Path anomaly detection using Hierarchical Temporal Memory
- 1. Anomoly Detection Using Hierarchical Temporal Memory and Grid Cells
- 2. The Problem ? The circle on the right is revolving in a circle. ? Detect any anomaly in the path the circle takes
- 3. Hierarchical Temporal Memory ? For our purpose. ? An algorithm learns to predict xt+1 from xt ? Accepts sparse binary tensors as input ? a.k.a Sparse Distributed Representation (SDR)
- 4. SDR and Encoders ? So, a SDR looks like this
- 5. SDR and Encoders ? So, a SDR looks like this ? How do we construct one
- 6. SDR and Encoders ? So, a SDR looks like this ? How do we convert integer/numbers into one
- 7. ScalarEncoder ? The simplest encoder ? Example: Encode scalars between 0 and 100 with 400 bits. ? SDR on the right represents 54
- 8. ScalarEncoder ? The simplest encoder ? Example: Encode scalars between 0 and 100 with 400 bits. ? SDR on the right represents 54 ? One SDR can represent may values ? What happens if > 100 or < 0 ? Encode as 0 or 100
- 9. ScalarEncoder ? We are NOT using it here
- 10. Grid Cells ? State of the art (for real) ? ScalarEncoder suc*s ? GeoSpatialEncoder is old
- 11. Grid Cells ? A grid cell fires in a grid like pattern ? With multiple grid cell units. A unique location can be represented
- 12. Grid Cells ? Implementation: ? ~50 lines of C++ ? And 2D vector math ? Won't board you with the details ? Encoding a 2D coord using such algorithm
- 13. Grid Cells
- 14. Demo ? The circle is the target ? Location encoded with grid cells ? Shows anomaly signal when moved out of ordinary path