Beruflich Dokumente
Kultur Dokumente
Experiments
Test Error Speed Increase
Conclusion
Motivation
Fast CrossValidation via Sequential Analysis Tammo Krueger Danny Panknin Mikio Braun Motivation Fast CV
Algorithm Meta-parameters Example Run
Cross-validation is an indispensable tool for applied ML but unfortunately very time consuming Example tted Q iteration:
Cross-Validation
10 10 10
parameter fold
50
max. iter.
Experiments
Test Error Speed Increase
Conclusion
Directly optimizing the error landscape to avoid calculations dicult due to noise Our approach: use increasing subsets of the training data
1 2 3
smaller subsets less training time more training data better error estimate relative behavior of parameter congurations converges
qqqqqqqqqqqqq
qqqqq
qqq
q qq q q q q q q q q q qq qq qqq qq q
qqqq
qqqqq
qqq
0.4
Size
q
Class. Error
0.3
Experiments
Test Error Speed Increase
0.2
1000
Conclusion
0.1 4 3 2 1 0 1 2
log()
2
qqqqqqqqqqqqqqqqqq qqqqqq q q qqqq q qq q qq q qq qqq q qq q q qq qq q q qq qqqqqq
3
qq qqqqqqqqqqqqqqq qqq qqqq qqqqqqq q qq qqqq q qq qqq q qq qq q qqq qqqqqqqq
0.4
0.3
0.2
0.1 4 0.5
qq q qqqqqqqqqqqqqqqqqq qqqq qqqqqq q q q qqqq qqq q qq qq q q qq q qqq q qq qqqq
5
qqq qqqqqqqqqqqqqqqq qqqqq q qq qqq qq q q qq qq q q qq qq q q q q qq q qq q qqqqqqq
6
qqqqqqqqqqqqqqq qqqqqqqqqq qq qqqqqqq q qqq q qqq q qq qq q qq q qqq qqqqqqq
Size
q
Class. Error
Fast CV
Algorithm Meta-parameters Example Run
50 100
0.3
250 500
0.2
750 1000
Experiments
Test Error Speed Increase
0.1 7 0.5
qqqqqqqqqqqqqqqq q qqqqq qq qqqq q q q qqq qqq q qq q qqq q qqq qqq q q qq qqqqqq
8
qqqqqqqqqqqqqqqqq qqqqq qqqq q qq qqqqqqq qqq qq qq q qq qq q q qq q qq qqqqqq
9
q qqqqqqqqqqqqqq qqqqq qqqq q q q qqq q qq qqq qq q q qq q qq q qqq q qqqq qqqqqqq
Conclusion
0.4
0.3
0.2
0.1 4 3 2 1 0 1 2 4 3 2 1 0 1 2 4 3 2 1 0 1 2
log()
Motivation Exploitation
Fast CrossValidation via Sequential Analysis Tammo Krueger Danny Panknin Mikio Braun Motivation Fast CV
Algorithm Meta-parameters Example Run
Observations: Individual runs are noisy, but at least we can see the tendency A lot of underperforming parameter congurations We can estimate the correct parameter on a suciently large subset of the data Exploitation: Transformation of the pointwise test errors of the congurations into a binary top or op scheme Dropping of signicant loser congurations along the way via tests from the sequential analysis framework Early stopping of the procedure, when we have seen enough data for a stable parameter estimation
Experiments
Test Error Speed Increase
Conclusion
1 0 0 0
2 0 1 1
3 0 0 1
4 0 0 0
steps 5 1 0 0 . . .
6 0 0 1 1 1 1
7 0 0 0 1 0 0
8 0 0 0 0 1 1
9 0 0 0 1 1 1
10 0 0 0 1 1 1
() ()
0.6 0.6 0.7 -0.8 -0.4 0.1 0.5 0.7 -0.9 -0.1 0.5 0.4 0.6 -0.3 0.0 pointwisePerformance matrix
0 0 1
1 0 1 1 1 1 1 1 1 0 1 1 trace matrix
Sa(0, 1, l, l)
c1 ck 15 Cumulative Sum
WINNER
Experiments
Test Error Speed Increase
c3 . . . ck2 ck1 ck
7 0 1 0 0
8 0 . . . 0 1 1
9 0 1 1 1
10 0 1 1 1
Conclusion
10
1 1 0 1 1 0 0 0 0 0 0 0 0 0 0 5 10 Step
1 0
LOSER
H0(0, 1, l, l)
15 20
= N/20 modelSize = 10 n = N 10
c1 c2 c3 . . .
0 0 0 0 0 1
0 1 1
0 0 1
0 0 0
1 0 0
0 0 1 1 1 1
0 0 0 1 0 0
0 0 0 0 1 1
ck2 ck1 ck
0.6 0.6 0.7 -0.8 -0.4 0.1 0.5 0.7 -0.9 -0.1 0.5 0.4 0.6 -0.3 0.0 pointwisePerformance matrix
. .
1 0 1 1 1 1 1 1 1 0 1 1 trace matrix
20
Sa(0, 1, l, l)
c1 ck
WINNER
15
c3 . . . ck2 ck1 ck
7 0 1 0 0
8 0 . . .
Motivation Fast CV
Algorithm Meta-parameters Example Run
0 1 1
10
Experiments
Test Error Speed Increase
1 1 0 0 0 0 0 0 0 0 0 0 5 10 Step
1 0
1 0
LOSER
H0(0, 1, l, l)
15 20
= modelSize = n=
Conclusion
(0 , 1 ) =argmax
0 ,1
H0 (0 , 1 , l , l )
q q q q
0.6
Pi
q
0.1 0.2
0.4
0.2
Experiments
Test Error Speed Increase
0.0
Change Point
10
15
Conclusion
600
500
400
Configuration
Status 300 flop top out 200 80 150 100 100 50 150 100 50 60 40 20 70 60 50 40 30 20 10 4 5 6 5 6 7
Experiments
Test Error Speed Increase
Conclusion
234
345
Step
Experimental Setup
Fast CrossValidation via Sequential Analysis Tammo Krueger Danny Panknin Mikio Braun Motivation Fast CV
Algorithm Meta-parameters Example Run
Experiments
Test Error Speed Increase
Conclusion
0.06
q q
Classification
Regression
q q q q q q
0.04
q q
q q
0.02
q q q q q q q q q q q q q q q q q q q q q q q q q
q q q q q q q q
method kreg sv
q
0.00
q q q
Experiments
Test Error Speed Increase
q q
0.02
Conclusion
0.04
q q
waveform
ringnorm
twonorm
pumadyn32nm
bank32nm
covtype
german
banana
heavisine
image
kin32nm
splice
doppler
bumps
blocks
Classification 140
q
Regression
q q q
120
q
100
q q
q q
80
method kreg sv
60
q q q q q q q q
40
Experiments
Test Error Speed Increase
20
q q q q q
Conclusion
pumadyn32nm
waveform
ringnorm
twonorm
bank32nm
covtype
german
banana
heavisine
image
kin32nm
splice
doppler
bumps
blocks
Motivation: we can estimate the correct parameter on a suciently large subset of the data Transformation: Race of congurations evaluated on linearly increasing subsets of the data At each step of this race:
Transform the test errors on individual data points of the remaining congurations into a binary top or op scheme 2 Drop signicant loser congurations along the way using tests from the sequential analysis framework 3 Apply distribution free testing techniques to decide, whether we have gathered enough evidence for a stable parameter estimation
1
Experiments
Test Error Speed Increase
Conclusion
Fast CrossValidation via Sequential Analysis Tammo Krueger Danny Panknin Mikio Braun Motivation Fast CV
Algorithm Meta-parameters Example Run
Experiments
Test Error Speed Increase
Conclusion
covtype
german
q q q q q q q q q q q
q q qq q q q q
image
ringnorm
splice
value
Fast CV
Algorithm Meta-parameters Example Run
method kreg sv
q q q q q q q q q q q q q q qq q q q q q q q qq q qq q qq q qq
Experiments
Test Error Speed Increase
twonorm
waveform
Conclusion
S1 S2 S3 S4 S5 S6 S7 S8 S9S10
S1 S2 S3 S4 S5 S6 S7 S8 S9S10
S1 S2 S3 S4 S5 S6 S7 S8 S9S10
variable
blocks
bumps
doppler
heavisine
q q
kin32nm
value
Fast CV
Algorithm Meta-parameters Example Run
method kreg sv
Experiments
Test Error Speed Increase
pumadyn32nm
Conclusion
S1 S2 S3 S4 S5 S6 S7 S8 S9S10
S1 S2 S3 S4 S5 S6 S7 S8 S9S10
S1 S2 S3 S4 S5 S6 S7 S8 S9S10
variable