Neural Network Approach to Discovering Temporal Correlations

1
Neural Network Approach to Discovering Temporal
Correlations

• S.A.Dolenko, Yu.V.Orlov, I.G.Persiantsev,
  Ju.S.ShugaiScobeltsyn Institute of Nuclear
  Physics,Moscow State UniversityE-mail
  yvo_at_radio-msu.net

2
Statement of the problem

• Discovering causal relationship behavior -
  event - What type of behavior has initiated the
  event? - What phenomenon has initiated the
  event?
• Application - geomagnetic storms
  forecasting SOHO - http//sohowww.nasacom.nasa.g
  ov
• Complexity of the task - What is the delay
  between the event and the moment of its
  initiation? - Can use passive observation only

Objective of the research Development of an
algorithm for discovering temporary correlations
3
Model assumptions

• Data Sequence of scene images
• Scene Set of objects
• Lifetime of objects gtgt Registration rate
• Object Set of features
• Phenomenon Unknown combination of features
• Event - Initiated by unknown phenomenon within
  Initiation duration - Search interval gtgt
  Initiation duration - Limited number of events
  types - Fixed (unknown) delay for a given type
  of event
• Find the most probable phenomenon and delay

4
Scheme of the algorithm
5
Model experiment 1 Single event
6
Model experiment 2 Two events
7
Approaching the Sun...
8
Future development

• NN experts specialization through competition
• Second hierarchical level - NN Supervisor
• Discovering temporal correlations Sun surface -
  Geomagnetic storms- Increasing forecast
  horizon- Improving forecast reliability
• Applications in seismology, medicine, finance,