Sparse Autoencoder를 통한 LLM의 Bias 줄이기 - 성에 따른 직업 4

2024.11.07 - [인공지능/XAI] - Sparse Autoencoder를 통한 LLM의 Bias 줄이기 - 성에 따른 직업 3

 

 

SAE는 7번 레이어에 붙어있습니다.

11레이어까지 있으니까 한 번 쭉 확인해봅시다....

Job	Gender Dominance	Female Percentage	Cosine Similarity with Woman	Cosine Similarity with Man
skincare specialist	Female	98.2	0.885237	0.851516
kindergarten teacher	Female	96.8	0.879833	0.844357
childcare worker	Female	94.6	0.908659	0.854653
secretary	Female	92.5	0.866899	0.826608
hairstylist	Female	92.4	0.881781	0.852713
dental assistant	Female	92	0.885272	0.84631
nurse	Female	91.3	0.892691	0.846891
school psychologist	Female	90.4	0.880224	0.829196
receptionist	Female	90	0.884083	0.84414
vet	Female	89.8	0.869386	0.852478
nutritionist	Female	89.6	0.886887	0.840343
maid	Female	88.7	0.919116	0.86704
therapist	Female	87.1	0.891367	0.849494
social worker	Female	86.8	0.873986	0.820296
sewer	Female	86.5	0.880158	0.854183
paralegal	Female	84.8	0.851667	0.838413
library assistant	Female	84.2	0.880276	0.830921
interior designer	Female	83.8	0.868622	0.829819
manicurist	Female	83	0.920434	0.916763
special education teacher	Female	82.8	0.870975	0.816348
police officer	Male	15.8	0.850793	0.803566
taxi driver	Male	12	0.879322	0.84247
computer architect	Male	11.8	0.88091	0.84602
mechanical engineer	Male	9.4	0.909123	0.881162
truck driver	Male	7.9	0.891951	0.859146
electrical engineer	Male	7	0.864773	0.823462
landscaping worker	Male	6.2	0.888274	0.854544
pilot	Male	5.3	0.876505	0.845214
repair worker	Male	5.1	0.896395	0.852561
firefighter	Male	5.1	0.898523	0.863758
construction worker	Male	4.2	0.888954	0.848474
machinist	Male	3.4	0.888064	0.871639
aircraft mechanic	Male	3.2	0.891025	0.853536
carpenter	Male	3.1	0.86811	0.843074
roofer	Male	2.9	0.847142	0.831337
brickmason	Male	2.2	0.867273	0.848192
plumber	Male	2.1	0.87239	0.842234
electrician	Male	1.7	0.894518	0.857185
vehicle technician	Male	1.2	0.880496	0.841039
crane operator	Male	1.1	0.89262	0.8612

이게 뭔....

 

Job	Gender Dominance	Female Percentage	Cosine Similarity with Woman	Cosine Similarity with Man
skincare specialist	Female	98.2	-0.28427	-0.18885
kindergarten teacher	Female	96.8	-0.29349	-0.15017
childcare worker	Female	94.6	0.183231	-0.09855
secretary	Female	92.5	-0.05228	-0.09362
hairstylist	Female	92.4	-0.15438	-0.00019
dental assistant	Female	92	-0.21715	-0.25152
nurse	Female	91.3	0.197028	0.026283
school psychologist	Female	90.4	-0.02939	-0.2249
receptionist	Female	90	-0.11885	-0.12152
vet	Female	89.8	-0.05999	0.101977
nutritionist	Female	89.6	-0.10836	-0.26641
maid	Female	88.7	0.452602	0.148875
therapist	Female	87.1	-0.03172	-0.11619
social worker	Female	86.8	-0.01379	-0.19423
sewer	Female	86.5	-0.03255	0.029441
paralegal	Female	84.8	-0.11245	0.103436
library assistant	Female	84.2	-0.06259	-0.19382
interior designer	Female	83.8	-0.14656	-0.16534
manicurist	Female	83	0.325327	0.609222
special education teacher	Female	82.8	-0.18254	-0.27691
police officer	Male	15.8	0.061902	-0.04618
taxi driver	Male	12	-0.14195	-0.17661
computer architect	Male	11.8	-0.07117	-0.07187
mechanical engineer	Male	9.4	-0.03773	0.142761
truck driver	Male	7.9	-0.04467	-0.04188
electrical engineer	Male	7	-0.31577	-0.31442
landscaping worker	Male	6.2	-0.31069	-0.2016
pilot	Male	5.3	0.000462	-0.00561
repair worker	Male	5.1	0.034276	-0.14766
firefighter	Male	5.1	0.030996	-0.07756
construction worker	Male	4.2	-0.15419	-0.11414
machinist	Male	3.4	-0.06887	0.178433
aircraft mechanic	Male	3.2	-0.17256	-0.20201
carpenter	Male	3.1	-0.37341	-0.0821
roofer	Male	2.9	-0.14815	0.04336
brickmason	Male	2.2	-0.14312	0.037151
plumber	Male	2.1	-0.11869	-0.05927
electrician	Male	1.7	-0.09495	-0.06853
vehicle technician	Male	1.2	-0.09524	-0.08894
crane operator	Male	1.1	-0.26544	-0.11965

내가 뭔갈 잘못하고 있나...?

 

이번엔 11레이어 

Job	Gender Dominance	Female Percentage	Cosine Similarity with Woman	Cosine Similarity with Man
skincare specialist	Female	98.2	0.952556	0.936987
kindergarten teacher	Female	96.8	0.951825	0.935702
childcare worker	Female	94.6	0.957717	0.932282
secretary	Female	92.5	0.93632	0.920797
hairstylist	Female	92.4	0.949957	0.937659
dental assistant	Female	92	0.950961	0.934148
nurse	Female	91.3	0.952866	0.933521
school psychologist	Female	90.4	0.945576	0.921782
receptionist	Female	90	0.95102	0.93408
vet	Female	89.8	0.942855	0.93628
nutritionist	Female	89.6	0.949447	0.927987
maid	Female	88.7	0.962526	0.939558
therapist	Female	87.1	0.952193	0.934614
social worker	Female	86.8	0.940401	0.916169
sewer	Female	86.5	0.949164	0.938944
paralegal	Female	84.8	0.939479	0.933251
library assistant	Female	84.2	0.943475	0.922672
interior designer	Female	83.8	0.94191	0.926157
manicurist	Female	83	0.968939	0.966691
special education teacher	Female	82.8	0.942118	0.917624
police officer	Male	15.8	0.925999	0.901801
taxi driver	Male	12	0.94509	0.928054
computer architect	Male	11.8	0.945797	0.928877
mechanical engineer	Male	9.4	0.961702	0.947846
truck driver	Male	7.9	0.951451	0.935774
electrical engineer	Male	7	0.93818	0.918885
landscaping worker	Male	6.2	0.954152	0.937349
pilot	Male	5.3	0.942039	0.930336
repair worker	Male	5.1	0.950627	0.929089
firefighter	Male	5.1	0.955618	0.939639
construction worker	Male	4.2	0.948317	0.930044
machinist	Male	3.4	0.953453	0.946909
aircraft mechanic	Male	3.2	0.951964	0.934727
carpenter	Male	3.1	0.945402	0.935081
roofer	Male	2.9	0.935538	0.928428
brickmason	Male	2.2	0.940855	0.932296
plumber	Male	2.1	0.947161	0.933655
electrician	Male	1.7	0.95383	0.937022
vehicle technician	Male	1.2	0.939626	0.920676
crane operator	Male	1.1	0.953499	0.938783

전혀 군집화가 안 보이네요..

Job	Gender Dominance	Female Percentage	Cosine Similarity with Woman	Cosine Similarity with Man
woman	Female	100	1	0.977756
skincare specialist	Female	98.2	0.952556	0.936987
kindergarten teacher	Female	96.8	0.951825	0.935702
childcare worker	Female	94.6	0.957717	0.932282
secretary	Female	92.5	0.93632	0.920797
hairstylist	Female	92.4	0.949957	0.937659
dental assistant	Female	92	0.950961	0.934148
nurse	Female	91.3	0.952866	0.933521
school psychologist	Female	90.4	0.945576	0.921782
receptionist	Female	90	0.95102	0.93408
vet	Female	89.8	0.942855	0.93628
nutritionist	Female	89.6	0.949447	0.927987
maid	Female	88.7	0.962526	0.939558
therapist	Female	87.1	0.952193	0.934614
social worker	Female	86.8	0.940401	0.916169
sewer	Female	86.5	0.949164	0.938944
paralegal	Female	84.8	0.939479	0.933251
library assistant	Female	84.2	0.943475	0.922672
interior designer	Female	83.8	0.94191	0.926157
manicurist	Female	83	0.968939	0.966691
special education teacher	Female	82.8	0.942118	0.917624
police officer	Male	15.8	0.925999	0.901801
taxi driver	Male	12	0.94509	0.928054
computer architect	Male	11.8	0.945797	0.928877
mechanical engineer	Male	9.4	0.961702	0.947846
truck driver	Male	7.9	0.951451	0.935774
electrical engineer	Male	7	0.93818	0.918885
landscaping worker	Male	6.2	0.954152	0.937349
pilot	Male	5.3	0.942039	0.930336
repair worker	Male	5.1	0.950627	0.929089
firefighter	Male	5.1	0.955618	0.939639
construction worker	Male	4.2	0.948317	0.930044
machinist	Male	3.4	0.953453	0.946909
aircraft mechanic	Male	3.2	0.951964	0.934727
carpenter	Male	3.1	0.945402	0.935081
roofer	Male	2.9	0.935538	0.928428
brickmason	Male	2.2	0.940855	0.932296
plumber	Male	2.1	0.947161	0.933655
electrician	Male	1.7	0.95383	0.937022
vehicle technician	Male	1.2	0.939626	0.920676
crane operator	Male	1.1	0.953499	0.938783
man	Male	0.1	0.977756	1

애초에 남자와 여자가 너무 가깝네요...?

 

음 한번 GPT -2 small SAE layer에서 cos 유사도를 구해보겠습니다.

prompt 1	prompt 2	cos similarity
man	woman	0.68
he	she	0.56
boy	girl	0.48
his	her	0.62
male	female	0.39

유사도가 가장 낮은 female과 male로 진행해 볼게요..

Job	Gender Dominance	Female Percentage	Cosine Similarity with female	Cosine Similarity with male
skincare specialist	Female	98.2	0.124757	0.134896
kindergarten teacher	Female	96.8	0.147743	0.160076
childcare worker	Female	94.6	0.141078	0.146944
secretary	Female	92.5	0.198375	0.201354
hairstylist	Female	92.4	0.128075	0.137902
dental assistant	Female	92	0.195335	0.204906
nurse	Female	91.3	0.244849	0.261608
school psychologist	Female	90.4	0.169121	0.154034
receptionist	Female	90	0.178872	0.179647
vet	Female	89.8	0.273882	0.289234
nutritionist	Female	89.6	0.177132	0.168246
maid	Female	88.7	0.238828	0.239321
therapist	Female	87.1	0.180114	0.187415
social worker	Female	86.8	0.167729	0.166011
sewer	Female	86.5	0.213534	0.227236
paralegal	Female	84.8	0.133382	0.142434
library assistant	Female	84.2	0.108575	0.109091
interior designer	Female	83.8	0.192761	0.199092
manicurist	Female	83	0.123101	0.1409
special education teacher	Female	82.8	0.136673	0.139805
police officer	Male	15.8	0.164144	0.15197
taxi driver	Male	12	0.167654	0.159419
computer architect	Male	11.8	0.144956	0.147659
mechanical engineer	Male	9.4	0.197546	0.207421
truck driver	Male	7.9	0.208168	0.221485
electrical engineer	Male	7	0.156188	0.155851
landscaping worker	Male	6.2	0.124597	0.128934
pilot	Male	5.3	0.210591	0.219529
repair worker	Male	5.1	0.128081	0.133228
firefighter	Male	5.1	0.213929	0.219438
construction worker	Male	4.2	0.186627	0.186755
machinist	Male	3.4	0.165033	0.17933
aircraft mechanic	Male	3.2	0.185949	0.195976
carpenter	Male	3.1	0.166936	0.179504
roofer	Male	2.9	0.21479	0.22808
brickmason	Male	2.2	0.130557	0.144164
plumber	Male	2.1	0.199274	0.211099
electrician	Male	1.7	0.128658	0.128711
vehicle technician	Male	1.2	0.168419	0.176294
crane operator	Male	1.1	0.164983	0.177588

흠 ...

이번엔 남자가 대부분 높네요...

 

그럼 이번엔 11번 레이어 에서 cos 유사도를 구해볼게요

prompt 1	prompt 2	cos similarity
man	woman	0.98
he	she	0.97
boy	girl	0.97
his	her	0.97
male	female	0.99

이건 그냥 다 높네요....

 

2번 레이어로 한번 볼게요

prompt 1	prompt 2	cos similarity
man	woman	0.88
he	she	0.88
boy	girl	0.87
his	her	0.80
male	female	0.92
man	he	0.75
woman	she	0.79

....?

애초에 직업에 대한 코사인 유사도보다 man, woman 이 더 높네요....

 

다시 미스트랄로....

mistral은 레이어가 깊어질수록 cos 유사도가 떨어지는 경향이 있어서 31번 레이어로 진행했습니다.

기본적으로 job에 대해 cos 유사도가 man보다 woman이 더 높네요 -> 그래서 모든 직업에서 여자의 cos 유사도가 더 높게 나오는 것 같습니다.

Job	Gender Dominance	Female Percentage	Cosine Similarity with Woman	Cosine Similarity with Man
woman	Female	100	1	0.478665
skincare specialist	Female	98.2	0.526142	0.337138
kindergarten teacher	Female	96.8	0.571344	0.319337
childcare worker	Female	94.6	0.653964	0.321974
secretary	Female	92.5	0.490099	0.289095
hairstylist	Female	92.4	0.475824	0.333437
dental assistant	Female	92	0.46657	0.21639
nurse	Female	91.3	0.576348	0.327447
school psychologist	Female	90.4	0.555125	0.294005
receptionist	Female	90	0.579533	0.30584
vet	Female	89.8	0.380612	0.295048
nutritionist	Female	89.6	0.530972	0.265065
maid	Female	88.7	0.426042	0.352122
therapist	Female	87.1	0.616198	0.312762
social worker	Female	86.8	0.598205	0.277394
sewer	Female	86.5	0.338412	0.328602
paralegal	Female	84.8	0.473365	0.355198
library assistant	Female	84.2	0.306555	0.190848
interior designer	Female	83.8	0.553535	0.26863
manicurist	Female	83	0.646296	0.64153
special education teacher	Female	82.8	0.528594	0.267602
police officer	Male	15.8	0.600642	0.303183
taxi driver	Male	12	0.583161	0.318098
computer architect	Male	11.8	0.515812	0.300871
mechanical engineer	Male	9.4	0.560198	0.285663
truck driver	Male	7.9	0.604801	0.328755
electrical engineer	Male	7	0.556409	0.275509
landscaping worker	Male	6.2	0.600496	0.296639
pilot	Male	5.3	0.509496	0.349558
repair worker	Male	5.1	0.641669	0.318395
firefighter	Male	5.1	0.512231	0.322519
construction worker	Male	4.2	0.611927	0.302671
machinist	Male	3.4	0.507236	0.3774
aircraft mechanic	Male	3.2	0.546825	0.296465
carpenter	Male	3.1	0.490861	0.403917
roofer	Male	2.9	0.465377	0.376143
brickmason	Male	2.2	0.471654	0.331363
plumber	Male	2.1	0.432962	0.369362
electrician	Male	1.7	0.538109	0.291336
vehicle technician	Male	1.2	0.480898	0.273057
crane operator	Male	1.1	0.558155	0.364165
man	Male	0	0.478665	1

그냥 널리 퍼져있네요....

Job: doctor
  Cosine similarity with man: 0.29202771186828613
  Cosine similarity with woman: 0.3873016834259033
  Cosine similarity with he: 0.18976394832134247
  Cosine similarity with she: 0.24517714977264404
  Cosine similarity with male: 0.26682591438293457
  Cosine similarity with female: 0.30686309933662415
Job: nurse
  Cosine similarity with man: 0.32744666934013367
  Cosine similarity with woman: 0.5763475894927979
  Cosine similarity with he: 0.32279425859451294
  Cosine similarity with she: 0.2484569251537323
  Cosine similarity with male: 0.2853856086730957
  Cosine similarity with female: 0.3248463273048401
Job: engineer
  Cosine similarity with man: 0.3628814220428467
  Cosine similarity with woman: 0.535031795501709
  Cosine similarity with he: 0.30105212330818176
  Cosine similarity with she: 0.24273112416267395
  Cosine similarity with male: 0.2686672806739807
  Cosine similarity with female: 0.31331366300582886
Job: teacher
  Cosine similarity with man: 0.3335043787956238
  Cosine similarity with woman: 0.5944691300392151
  Cosine similarity with he: 0.3386254608631134
  Cosine similarity with she: 0.24903792142868042
  Cosine similarity with male: 0.2746357321739197
  Cosine similarity with female: 0.33448266983032227
Job: scientist
  Cosine similarity with man: 0.3251022696495056
  Cosine similarity with woman: 0.6097099781036377
  Cosine similarity with he: 0.37819525599479675
  Cosine similarity with she: 0.2599332332611084
  Cosine similarity with male: 0.2557474374771118
  Cosine similarity with female: 0.3047844469547272

모델이 오히려 논문에서 나온 것 보다 출력이 잘 안나오네요...

이렇게 해야 나오네여

일단 이제 편향이 있는 것은 확인 했으니 편향을 만드는 feature를 찾아봅시다.