Summary: (sat jan 23 1999 18 15 34) Exp1: predict voxels in left broca(t+1) from left broca (t), use 6hidden units, 15k epochs (sun jan 24 1999 10 17 32) Exp2: predict voxels in left broca(t+1) from left broca (t), use 8 hidden units, 55k epochs Conclusion: Both networks predict better than random. Have average RMS error per prediction of 35.5 and 35.1 respectively, compared to RMS error per prediction of 47.3 using mean value of each voxel to predict and an RMS error per prediction of 44.68 using v_i(t+1) <-- v_i(t) ;;;----------------------------------- **Exp1** (sat jan 23 1999 18 15 34) ((attempting to learn output from (input) using 555 examples plus 238 held out) (network has 6 hidden units training for 15000 epochs with batch? 1 var-momentum? 1 learning.rate 0.01 momentum 0.7)) ((learned network stored on "/afs/cs/project/theo-35/fmri/backprop/train-1-23-99-6h") (training and holdout errors (9551880 0.0012710507 0.001280642 14850))) Performance (first the network, then using mean voxel value as prediction): Network: Predictions: 238 (holdout examples); RMS error per example for each output unit: (32.06 30.77 32.46 28.80 32.08 31.71 38.89 33.36 39.20 36.70 42.91 32.37 38.02 41.83 41.51 ) Average RMS error per example: 35.51 Mean voxel values as predictions: Predictions: 238; RMS error per example for each output unit: (37.75 37.92 45.21 36.05 35.48 38.16 51.80 45.45 57.66 44.69 67.98 43.86 55.17 68.80 52.28 ) Average RMS error per example: 47.88 Sorted mean-slopes based on 238 examples, along with (min-slopes max-slopes) highest magnitude derivatives: dOUT(11)/dIN(11) 0.24 (min-slopes: 0.22) (max-slopes: 0.25) dOUT(11)/dIN(14) 0.23 (min-slopes: 0.21) (max-slopes: 0.24) dOUT(14)/dIN(11) 0.23 (min-slopes: 0.20) (max-slopes: 0.24) dOUT(14)/dIN(14) 0.22 (min-slopes: 0.19) (max-slopes: 0.23) dOUT(14)/dIN(15) 0.19 (min-slopes: 0.16) (max-slopes: 0.20) dOUT(11)/dIN(15) 0.19 (min-slopes: 0.17) (max-slopes: 0.20) dOUT(9)/dIN(14) 0.16 (min-slopes: 0.15) (max-slopes: 0.17) dOUT(15)/dIN(11) 0.16 (min-slopes: 0.13) (max-slopes: 0.18) dOUT(15)/dIN(14) 0.15 (min-slopes: 0.12) (max-slopes: 0.17) dOUT(9)/dIN(11) 0.15 (min-slopes: 0.13) (max-slopes: 0.16) dOUT(13)/dIN(11) 0.15 (min-slopes: 0.12) (max-slopes: 0.17) dOUT(13)/dIN(14) 0.15 (min-slopes: 0.12) (max-slopes: 0.16) dOUT(9)/dIN(9) 0.14 (min-slopes: 0.13) (max-slopes: 0.14) dOUT(15)/dIN(15) 0.14 (min-slopes: 0.11) (max-slopes: 0.16) lowest magnitude derivatives: dOUT(6)/dIN(6) -.01 (min-slopes: -.01) (max-slopes: -.01) dOUT(3)/dIN(2) -.01 (min-slopes: -.01) (max-slopes: -.01) dOUT(6)/dIN(12) -.01 (min-slopes: -.01) (max-slopes: -.01) dOUT(13)/dIN(6) 0.01 (min-slopes: 0.01) (max-slopes: 0.01) dOUT(11)/dIN(7) -.01 (min-slopes: -.02) (max-slopes: -.00) dOUT(1)/dIN(5) 0.01 (min-slopes: 0.01) (max-slopes: 0.01) dOUT(14)/dIN(7) -.01 (min-slopes: -.01) (max-slopes: -.00) dOUT(4)/dIN(6) -.01 (min-slopes: -.01) (max-slopes: -.01) dOUT(5)/dIN(6) -.01 (min-slopes: -.01) (max-slopes: -.01) dOUT(5)/dIN(12) -.01 (min-slopes: -.01) (max-slopes: -.01) dOUT(13)/dIN(7) -.01 (min-slopes: -.01) (max-slopes: -.01) dOUT(4)/dIN(12) -.01 (min-slopes: -.01) (max-slopes: -.01) dOUT(11)/dIN(6) 0.01 (min-slopes: 0.00) (max-slopes: 0.01) dOUT(3)/dIN(1) 0.01 (min-slopes: 0.00) (max-slopes: 0.01) dOUT(12)/dIN(7) 0.01 (min-slopes: 0.00) (max-slopes: 0.01) dOUT(1)/dIN(8) 0.01 (min-slopes: 0.00) (max-slopes: 0.01) dOUT(1)/dIN(9) 0.00 (min-slopes: -.00) (max-slopes: 0.01) dOUT(1)/dIN(10) 0.00 (min-slopes: 0.00) (max-slopes: 0.01) dOUT(15)/dIN(7) -.00 (min-slopes: -.00) (max-slopes: 0.00) ;;;----------------------------------- **Exp2** (sun jan 24 1999 10 17 32) ((attempting to learn output from (input) using 555 examples plus 238 held out) (network has 8 hidden units training for 35000 epochs with batch? 1 var-momentum? 1 learning.rate 0.01 momentum 0.7)) ((learned network stored on "/afs/cs/project/theo-35/fmri/backprop/train-1-24-99-8h") (training and holdout errors (9551880 0.0012161442 0.0013369615 34550))) Performance (first the network, then using mean voxel value as prediction): These predictions are over all available data. Train and holdout errors are similar. Network: Predictions: 793; RMS error per example for each output unit: (32.21 29.90 32.54 28.63 32.68 30.37 37.89 31.90 40.81 36.91 41.47 32.71 36.89 40.94 40.95 ) Average RMS error per example: 35.12 Mean voxel values as predictions: Predictions: 793; RMS error per example for each output unit: (38.25 37.14 43.21 36.56 37.07 37.93 51.65 43.40 57.38 44.22 66.96 43.86 51.42 65.82 54.18 ) Average RMS error per example: 47.27 For comparison, here is the corresponding data for network of EXP1 Predictions: 793 (all examples); RMS error per example: (33.12 30.47 32.69 28.83 33.42 30.68 37.96 32.20 40.93 37.15 41.70 33.00 37.19 40.89 41.35 ) RMS error per example per unit: 35.44 And here are predictions based on Voxel_i(t+1) <-- Voxel_i(t) Predictions: 793; RMS error per example for each output unit: (45.59 40.07 44.81 42.78 42.91 39.08 46.56 37.96 47.52 47.12 48.99 40.20 45.05 48.11 53.50 ) Average RMS error per example: 44.68 Sorted mean-slopes based on 238 examples, along with (min-slopes max-slopes) dOUT(7)/dIN(7) 0.40 (min-slopes: 0.32) (max-slopes: 0.49) dOUT(3)/dIN(3) 0.27 (min-slopes: 0.22) (max-slopes: 0.30) dOUT(9)/dIN(9) 0.26 (min-slopes: 0.23) (max-slopes: 0.27) dOUT(11)/dIN(11) 0.25 (min-slopes: 0.20) (max-slopes: 0.27) dOUT(15)/dIN(13) 0.25 (min-slopes: 0.22) (max-slopes: 0.25) dOUT(8)/dIN(8) 0.24 (min-slopes: 0.21) (max-slopes: 0.25) dOUT(3)/dIN(8) 0.24 (min-slopes: 0.20) (max-slopes: 0.26) dOUT(14)/dIN(11) 0.23 (min-slopes: 0.20) (max-slopes: 0.24) dOUT(14)/dIN(13) 0.22 (min-slopes: 0.19) (max-slopes: 0.23) dOUT(14)/dIN(15) 0.22 (min-slopes: 0.19) (max-slopes: 0.23) dOUT(13)/dIN(13) 0.22 (min-slopes: 0.19) (max-slopes: 0.22) dOUT(11)/dIN(15) 0.21 (min-slopes: 0.18) (max-slopes: 0.22) dOUT(9)/dIN(11) 0.19 (min-slopes: 0.17) (max-slopes: 0.20) dOUT(7)/dIN(5) 0.18 (min-slopes: 0.15) (max-slopes: 0.22) dOUT(9)/dIN(8) 0.18 (min-slopes: 0.17) (max-slopes: 0.19) dOUT(15)/dIN(1) 0.17 (min-slopes: 0.16) (max-slopes: 0.18) .... dOUT(12)/dIN(2) 0.01 (min-slopes: 0.00) (max-slopes: 0.01) dOUT(12)/dIN(8) 0.01 (min-slopes: 0.00) (max-slopes: 0.01) dOUT(3)/dIN(13) 0.01 (min-slopes: -.01) (max-slopes: 0.02) dOUT(3)/dIN(11) -.01 (min-slopes: -.01) (max-slopes: 0.00) dOUT(5)/dIN(8) -.00 (min-slopes: -.01) (max-slopes: -.00) dOUT(8)/dIN(15) 0.00 (min-slopes: 0.00) (max-slopes: 0.01) dOUT(14)/dIN(5) 0.00 (min-slopes: -.01) (max-slopes: 0.01) dOUT(6)/dIN(9) 0.00 (min-slopes: -.00) (max-slopes: 0.01) dOUT(5)/dIN(6) -.00 (min-slopes: -.00) (max-slopes: -.00) dOUT(1)/dIN(9) 0.00 (min-slopes: -.00) (max-slopes: 0.01) dOUT(4)/dIN(2) 0.00 (min-slopes: 0.00) (max-slopes: 0.00) dOUT(6)/dIN(10) -.00 (min-slopes: -.00) (max-slopes: -.00) dOUT(5)/dIN(11) 0.00 (min-slopes: -.00) (max-slopes: 0.00) dOUT(7)/dIN(6) 0.00 (min-slopes: -.01) (max-slopes: 0.01) dOUT(14)/dIN(12) 0.00 (min-slopes: -.00) (max-slopes: 0.00) dOUT(5)/dIN(2) 0.00 (min-slopes: -.00) (max-slopes: 0.00) dOUT(12)/dIN(10) -.00 (min-slopes: -.00) (max-slopes: 0.00)