File size: 6,589 Bytes
f6cc031
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
<Poster Width="1734" Height="1227">
	<Panel left="31" right="156" width="408" height="237">
		<Text>1. C ONTRIBUTIONS</Text>
		<Text>The key contributions of our work are:</Text>
		<Text>• Analysis of the spatial receptive field (RF) designs for</Text>
		<Text>pooled features.</Text>
		<Text>• Evidence that spatial pyramids may be suboptimal in</Text>
		<Text>feature generation.</Text>
		<Text>• An algorithm that jointly learns adaptive RF and</Text>
		<Text>the classifiers, with an efficient implementation using</Text>
		<Text>over-completeness and structured sparsity.</Text>
	</Panel>

	<Panel left="451" right="157" width="828" height="238">
		<Text>2. T HE P IPELINE</Text>
		<Figure left="532" right="194" width="667" height="137" no="1" OriWidth="0.555491" OriHeight="0.0848972
" />
		<Text>State-of-the-art classification algorithms take a two-layer pipeline: the coding layer learns activations from local</Text>
		<Text>image patches, and the pooling layer aggregates activations in multiple spatial regions. Linear classifiers are learned</Text>
		<Text>from the pooled features.</Text>
	</Panel>

	<Panel left="1293" right="156" width="407" height="239">
		<Text>3. N EUROSCIENCE I NSPIRATION</Text>
		<Figure left="1347" right="199" width="314" height="178" no="2" OriWidth="0" OriHeight="0
" />
	</Panel>

	<Panel left="30" right="408" width="409" height="408">
		<Text>4. S PATIAL P OOLING R EVISITED</Text>
		<Text>• Much work has been done on the coding part, while</Text>
		<Text>the spatial pooling methods are often hand-crafted.</Text>
		<Text>• Sample performances on CIFAR-10 with different re-</Text>
		<Text>ceptive field designs:</Text>
		<Figure left="102" right="553" width="280" height="132" no="3" OriWidth="0" OriHeight="0
" />
		<Text>Note the suboptimality of SPM - random selection</Text>
		<Text>from an overcomplete set of spatially pooled features</Text>
		<Text>consistently outperforms SPM.</Text>
		<Text>• We propose to learn the spatial receptive fields as well</Text>
		<Text>as the codes and the classifier.</Text>
	</Panel>

	<Panel left="31" right="830" width="406" height="364">
		<Text>5. N OTATIONS</Text>
		<Text>• I: image input.</Text>
		<Text>• A1 , · · · , AK : code activation as matrices, with Akij : ac-</Text>
		<Text>tivation of code k at position (i, j).</Text>
		<Text>• Ri : RF of the i-th pooled feature.</Text>
		<Text>• op(·): pooling operator, such as max(·).</Text>
		<Text>• f (x, θ): the classifier based on pooled features x.</Text>
		<Text>• A pooled feature xi is defined by choosing a code in-</Text>
		<Text>dexed by ci and a spatial RF Ri :</Text>
		<Text>The vector of pooled features x is then determined</Text>
		<Text>by the set of parameters C = {c1 , · · · , cM } and R =</Text>
		<Text>{R1 , · · · , RM }.</Text>
	</Panel>

	<Panel left="452" right="410" width="406" height="361">
		<Text>6. T HE L EARNING P ROBLEM</Text>
		<Text>N{(In , yn )}n=1 ,• Given a set of training datawe jointly</Text>
		<Text>learn the classifier and the pooled features as (assum-</Text>
		<Text>ing that coding is done in an unsupervised way):</Text>
		<Text>• Advantage: pooled features are tailored towards the</Text>
		<Text>classification task (also reduces redundancy).</Text>
		<Text>• Disadvantage: may be intractable - an exponential</Text>
		<Text>number of possible receptive fields.</Text>
		<Text>• Solution: reasonably overcomplete receptive field</Text>
		<Text>candidates + sparsity constraints to control the num-</Text>
		<Text>ber of final features.</Text>
	</Panel>

	<Panel left="452" right="785" width="406" height="411">
		<Text>7. O VERCOMPLETE RF</Text>
		<Text>• We propose to use overcomplete receptive field can-</Text>
		<Text>didates based on regular grids:</Text>
		<Figure left="487" right="879" width="102" height="100" no="4" OriWidth="0.104624" OriHeight="0.0817694
" />
		<Text> (a) Base</Text>
		<Figure left="606" right="878" width="102" height="101" no="5" OriWidth="0.104624" OriHeight="0.080429
" />
		<Text> (b) SPM</Text>
		<Figure left="725" right="878" width="102" height="101" no="6" OriWidth="0.103468" OriHeight="0.0808758
" />
		<Text> (c) Ours</Text>
		<Text>• The structured sparsity regularization is adopted to</Text>
		<Text>select only a subset of features for classification:</Text>
	</Panel>

	<Panel left="871" right="409" width="407" height="787">
		<Text>8. G REEDY F EATURE S ELECTION</Text>
		<Text>• Directly perform optimization is still time and mem-</Text>
		<Text>ory consuming.</Text>
		<Text>• Following [Perkins JMLR03], We adopted an incre-</Text>
		<Text>mental, greedy approach to select features based on</Text>
		<Text>their scores:</Text>
		<Text>• After each increment, the model is retrained only with</Text>
		<Text>respect to an active subset of selected features to en-</Text>
		<Text>sure fast re-training:</Text>
		<Figure left="983" right="753" width="193" height="160" no="7" OriWidth="0.212717" OriHeight="0.139857
" />
		<Text>• Benefit of overcompleteness in spatial pooling + fea-</Text>
		<Text>ture selection: higher performance with smaller code-</Text>
		<Text>books and lower feature dimensions.</Text>
		<Figure left="981" right="1008" width="222" height="169" no="8" OriWidth="0.244509" OriHeight="0.150134
" />
	</Panel>

	<Panel left="1293" right="410" width="405" height="540">
		<Text>9. R ESULTS</Text>
		<Text>• Performance comparison on CIFAR-10 with state-of-</Text>
		<Text>the-art approaches:</Text>
		<Figure left="1332" right="506" width="354" height="213" no="9" OriWidth="0.343353" OriHeight="0.152368
" />
		<Text>• Result on MNIST and the 1-vs-1 saliency map ob-</Text>
		<Text>tained from our algorithm:</Text>
		<Figure left="1327" right="790" width="359" height="155" no="10" OriWidth="0.358382" OriHeight="0.120197
" />
	</Panel>

	<Panel left="1294" right="964" width="404" height="230">
		<Text>10. R EFERENCES</Text>
		<Text>• A Coates and AY Ng. The importance of encoding</Text>
		<Text>versus training with sparse coding and vector quanti-</Text>
		<Text>zation. ICML 2011.</Text>
		<Text>• S Perkins, K Lacker, and J Theiler. Grafting: fast, incre-</Text>
		<Text>mental feature selection by gradient descent in func-</Text>
		<Text>tion space. JMLR, 3:1333–1356, 2003.</Text>
		<Text>• DH Hubel and TN Wiesel. Receptive fields, binocu-</Text>
		<Text>lar interaction and functional architecture in the cat’s</Text>
		<Text>visual cortex. J. of Physiology, 160(1):106–154, 1962.</Text>
	</Panel>

</Poster>