Deep neural networks with interpretability
View Patent ↗A feature sub-network trainer improves robustness of interpretability of a deep neural network (DNN) by increasing the likelihood that the DNN will converge to a global minimum of a cost function of the DNN. After determining a plurality of correctly classified examples of a pre-trained DNN, the trainer extracts from the pre-trained DNN a feature sub-network that includes an input layer of the DNN and one or more subsequent sparsely-connected layers of the DNN. The trainer averages output signals from the sub-network to form an average representation of each class identifiable by the DNN. The trainer relabels each correctly classified example with the appropriate average representation, and then trains the feature sub-network with the relabeled examples. In one demonstration, the feature sub-network trainer improved classification accuracy of a seven-layer convolutional neural network, trained with two thousand examples, from 75% to 83% by reusing the training examples.
1. A deep neural network classification method, comprising:
determining a plurality of correctly classified examples of a pre-trained neural network;
extracting, from the pre-trained neural network, a feature sub-network having a plurality of sub-network output nodes;
averaging, for each class of a plurality of classes identifiable by the pre-trained neural network, a plurality of output signals from each of the plurality of sub-network output nodes to form one of a plurality of averaged signals that are in one-to-one correspondence with the plurality of sub-network output nodes, the plurality of averaged signals forming an average representation of said each class, each of the plurality of output signals being generated by inputting one of the plurality of correctly classified examples of said each class into the feature sub-network;
forming a plurality of relabeled examples by relabeling each correctly classified example, of the plurality of correctly classified examples, with the average representation of the class to which said each correctly classified example belongs; and
training the feature sub-network with the plurality of relabeled examples.
2. The method of claim 1 , further comprising joining the feature sub-network, after said training, with a remaining dense sub-network of the pre-trained neural network.
3. The method of claim 1 , wherein said determining includes:
inputting a plurality of examples to the pre-trained neural network; and
comparing, for each example of the plurality of examples, a corresponding output of the pre-trained neural network to a label of said each example.
4. The method of claim 3 , further comprising generating the pre-trained neural network by pre-training an untrained neural network with the plurality of examples.
5. The method of claim 4 , wherein said pre-training the untrained neural network includes pre-training the untrained neural network to a selected level of classification accuracy.
6. The method of claim 1 , wherein said training the feature sub-network includes training the feature sub-network via backpropagation.
7. The method of claim 1 , wherein:
the pre-trained neural network includes one or more hidden layers of a plurality of layers; and
extracting the feature sub-network includes selecting one of the one or more hidden layers as an output layer of the feature sub-network.
8. The method of claim 7 , at least one of a plurality of layers of the feature sub-network being a sparsely connected layer.
9. The method of claim 7 , at least one of a plurality of layers of the feature sub-network being a convolutional layer.
10. The method of claim 7 , wherein:
the layers of the pre-trained neural network following the output layer of the feature sub-network form a dense sub-network; and
at least one of the layers of the dense sub-network is a fully connected layer.
11. A system, comprising:
a processor;
a memory in electronic communication with the processor;
and a sub-network training engine, implemented as machine-readable instructions that are stored in the memory and, when executed by the processor, control the system to:
determine a plurality of correctly classified examples of the pre-trained neural network,
extract, from the pre-trained neural network, a feature sub-network having a plurality of sub-network output nodes,
average, for each class of a plurality of classes identifiable by the pre-trained neural network, a plurality of output signals from each of the plurality of sub-network output nodes to form one of a plurality of averaged signals that are in one-to-one correspondence with the plurality of sub-network output nodes, the plurality of averaged signals forming an average representation of said each class, each of the plurality of output signals being generated by inputting one of the plurality of correctly classified examples of said each class into the feature sub-network,
form a plurality of relabeled examples by relabeling each correctly classified example, of the plurality of correctly classified examples, with the average representation of the class to which said each correctly classified example belongs, and
train the feature sub-network with the plurality of relabeled examples.
12. The system of claim 11 , the sub-network training engine including additional machine-readable instructions that, when executed by the processor, control the system to join the feature sub-network with a remaining dense sub-network of the pre-trained neural network.
13. The system of claim 11 , wherein the machine-readable instructions that, when executed by the processor, control the system to determine the correctly classified examples include machine-readable instructions that, when executed by the processor, control the system to:
input a plurality of examples to the pre-trained neural network, and
compare, for each example of the plurality of examples, a corresponding output of the pre-trained neural network to a label of said each example.
14. The system of claim 13 , the sub-network training engine including additional machine-readable instructions that, when executed by the processor, control the system to generate the pre-trained neural network by pre-training an untrained neural network with the plurality of examples.
15. The system of claim 14 , wherein the machine-readable instructions that, when executed by the processor, control the system to pre-train the untrained neural network include machine-readable instructions that, when executed by the processor, control the system to pre-train the untrained neural network to a selected level of training accuracy.
16. The system of claim 11 , wherein the machine-readable instructions that, when executed by the processor, control the system to train the feature sub-network include machine-readable instructions that, when executed by the processor, control the system to train the feature sub-network via backpropagation.
17. The system of claim 11 , wherein:
the pre-trained neural network includes one or more hidden layers of a plurality of layers; and
the machine-readable instructions that, when executed by the processor, control the system to extract the feature sub-network include machine-readable instructions that, when executed by the processor, control the system to select one of the one or more hidden layers as an output layer of the feature sub-network.
18. The system of claim 17 , at least one of a plurality of layers of the feature sub-network being a sparsely connected layer.
19. The system of claim 17 , at least one of a plurality of layers of the feature sub-network being a convolutional layer.
20. The system of claim 17 , wherein:
the layers of the pre-trained neural network following the output layer of the feature sub-network form a dense sub-network; and
at least one of the layers of the dense sub-network is a fully connected layer.