Download presentation
Presentation is loading. Please wait.
Published byLilian Atkinson Modified over 9 years ago
1
Practical Heirarchical Temporal Memory for Time Series Prediction
Author: Nicholas Hainsey Faculty Advisor: Dr. C. David Shaffer
2
Heirarchical Temporal Memory
Neural network Created by Jeff Hawkins Designed to mimic the human neocortex Network A B C INPUT Implemented one region of the neocortex Prediction Time
3
Input Encoding 1 A B C 1 01110 01000 01000 01000 01110 Time A B C
1 A B C Time 1 A B C Time
4
Spatial Pooler HTM Region
Input bits
5
Spatial Pooler Proximal dendrite Input bits
6
Spatial Pooler Overlap Score Input bits
7
Spatial Pooler Goal: Each input will activate a small percentage of the columns Similar inputs will activate overlapping sets of columns
8
Temporal Pooler Inactive Active Predicted Getting Started
Dystal Dendrites, Interconnections between active and predicted Make one of the B cells unpredicted then show bursting How does temporal pooler learn Inactive Active Predicted
9
Temporal Pooler Inactive Active Predicted Getting Started
Dystal Dendrites, Interconnections between active and predicted Make one of the B cells unpredicted then show bursting How does temporal pooler learn Inactive Active Predicted
10
HTM Implementations HTMCLA HTM-CLA-Visualizer NuPIC
A C++ implementation based off Numenta’s CLA white paper HTM-CLA-Visualizer Java interface for visualization of HTMs NuPIC Created by Numenta, used in Nustudio
11
Nustudio
12
Nustudio
13
Nustudio Run Simulation Stop
14
Nustudio Connect to server
15
Nustudio Connect to server
16
Nustudio Run simulation from server
17
Nustudio Pos (z): 2 Was Predicted: True Is Active: True Activation Rate: .050 Prediction Rate: .500
18
Nustudio
19
Nustudio P1: Mean P2: Standard Deviation
20
Nustudio
21
Predictions with noise
Learning SD: SD: SD: 20.0
22
Noise with learning No Noise 5.0 SD 10.0 SD 20.0 SD
23
Conclusions Additions to Nustudio Noise Comparison
Constant simulation from file Live simulation from server Adding noise to incoming data Viewing individual regions of the HTM at any step Noise Comparison Still seems stable under varying levels of noise
24
Future Work More robust test of noisy data More customization of noise
More distributions to choose from More control over where noise is applied Ability to export prediction data
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.