Making sure we're alert at the wheel
Neural networks are scientists' attempt to model the human brain in the hopes of achieving artificial intelligence (AI). Neural networks are used to model complex input-output relationships. Like humans, neural networks acquire knowledge by learning and store this knowledge in the form of neuron weighting factors. The Neural Networks Laboratory at the University of Milan in Italy is employing neural networks with the goal of reducing automobile accident rates. Continuous monitoring of drivers' electrocardiograms, respiration, galvanic skin response and skin temperature provided data regarding their level of alertness. On the other hand, automobile parameters such as speed, acceleration, steering angle and so on were also recorded. These two streams of data provided the input to a hybrid Multi-Layer Perceptron (MLP) neural network. Since neural networks must be trained, the next phase comprised intensive training. The back-propagation technique was employed wherein input is fed to the MLP and the weighting factors are adjusted until the desired output is achieved. Application of Probably Approximately Correct (PAC) training theory helped define the necessary amount of training. When training was complete, rules were established to govern the interplay between sub-symbolic and symbolic entities in the MLP. The trade-off between computing time and performance was intelligently balanced though the application of a simulated annealing optimisation scheme. With respect to the road, the procedures requiring the driver's utmost attention had to be defined using the parameters available. Bringing it all together, the system monitors the driver's health status during execution of such procedures and based on the output of the network provides valuable visual feedback. The system can only be classified as a prototype at this stage. Partnerships with organisations with complementary skills are necessary in order to move forward.
