Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Functional Models of Mouse Visual CortexEkaterina Taralova, Tony Jebara, Rafael Yuste
Data Model Validation
2
Data
Data collected at 3.7fps, video playing at 30fps.3
Average of spontaneous and evoked Calcium activity (7.6K samples at 3.7 fps).
Automatically extracted ROIs (cell hypothesis) using Foopsi.
~100 neurons
Data
4
Automatically inferred spikes using Foopsi for spontaneous and evoked activity. Gratings start after red line.
Data
5
• Each neuron is a node in a graph • Edges encode conditional functional dependencies • No hidden nodes • Static model
CRF model
6
1
13
2
1434
38102
3
5
61 63
74
4
27
40
118
2641
6
1273
86
132
7
103109 121
8
1851
9
68 12810
11
105108
93
57 79
15
2448
88
16 17
112
115
117
19
33
20
31
85
21
22
45
23
69106
3237
50
127
25
28
71 29
30
46
66
83 110
35
47
113
36
44
119
39
131
84
42
62 130
43
53 96
100 95 124
5467
49
89
133
557652
97 126
111
707256
7875
58
77
59
60
64
125 129
65
81
107
123
135 91
9287
120
98
80
99
82
90
114
122 94101
104
116 134
1 2 3
204146
54
111
4
31
5 67
8
39 4758 5967
9
10
11
84
12
102
1314 15 16 1718 19
2745 48 80
86
90 99
103
118
121
124130
21
73
22
25
106
11523
5068
24
74
2628
40
29
30
113
32 33 34
3536
98
37 38
66 85105
110 112
125
42
43
123
44
49 51
52
53
57 83 87 95 107 122 128
55 56
72
60 61 62 63 64 65 69 70 71
75
76 77
78
79 8182
117
88 89 91 92 93 94 96 97 100 101 104 108 109 114 116 119 120 126 127 129 131 132 133 134 135
CRF model
train likelihood: -5.3206 test likelihood: -6.4686 true test likelihood: -15.7394
Loopy model:train likelihood: -3.0674 test likelihood: -4.8702 true test likelihood: -16.0309
Tree model:
7
Accuracy of inferring individual neurons, given the rest. Spontaneous activity model.
Validation
Accuracy of inferring individual neurons, given the rest. 90 degree gratings model.
Validation
Mean accuracy of inferring individual neurons over 120 test frames.
Validation
10
Automatically inferred spikes using Foopsi for spontaneous and evoked activity. Gratings start after red line.
Data
11
0 degree (vertical)
Accuracy of inferring individual neurons, given the rest. 0 degree model.
Validation
Accuracy of inferring individual neurons, given the rest. 135 degree model.
Validation
Accuracy of inferring individual neurons, given the rest. 45 degree model.
Validation
Sampled neural activity from tree model
15
Neurons that fire > 50% of the time, per stimulus model
Spontaneous 0 degree
135 degree45 degree 16
• Models for natural image stimuli. • Models across V1 layers, 300-500 neurons per layer, ~4K
per animal (Paul Allen Institute dataset). • Encode expert knowledge during structure learning.
Work in progress
17
8012
0
1
35
V1 orientation tuning
18
V1 orientation tuning
135 degree, layer 1 tree model, orientation selectivity
90 degree, layer 1 tree model, orientation selectivity
19
• Encode expert knowledge during structure learning.
Normalized Projection Density
= (NPV)Source volume
Target volume= Target fluor.
Target vol.
Source fluor.
Source vol.
AL AM LM LI VI P PM POR RLLPLDLGd
SC
ALAM
LMLI
VI
P
PM
POR
RL
LGd
55
5
5
1 2/3 1 2/3 1 2/3 1 2/3 1 2/3 1 2/3 1 2/3 1 2/3 1 2/3
20
LGN
LM
LMRL
RL
V1
“Wiring diagram”
Glickfeld et al 201321
90o Stimulus
2/3 2/35 4 2/3 4 5
124 101 174 112 115 127
31 63 58 83 98
43 33 38 44
54 58 89
46 47
87
2/3 2/35 4 2/3 4 5
150 113 182 110 102 103
41 63 62 84 105
32 40 30 45
47 49 97
48 60
73
180o Stimulus
Glickfeld et al 2013
“Wiring diagram”
22
Functional Models of Mouse Visual CortexEkaterina Taralova, Tony Jebara, Rafael Yuste
Thank you!
24