R matrix is the matrix from x_t to prediction x_t+1: R = A*C +S which has the dimension of (sensor_numbers, sensor_nukmbers)
Statistics:
In file:
A: Layer2_time1109
B: Layer2_time1110
C:Layer3_time1107
D:Layer3_time1106
A = np.array([
[1.131289, 0.240440, 0.064601, -0.053835],
[-0.076714, 1.092745, 0.080804, -0.061647 ],
[0.077544, 0.048549, 1.096725, -0.094023 ],
[-0.094524, 0.252590, -0.007061, 1.373786 ],
])
B = np.array([
[1.124227, 0.246003, 0.056974, -0.043832],
[-0.070387, 1.087726, 0.087623, -0.070522],
[0.077246, 0.048774, 1.096448, -0.093626],
[-0.089635, 0.248712, -0.002097, 1.367048],
])
print("A: \n", A)
print("B: \n", B)
w, v = LA.eig(A)
w
>> array([1.3770925 +0.j , 1.11471654+0.14145481j,
1.11471654-0.14145481j, 1.08801942+0.j ])
w1, v1 = LA.eig(B)
w1
>>array([1.3542298 +0.j , 1.10916283+0.13574698j,
1.10916283-0.13574698j, 1.10289353+0.j ])
print(v[:,0])
print(v1[:,0])
>> [ 0.44210381+0.j 0.16677499+0.j 0.41236512+0.j -0.77890008+0.j]
[-0.47703554+0.j -0.21311281+0.j -0.44698785+0.j 0.72610047+0.j]
C = np.array([
[1.149754, 0.029372, 0.161366, -0.149284 ],
[-0.062357, 1.138778, 0.059344, 0.033338 ],
[0.052940, 0.023090, 1.126736, -0.095034 ],
[-0.111944, 0.365037, -0.083150, 1.326927 ],
])
Eigenvalue: array([1.47908338+0.j , 1.01431406+0.j ,
1.12439878+0.05620464j, 1.12439878-0.05620464j])
First EigenVector: [-0.49660374+0.j 0.12065799+0.j -0.28539083+0.j 0.81078878+0.j]
D= np.array([
[1.145186, 0.032887, 0.158309, -0.144351 ],
[-0.060468, 1.137329, 0.060610, 0.031261 ],
[0.048988, 0.026144, 1.124093, -0.090753 ],
[-0.109832, 0.363408, -0.081747, 1.324672 ],
])
Eigenvalue: array([1.46905131+0.j , 1.01402966+0.j ,
1.12409952+0.05607323j, 1.12409952-0.05607323j])
First EigenVector: [-0.48852883+0.j 0.1158168 +0.j -0.27620538+0.j 0.8195344 +0.j]
The A and B is the R matrix for two near frame (frame 1109 and 1110), and we see that it’s eigenvactor for the largetest eigenvalue v[:,0] and v1[:,0] has this dramatic change in direction for this two near frame.(Because simply this two near frame this should be similar! Thus the last layer this near frame eigenvectors is similar!) However the reason for the dramatic change is because the layer 3 eigenvector has effect on this layer 2 eigenvector direction change! We can see the direction from [ 0.44210381+0.j 0.16677499+0.j 0.41236512+0.j -0.77890008+0.j] to [-0.47703554+0.j -0.21311281+0.j -0.44698785+0.j 0.72610047+0.j] is somehow pushed into another direction by the Layer 3 eigenvector [-0.49660374+0.j 0.12065799+0.j -0.28539083+0.j 0.81078878+0.j] © and [-0.48852883+0.j 0.1158168 +0.j -0.27620538+0.j 0.8195344 +0.j] (D).
Configuration weight decay factor (0.8, 0.2)
From this I suppose this Higher level is trying to push the lower level eigenvectors for the largest Eigenvalue to its own (higher level) Eigenvalue directions!
0.2, 0.8 Configuration: More from brain(higher level), more instruction, not work because every layer should do it’s behaviour it cannot do brain’s instruction; it need it self to actually carry out stuff, the brain only serves as motivation, guidance signal.
Also : this is less information from the outside, but more information from the inside.