuron layers including at least one hidden neuron layer other than a hidden neuron layer immediately preceding the output layer, wherein the one mapping function generates at least one pseudo-random bit sequence of the set of pseudo-random bit sequences based on the received output signals of said output layer and intermediate information of the artificial neural network represented by the hidden layer output signals received additionally to the output signals of said output layer, each of the plurality of mapping functions being fed as input, at least one selected from the group consisting of: signals received directly from at least two of the plurality of hidden neuron, and signals received directly from the output neural layer and the at least one of the plurality of hidden neuron layers, wherein each one of the plurality of hidden neuron layers is used in direct feeding in generating the set of pseudo-random bit sequences;
the at least one computer processor further adapted for determining a hyper-vector using said set of pseudo-random bit sequences by concatenating the set of pseudo-random bit sequences generated by the plurality of mapping functions; and
a storage module adapted for storing said hyper-vector in an associative memory, in which a distance between different hyper-vectors is determinable.
14. The machine-learning system according to claim 13 , wherein the at least one computer processor is further
for combining different hyper-vectors stored in said associative memory for deriving said cognitive query and candidate answers, each of which is obtained from said sensor input signals.
15. The machine-learning system according to claim 14 , wherein the at least one computer processor is further
adapted for measuring a distance between said hyper-vector representing said cognitive query and said hyper-vectors representing said candidate answers.
16. The machine-learning system according to claim 15 , wherein the at least one computer processor is further
adapted for selecting said hyper-vector relating to said candidate answers having a minimum distance from said hyper-vector representing said cognitive query.
17. The machine-learning system according to claim 14 , wherein said sensor for feeding sensor input data to an input layer of an artificial neural network is also adapted for
feeding said sensor input data to input layers of a plurality of artificial neural networks.
18. The machine-learning system according to claim 13 , wherein said generating a set of pseudo-random bit sequences also comprises
using said output signals of said output layer and said hidden layer output signals of a plurality of said hidden neuron layers as input data for one mapping function.
19. The machine-learning system according to claim 13 , wherein said neural network is a convolutional neural network.
20. The machine-learning system according to claim 13 , wherein the at least one computer processor is further
adapted for generating a plurality of random hyper-vectors as role vectors at an end of a training of said neural network to be stored in said associative memory.
21. The machine-learning system according to claim 17 , wherein the at least one computer processor is further
adapted generating continuously a plurality of pseudo-random hyper-vectors as filler vectors based on output data of said plurality of neural networks for each new set of sensor input data, said pseudo-random hyper-vectors to be stored in said associative memory.
22. The machine-learning system according to claim 14 , wherein the at least one computer processor is further adapted for
binding different hyper-vectors by a vector-element-wise binary XOR operation.
23. The machine-learning system according to claim 14 , wherein the at least one computer processor is further adapted for
bundling different hyper-vectors by a vector-element-wise binary average.
24. The machine-learning system according to claim 20 , wherein said role vectors are stored in said associative memory at said end of a training of said artificial neural network if a stochastic gradient descent value determined between two sets of input vectors of said sensor input signals to said neural network remain below a predefined threshold value.
25. A computer program product for answering a cognitive query from sensor input signals, said computer program product comprising a computer readable storage medium having program instructions embodied therewith, said program instructions being executable by a computing system to cause the computing system to:
feed sensor input signals to an input layer of an artificial neural network comprising a plurality of hidden neuron layers and an output neural layer;
determine hidden layer output signals from each of said plurality of hidden neuron layers and output signals from said output neural layer;
generate a set of pseudo-random bit sequences by applying a plurality of mapping functions by at least feeding said output signals of said output layer and feeding said hidden layer output signals of at least one of said hidden neuron layers directly from said at least one of said hidden neuron layers as input data to one mapping function, the one mapping function receiving the output signals of the output layer and the hidden layer output signals of at least one of the hidden neuron layers additional to receiving the output signals of the output layer, the at least one of the hidden neuron layers including at least one hidden neuron layer other than a hidden neuron layer immediately preceding the output layer, wherein the one mapping function generates at least one pseudo-random bit sequence of the set of pseudo-random bit sequences based on the received output signals of said output layer and intermediate information of the artificial neural network represented by the hidden layer output signals received additionally to the output signals of said output layer, each of the plurality of mapping functions being fed as input, at least one selected from the group consisting of: signals received directly from at least two of the plurality of hidden neuron, and signals received directly from the output neural layer and the at least one of the plurality of hidden neuron layers, wherein each one of the plurality of hidden neuron layers is used in direct feeding in generating the set of pseudo-random bit sequences;
determine a hyper-vector using said set of pseudo-random bit sequences by concatenating the set of pseudo-random bit sequences generated by the plurality of mapping functions; and
store said hyper-vector in an associative memory, in which a distance between different hyper-vectors is determinable.