An Asymptotic Analysis of Generative, Discriminative, and Pseudolikelihood Estimators

author: Percy Liang, Berkeley , University of California

Description

Statistical and computational concerns have motivated parameter estimators based on various forms of likelihood, e.g., joint, conditional, and pseudolikelihood. In this paper, we present a unified framework for studying these estimators, which allows us to compare their relative (statistical) efficiencies. Our asymptotic analysis suggests that modeling more of the data tends to reduce variance, but at the cost of being more sensitive to model misspecification. We present experiments validating our analysis.

You might be experiencing some problems with Your Video player.

*Slides*
0:00	An Asymptotic Analysis of Estimators: Generative, Discriminative, Pseudolikelihood
0:07	Goal (1)
0:30	Goal (2)
0:50	Goal (3)
1:04	Goal (4)
1:16	Goal (5)
1:43	Existing intuitions (1)
2:01	Existing intuitions (2)
2:07	Existing intuitions (3)
2:18	Existing intuitions (4)
2:45	Existing intuitions (5)
2:53	Model-based estimators and neighborhoods (1)
3:26	Model-based estimators and neighborhoods (2)
4:04	Model-based estimators and neighborhoods (3)
4:09	Model-based estimators and neighborhoods (4)
4:14	Model-based estimators and neighborhoods (5)
4:28	Model-based estimators and neighborhoods (6)
4:39	Model-based estimators and neighborhoods (7)
4:43	Model-based estimators and neighborhoods (8)
4:44	Composite likelihood estimators (1)
5:19	Composite likelihood estimators (2)
5:21	Composite likelihood estimators (3)
5:41	Composite likelihood estimators (4)
5:45	Composite likelihood estimators (5)
6:15	Review of exponential families (1)
6:59	Review of exponential families (2)
7:34	Sketch of arguments for comparing estimators (1)
7:59	Sketch of arguments for comparing estimators (2)
8:23	Sketch of arguments for comparing estimators (3)
8:49	Sketch of arguments for comparing estimators (4)
9:02	Sketch of arguments for comparing estimators (5)
9:10	Sketch of arguments for comparing estimators (6)
9:50	Sketch of arguments for comparing estimators (7)
10:00	Sketch of arguments for comparing estimators (8)
10:26	Sketch of arguments for comparing estimators (9)
10:44	Sketch of arguments for comparing estimators (10)
10:51	Sketch of arguments for comparing estimators (11)
11:10	Overview of asymptotic analysis (1)
11:59	Overview of asymptotic analysis (2)
12:39	Overview of asymptotic analysis (3)
12:55	Overview of asymptotic analysis (4)
13:12	Overview of asymptotic analysis (5)
13:36	Well-specified case (1)
13:59	Well-specified case (2)
14:07	Well-specified case (3)
14:25	Well-specified case (4)
14:38	Well-specified case: comparing two estimators (1)
15:08	Well-specified case: comparing two estimators (2)
15:31	Well-specified case: comparing two estimators (3)
15:51	Well-specified case: comparing two estimators (4)
16:00	Well-specified case: comparing two estimators (5)
16:16	Well-specified case: comparing two estimators (6)
16:27	Multiple components (1)
16:47	Multiple components (2)
16:53	Multiple components (3)
17:07	Multiple components (4)
17:17	Multiple components (5)
17:28	Multiple components (6)
17:41	Multiple components (7)
18:22	Misspecified case (1)
18:38	Misspecified case (2)
18:49	Misspecified case (3)
19:01	Misspecified case (4)
19:23	Verifying the error rates empirically (1)
19:53	Verifying the error rates empirically (2)
20:19	Verifying the error rates empirically (3)
20:28	Verifying the error rates empirically (4)
20:45	Verifying the error rates empirically (5)
21:07	Verifying the error rates empirically (6)
21:26	Verifying the error rates empirically (7)
21:34	Verifying the error rates empirically (8)
21:55	Verifying the error rates empirically (9)
21:59	Application: part-of-speech tagging (1)
22:18	Application: part-of-speech tagging (2)
22:25	Application: part-of-speech tagging (3)
23:11	Application: part-of-speech tagging (4)
23:28	Summary (1)
23:46	Summary (2)
23:55	Summary (3)

Lecture rating

People found this lecture:
Worth seeing
because it is:
Valuable and informative
Well presented
Easily understandable
Acceptably recorded
You need to login to cast your vote.

Report a problem or upload files

If you have found a problem with this lecture or would like to send us extra material, articles, exercises, etc., please use our ticket system to describe your request and upload the data.
Enter your e-mail into the 'Cc' field, and we will keep you updated with your request's status.