Hierarchical Temporal Memories

We call Hierarchical Temporal Memories (HTMs) a machine learning approach, whose key characteristic is that it captures “causes” that lead to certain sensory input, and organizes them in a hierarchical fashion. As an intuitive example, frequently used by the human perceptive mechanisms: a car in a visual sensor is a high level cause, which can be decomposed in low level causes such as wheels, engine, steering wheel, etc. The proposed method of representation is domain independent, with domain knowledge addable in the supervised learning mode, if necessary.

The purpose of the project is to:

• design an application domain fitting the data gathered by wireless sensor networks in environmental monitoring applications
• design the appropriate HTM structure (as the student will see, or already knows from other machine learning techniques, proper choice of parameters can influence the performance of the modeling aproach)
• train the HTM structure with the vast amount of measurement data we have from past monitoring applications
• usage of the HTM structure for pattern (i.e event) recognition applications which have direct benefit in recognizing natural phenomena ahead of time
• evaluation of effectiveness and performance of approach

Benefits

• usage of a novel (but still controversial) technique in machine learning, a technique based on biological knowledge about how the human brain works
• knowledge augmentation in the areas of data mining, modeling, machine learning, sensor networks, environmental engineering

Prerequisites

• Basic skills in Python
• Interest to experiment, analyze and improve machine learning algorithms