% dependent" experiments are marked in the data files.  The reported% performance is an average over 10 runs with this single division of the% data set.% % A standard back-propagation network was used for all experiments.  The% network had 60 inputs and 2 output units, one indicating a cylinder and the% other a rock.  Experiments were run with no hidden units (direct% connections from each input to each output) and with a single hidden layer% with 2, 3, 6, 12, or 24 units.  Each network was trained by 300 epochs over% the entire training set.% % The weight-update formulas used in this study were slightly different from% the standard form.  A learning rate of 2.0 and momentum of 0.0 was used.% Errors less than 0.2 were treated as zero.  Initial weights were uniform% random values in the range -0.3 to +0.3.% % RESULTS: % % For the angle independent experiments, Gorman and Sejnowski report the% following results for networks with different numbers of hidden units:% % Hidden	% Right on	Std.	% Right on	Std.% Units	Training set	Dev.	Test Set	Dev.% ------	------------	----	----------	----% 0	89.4		2.1	77.1		8.3% 2	96.5		0.7	81.9		6.2% 3	98.8		0.4	82.0		7.3% 6	99.7		0.2	83.5		5.6% 12	99.8		0.1	84.7		5.7% 24	99.8		0.1	84.5		5.7% % For the angle-dependent experiments Gorman and Sejnowski report the% following results:% % Hidden	% Right on	Std.	% Right on	Std.% Units	Training set	Dev.	Test Set	Dev.% ------	------------	----	----------	----% 0	79.3		3.4	73.1		4.8% 2	96.2		2.2	85.7		6.3% 3	98.1		1.5	87.6		3.0% 6	99.4		0.9	89.3		2.4% 12	99.8		0.6	90.4		1.8% 24     100.0		0.0	89.2		1.4% % Not surprisingly, the network's performance on the test set was somewhat% better when the aspect angles in the training and test sets were balanced.% % Gorman and Sejnowski further report that a nearest neighbor classifier on% the same data gave an 82.7% probability of correct classification.% % Three trained human subjects were each tested on 100 signals, chosen at% random from the set of 208 returns used to create this data set.  Their% responses ranged between 88% and 97% correct.  However, they may have been% using information from the raw sonar signal that is not preserved in the% processed data sets presented here.% % REFERENCES: % % 1. Gorman, R. P., and Sejnowski, T. J. (1988).  "Analysis of Hidden Units% in a Layered Network Trained to Classify Sonar Targets" in Neural Networks,% Vol. 1, pp. 75-89.%%%%% Relabeled values in attribute 'Class'%    From: R                       To: Rock                %    From: M                       To: Mine