baseline runs/rnn_v31/rnn.pt feedforward=False params=327k clock runs/rnn_v31_clock/rnn.pt feedforward=False params=327k clock_priv runs/rnn_v31_clock_priv/rnn.pt feedforward=False params=327k clock_emerge runs/rnn_v31_clock_emerge/rnn.pt feedforward=False params=327k split default E=35 T=200 band=(0.13, 0.63) hidden_runs=101 split short E=30 T=200 band=(0.13, 0.45) hidden_runs=90 split long E=30 T=200 band=(0.13, 0.78) hidden_runs=88 position probe on TRUE hidden frames: mean decoded y 0.384 (band interior [0.21, 0.55]), 1.2% fall outside (a)+(f) probing h ... ============================================================================== OBJECT PERMANENCE -- v3 stage two ============================================================================== (a) ball position from a frozen probe, by visibility (R^2 / rmse, in world units) Read the rmse column across bins and the R^2 column down them: the hidden bin's variance is smaller than the visible bin's, so the two columns disagree on purpose. probe = linear feature visible partial hidden z 0.94/0.056 0.94/0.015 0.72/0.134 0.75/0.119 0.11/0.222 -0.02/0.102 h[baseline] 1.00/0.016 0.99/0.008 0.89/0.085 0.89/0.078 0.52/0.163 0.09/0.096 h[clock] 0.99/0.018 0.99/0.006 0.87/0.092 0.87/0.084 0.31/0.196 0.21/0.090 h[clock_priv] 1.00/0.015 0.99/0.007 0.89/0.082 0.89/0.081 0.42/0.180 0.14/0.094 h[clock_emerge] 0.99/0.019 0.98/0.009 0.87/0.092 0.89/0.080 0.28/0.200 0.22/0.089 probe = knn feature visible partial hidden z 0.99/0.023 0.95/0.014 0.83/0.104 0.86/0.090 -0.26/0.265 -0.25/0.113 h[baseline] 0.99/0.020 0.85/0.024 0.83/0.104 0.82/0.102 0.50/0.167 -0.18/0.110 h[clock] 0.99/0.019 0.84/0.025 0.81/0.110 0.79/0.108 0.51/0.166 -0.01/0.101 h[clock_priv] 0.99/0.017 0.86/0.023 0.84/0.101 0.83/0.097 0.56/0.157 -0.05/0.103 h[clock_emerge] 0.98/0.030 0.83/0.026 0.82/0.108 0.81/0.104 0.40/0.183 0.17/0.092 frames per bin: {'visible': 1703, 'partial': 2162, 'hidden': 3135} decay of position-from-h with frames hidden (rmse, world units; one frozen kNN probe) k 1 4 8 11 14 17 21 24 z (lin) 0.291 0.272 0.250 0.244 0.245 0.245 0.256 0.266 z 0.306 0.284 0.286 0.291 0.296 0.324 0.303 0.272 h[baseline] (lin) 0.189 0.160 0.173 0.187 0.198 0.219 0.224 0.224 h[baseline] 0.189 0.216 0.219 0.233 0.250 0.273 0.240 0.240 h[clock] (lin) 0.205 0.180 0.197 0.221 0.232 0.238 0.235 0.214 h[clock] 0.185 0.220 0.214 0.236 0.271 0.268 0.245 0.236 h[clock_priv] (lin) 0.170 0.157 0.166 0.180 0.183 0.201 0.203 0.219 h[clock_priv] 0.165 0.195 0.204 0.224 0.261 0.278 0.255 0.255 h[clock_emerge] (lin) 0.192 0.183 0.193 0.215 0.217 0.235 0.211 0.221 h[clock_emerge] 0.176 0.205 0.225 0.255 0.281 0.268 0.237 0.241 no memory 0.021 0.085 0.171 0.234 0.295 0.353 0.419 0.478 (no memory = predict the ball's last visible position; its error grows as speed x k and is the line every h curve has to get under. The `z` row is not zero because a tracking paddle leaks some ball x; `h[ff]` is the honest floor.) (f) what else is in h on fully hidden frames (linear R^2) feature ball_vx ball_vy frames_hidden z -0.005 -0.183 -0.577 h[baseline] 0.005 -0.324 -0.232 h[clock] 0.076 -0.063 -0.336 h[clock_priv] -0.011 0.039 -0.220 h[clock_emerge] -0.113 0.460 -0.295 report -> runs/rnn_v31_clock_permanence/report.json