Story of the Week  

Neural Networks: Do they make sense?


Artificial Neural Networks (ANNs) have established themselves as powerful tools for forecasting and modeling complex systems in engineering and sciences. However, one of the major criticisms of ANNs is their being black box models, which do not consider the underlying physics of the problem. This criticism mainly stems from the fact that no satisfactory explanation of their internal behavior has been offered yet. However, if the nonlinear function being mapped by the ANN is explored further, one may be able to shed some light into the physical processes of the system being modeled that are inherent in a trained ANN. It is quite possible that while training of an ANN, different hidden neurons do learn the individual relationships inherent in different components of a physical process. However, identifying components of a physical process in a trained ANN is an area of research that is in its infancy and requires attention of researchers.

The researchers working in the area of hydrology and water resources engineering at IIT Kanpur are trying to unearth the physical processes hidden inside the trained ANN rainfall-runoff models. The rainfall-runoff process is an extremely complex, dynamic, non-linear, and fragmented physical process that is not clearly understood and is very difficult to model. In order to investigate whether a trained ANN model makes physical sense or not, the daily average rainfall (mm) and runoff (cubic meters per second) data taken from a real watershed were employed to first develop an ANN model for the rainfall-runoff process. The developed ANN model was then explored at the hidden nodes and the outputs from the hidden nodes were compared with some of the components of the conceptual rainfall-runoff model developed on the same data set. The results obtained suggest that the distributed structure of an ANN is able to capture certain physical behavior of the rainfall runoff process during training. It was found that the various hidden nodes in the trained ANN rainfall-runoff model were capable of modeling various components of the hydrologic system such as infiltration, base flow, and delayed and quick surface flow, and represent the rising limb and different portions of the falling limb of a flow hydrograph. These findings are indeed encouraging and demonstrate that the trained ANN models do make sense, if explored properly. A continued research effort in a variety of areas in engineering and sciences is needed to understand the internal workings of the trained ANN models to remove the stigma of them being black-box models.

For more details of the above study, please contact:

Dr. Ashu Jain
Associate Professor
Department of Civil Engineering
Indian Institute of Technology Kanpur

Previous Story

Neural Networks: Do they make sense?

Magnetorheological Abrasive Flow Finishing (MRAFF)

Fabrication of Photonic Devices

Atomic and molecular clusters: designer materials for the nanoworld

Economical Road Design

Molecular Starbursts

2004 Sumatra Earthquake and Tsunami

The Flow Induced Vibrations

Novel Ceramics

The Complexity of Some Isomorphism Problems


Ordered Peptide Aggregation

Asymmetric Synthesis