Systems and methods for real-time data processing and for emergency planning
Systems and methods are described herein for real-time data processing and for emergency planning. Scenario test data may be collected in real-time based on monitoring local or regional data to ascertain any anomaly phenomenon that may indicate an imminent danger or of concern. A computer-implemented method may include filtering a plurality of different test scenarios to identify a sub-set of test scenarios from the plurality of different test scenarios that may have similar behavior characteristics. A sub-set of test scenarios is provided to a trained neural network to identify one or more sub-set of test scenarios. The one or more identified sub-set of test scenarios may correspond to one or more anomaly test scenarios from the sub-set of test scenarios that is most likely to lead to an undesirable outcome. The neural network may be one of: a conventional neural network and a modular neural network.
1. A modular neural network system for real-time processing of data, comprising:
a plurality of graphical processors, executing stored programming codes stored in a non-transitory computer storage medium to process scenario data, wherein the programming codes are configured as:
a first processing segment and a second processing segment, each of the first and the second processing segment comprising respective plurality of processing bulks, wherein:
a first bulk of the first and the second processing segment is a front portion and a last bulk of the first and the second processing segment is a back portion,
the first bulk has a shallower data processing block depth than subsequent processing bulks, such that processing block depth increases as a flow of the scenario data increases towards the back portion;
each processing bulk of the first and the second processing segment comprises a plurality of processing blocks for processing certain micro tasks;
each processing block having a different kernel size at a same level for capturing features with different data sizes in the scenario data;
the kernels are selected from one of: a convolution kernel, a transpose convolution kernel;
such that the first segment is responsible for learning special tasks, which the learning will be used in the second segment to reconstruct a volumetric input.
2. The modular neural network of claim 1 , wherein the bulks and blocks in the second segment are transpose blocks and transpose bulks.
3. The modular neural network of claim 2 , wherein the bulks in the first segment is connected to corresponding bulks of the second segment.
4. The modular neural network of claim 1 , wherein the modular neural network utilizes a modular topology formed by smaller network nodes each utilizing different techniques to perform filtering the plurality of different test scenarios to identify a sub-set of scenarios from the plurality of different scenarios having similar behavior characteristics.
5. The modular neural network of claim 4 , wherein the techniques used in the modular topology comprises two or more of: highly-clustered non-regular (HCNR) topology, repeated blocks topology, multi-architectural topology,
the repeated blocks topology further comprises a multi-path node, a modular node, a sequential node and a recursive node.
6. The modular neural network of claim 1 , wherein the modular neural network comprises an aggregate network formed by the plurality of processing blocks for processing the micro tasks, wherein the aggregate network having an architecture of a shallow block depth at the front portion, and increases depth in the block as a flow of the scenario test data increases towards the back portion, wherein each block having the different kernel size at a same level for capturing features with different data sizes in the scenario test data.
7. The modular neural network of claim 6 , wherein processing channels increases in each processing block, wherein the plurality of processing blocks comprise 3DConv with variable number of strides, filters, and kernel size depending on where in the architecture the plurality of processing blocks appear.
8. The modular neural network of claim 6 , wherein the processing blocks utilizes Tanh (Hyperbolic Tangent) as an activation function to produce a range of values between −1 and 1 for normalized distance field.
9. The modular neural network of claim 6 , wherein the architecture comprises at least 4 bulks in the first segment, the first bulk includes 16 blocks, each block has 4 filters, stride of size (t:1; w:1; h:1), wherein the kernel size varies from one block to another in order to capture entities with various scale and size.
10. The modular neural network of claim 9 , wherein the kernel size takes on any of following sizes: [t, w, h]→[1, 3, 3], [2, 5, 5], [2, 7, 7], [2, 9, 9], [2, 11, 11].
11. The modular neural network of claim 10 , wherein the kernel size follows a same circular assignment scheme in a previous processing bulk.
12. The modular neural network of claim 9 , wherein all the 4 filters for the last processing bulk are concatenated, and there are no inner connections in a bulk and every block in the processing bulk shares a same input.
13. The modular neural network of claim 6 , wherein the aggregated network being an alternate embodiment of the modular neural network for real-time data processing to detect anomaly scenarios.