Sparse Autoencoder를 통한 LLM의 Bias 줄이기 - 성에 따른 직업 3

2024.11.05 - [인공지능/XAI] - Sparse Autoencoder를 통한 LLM의 Bias 줄이기 - 성에 따른 직업 2

 

이번에는 SAE의 Feature에서 COS 유사도를 구하는 것이 아니라 그 이후의 레이어에서 확인해 보도록 하겠습니다.

 

Job	Gender Dominance	Female Percentage	Cosine Similarity with Woman	Cosine Similarity with Man
skincare specialist	Female	98.2	0.651186	0.556518
kindergarten teacher	Female	96.8	0.649491	0.514379
childcare worker	Female	94.6	0.71621	0.519848
secretary	Female	92.5	0.583385	0.54581
hairstylist	Female	92.4	0.634154	0.553486
dental assistant	Female	92	0.594918	0.44044
nurse	Female	91.3	0.712285	0.57134
school psychologist	Female	90.4	0.640011	0.494764
receptionist	Female	90	0.63295	0.544817
vet	Female	89.8	0.521638	0.586588
nutritionist	Female	89.6	0.620365	0.503985
maid	Female	88.7	0.557642	0.581538
therapist	Female	87.1	0.607497	0.489134
social worker	Female	86.8	0.705663	0.54845
sewer	Female	86.5	0.637374	0.63115
paralegal	Female	84.8	0.625037	0.606166
library assistant	Female	84.2	0.422588	0.460436
interior designer	Female	83.8	0.637165	0.494837
manicurist	Female	83	0.644541	0.701974
special education teacher	Female	82.8	0.661597	0.507501
police officer	Male	15.8	0.70545	0.507632
taxi driver	Male	12	0.642695	0.500626
computer architect	Male	11.8	0.651701	0.564702
mechanical engineer	Male	9.4	0.637769	0.532792
truck driver	Male	7.9	0.667702	0.494411
electrical engineer	Male	7	0.628306	0.492114
landscaping worker	Male	6.2	0.669272	0.483533
pilot	Male	5.3	0.629289	0.603345
repair worker	Male	5.1	0.678439	0.516247
firefighter	Male	5.1	0.653101	0.52751
construction worker	Male	4.2	0.718746	0.525603
machinist	Male	3.4	0.644379	0.622361
aircraft mechanic	Male	3.2	0.620817	0.460778
carpenter	Male	3.1	0.654313	0.653469
roofer	Male	2.9	0.624167	0.607007
brickmason	Male	2.2	0.579816	0.553472
plumber	Male	2.1	0.629519	0.639175
electrician	Male	1.7	0.642468	0.532843
vehicle technician	Male	1.2	0.614766	0.479526
crane operator	Male	1.1	0.673586	0.602639

여기서도 여자가 전부다 COS 유사도가 높은 편향이 있네요.....

 

TSNE 또한 딱히 안 보이네요..

흠 UMAP 또한 아무것도 안보이네요...

 

Job: doctor
  Cosine similarity with man: 0.6425841450691223
  Cosine similarity with woman: 0.6595352292060852
  Cosine similarity with he: 0.5585692524909973
  Cosine similarity with she: 0.6580565571784973
Job: nurse
  Cosine similarity with man: 0.5713397264480591
  Cosine similarity with woman: 0.7122852802276611
  Cosine similarity with he: 0.5225833654403687
  Cosine similarity with she: 0.5876948833465576
Job: engineer
  Cosine similarity with man: 0.6049615144729614
  Cosine similarity with woman: 0.6559303402900696
  Cosine similarity with he: 0.5704656839370728
  Cosine similarity with she: 0.6277342438697815
Job: teacher
  Cosine similarity with man: 0.5859898924827576
  Cosine similarity with woman: 0.667400598526001
  Cosine similarity with he: 0.5594150424003601
  Cosine similarity with she: 0.6244168281555176
Job: scientist
  Cosine similarity with man: 0.5397614240646362
  Cosine similarity with woman: 0.6615307331085205
  Cosine similarity with he: 0.5063120126724243
  Cosine similarity with she: 0.5763662457466125

진짜 답이 안보이네요...

 

일단 여기까지는 미스트랄 모델이었으니 다른 모델로 변경해보겠습니다.

 

이 밑에 부턴 GPT-2 small 모델입니다.

Job	Gender Dominance	Female Percentage	Cosine Similarity with Woman	Cosine Similarity with Man
skincare specialist	Female	98.2	0.195972	0.204778
kindergarten teacher	Female	96.8	0.220412	0.233106
childcare worker	Female	94.6	0.21226	0.196284
secretary	Female	92.5	0.317799	0.327826
hairstylist	Female	92.4	0.197492	0.212216
dental assistant	Female	92	0.292716	0.315205
nurse	Female	91.3	0.367193	0.412158
school psychologist	Female	90.4	0.252426	0.209854
receptionist	Female	90	0.291924	0.294703
vet	Female	89.8	0.41986	0.476376
nutritionist	Female	89.6	0.28915	0.251744
maid	Female	88.7	0.414578	0.406934
therapist	Female	87.1	0.288946	0.296861
social worker	Female	86.8	0.264815	0.248089
sewer	Female	86.5	0.320507	0.360573
paralegal	Female	84.8	0.195828	0.215684
library assistant	Female	84.2	0.18374	0.167661
interior designer	Female	83.8	0.292864	0.308887
manicurist	Female	83	0.313635	0.443759
special education teacher	Female	82.8	0.211079	0.203801
police officer	Male	15.8	0.247464	0.196906
taxi driver	Male	12	0.253791	0.227682
computer architect	Male	11.8	0.229394	0.219098
mechanical engineer	Male	9.4	0.301834	0.319501
truck driver	Male	7.9	0.310374	0.335189
electrical engineer	Male	7	0.242503	0.236696
landscaping worker	Male	6.2	0.191117	0.189157
pilot	Male	5.3	0.317637	0.349418
repair worker	Male	5.1	0.214539	0.207093
firefighter	Male	5.1	0.329568	0.336709
construction worker	Male	4.2	0.301041	0.298139
machinist	Male	3.4	0.246829	0.273729
aircraft mechanic	Male	3.2	0.278695	0.294949
carpenter	Male	3.1	0.251271	0.283897
roofer	Male	2.9	0.324939	0.361726
brickmason	Male	2.2	0.201017	0.226547
plumber	Male	2.1	0.308215	0.334688
electrician	Male	1.7	0.205533	0.195738
vehicle technician	Male	1.2	0.263219	0.261841
crane operator	Male	1.1	0.264234	0.276296

뭔가 고르게 분포된 이 느낌....

확실하게 남자가 많은 직업은 남자한테 편향적이다! 이런 것은 아니네요

 

TSNE와 UMAP 또한 알 수가 없네요....

Job: doctor
  Cosine similarity with man: 0.3023069500923157
  Cosine similarity with woman: 0.3201085031032562
  Cosine similarity with he: 0.235477015376091
  Cosine similarity with she: 0.2189578413963318
Job: nurse
  Cosine similarity with man: 0.4121583104133606
  Cosine similarity with woman: 0.36719265580177307
  Cosine similarity with he: 0.34656408429145813
  Cosine similarity with she: 0.3177404999732971
Job: engineer
  Cosine similarity with man: 0.2845001220703125
  Cosine similarity with woman: 0.28417065739631653
  Cosine similarity with he: 0.21660952270030975
  Cosine similarity with she: 0.199890598654747
Job: teacher
  Cosine similarity with man: 0.405282199382782
  Cosine similarity with woman: 0.4033980965614319
  Cosine similarity with he: 0.3337845802307129
  Cosine similarity with she: 0.3021133244037628
Job: scientist
  Cosine similarity with man: 0.23807066679000854
  Cosine similarity with woman: 0.26343828439712524
  Cosine similarity with he: 0.1863318830728531
  Cosine similarity with she: 0.17157433927059174

흠.....

 

그럼 이번엔 SAE layer의 다음 레이어인 8번 레이어로 가보겠습니다.

Job	Gender Dominance	Female Percentage	Cosine Similarity with Woman	Cosine Similarity with Man
skincare specialist	Female	98.2	0.830806	0.8029
kindergarten teacher	Female	96.8	0.822819	0.789158
childcare worker	Female	94.6	0.868605	0.81984
secretary	Female	92.5	0.81239	0.773908
hairstylist	Female	92.4	0.816338	0.786998
dental assistant	Female	92	0.850209	0.816433
nurse	Female	91.3	0.866288	0.812078
school psychologist	Female	90.4	0.838778	0.792577
receptionist	Female	90	0.844404	0.80642
vet	Female	89.8	0.827386	0.810253
nutritionist	Female	89.6	0.84587	0.802108
maid	Female	88.7	0.89618	0.827164
therapist	Female	87.1	0.852225	0.817951
social worker	Female	86.8	0.826142	0.779234
sewer	Female	86.5	0.833624	0.806207
paralegal	Female	84.8	0.772227	0.765412
library assistant	Female	84.2	0.833513	0.783677
interior designer	Female	83.8	0.827806	0.79507
manicurist	Female	83	0.865115	0.874826
special education teacher	Female	82.8	0.822208	0.772009
police officer	Male	15.8	0.822562	0.784152
taxi driver	Male	12	0.838859	0.810866
computer architect	Male	11.8	0.840573	0.813889
mechanical engineer	Male	9.4	0.869139	0.851203
truck driver	Male	7.9	0.859283	0.838243
electrical engineer	Male	7	0.810291	0.77179
landscaping worker	Male	6.2	0.833979	0.809405
pilot	Male	5.3	0.842505	0.814605
repair worker	Male	5.1	0.859992	0.819944
firefighter	Male	5.1	0.860235	0.832443
construction worker	Male	4.2	0.844144	0.809736
machinist	Male	3.4	0.833735	0.833825
aircraft mechanic	Male	3.2	0.847732	0.824421
carpenter	Male	3.1	0.805981	0.79008
roofer	Male	2.9	0.794447	0.777306
brickmason	Male	2.2	0.814994	0.809293
plumber	Male	2.1	0.8419	0.810002
electrician	Male	1.7	0.847812	0.820417
vehicle technician	Male	1.2	0.837883	0.809716
crane operator	Male	1.1	0.85486	0.827914

ㅋㅋㅋㅋㅋ.ㅋ.....

이번에도 여자가 다 높게 나오네요 ㅎㅎ.......

TSNE 시각화에선 그래도 남자와 여자가 군집화 된 것을 볼 수 있습니다...?

UMAP도 구분 된 것 같은 모습으로 보이네요 

 

그럼 이번엔 PCA를 진행 후 COS 유사도를 진행해보겠습니다.

Job	Gender Dominance	Female Percentage	Cosine Similarity with Woman	Cosine Similarity with Man
skincare specialist	Female	98.2	-0.33386	-0.24223
kindergarten teacher	Female	96.8	-0.29054	-0.20816
childcare worker	Female	94.6	0.090889	-0.07886
secretary	Female	92.5	-0.05085	-0.08793
hairstylist	Female	92.4	-0.22697	-0.09444
dental assistant	Female	92	-0.13962	-0.20142
nurse	Female	91.3	0.279483	0.0298
school psychologist	Female	90.4	-0.03985	-0.18175
receptionist	Female	90	-0.09819	-0.12192
vet	Female	89.8	-0.02614	0.058945
nutritionist	Female	89.6	-0.11909	-0.24243
maid	Female	88.7	0.515144	0.152515
therapist	Female	87.1	-0.00373	-0.05599
social worker	Female	86.8	-0.04072	-0.14932
sewer	Female	86.5	-0.01937	0.005589
paralegal	Female	84.8	-0.11654	0.030654
library assistant	Female	84.2	-0.10128	-0.23378
interior designer	Female	83.8	-0.12399	-0.14279
manicurist	Female	83	0.099113	0.419571
special education teacher	Female	82.8	-0.22952	-0.35435
police officer	Male	15.8	0.073392	-0.00112
taxi driver	Male	12	-0.16622	-0.15999
computer architect	Male	11.8	-0.04169	-0.01093
mechanical engineer	Male	9.4	-0.01881	0.112229
truck driver	Male	7.9	-0.01101	0.06296
electrical engineer	Male	7	-0.32179	-0.31583
landscaping worker	Male	6.2	-0.34693	-0.23745
pilot	Male	5.3	0.043929	0.0348
repair worker	Male	5.1	0.041328	-0.10179
firefighter	Male	5.1	-0.0164	-0.03699
construction worker	Male	4.2	-0.19002	-0.15065
machinist	Male	3.4	-0.07069	0.186568
aircraft mechanic	Male	3.2	-0.17663	-0.12196
carpenter	Male	3.1	-0.31704	-0.10918
roofer	Male	2.9	-0.15045	-0.06005
brickmason	Male	2.2	-0.09472	0.100032
plumber	Male	2.1	-0.04168	-0.08663
electrician	Male	1.7	-0.17601	-0.10355
vehicle technician	Male	1.2	-0.12613	-0.07126
crane operator	Male	1.1	-0.18231	-0.10622

여기도 편향이 그렇게 강하다! 라고 볼 순 없네요,,,>?ㅜㅜㅜㅜ

 

또 다른 아이디어가 떠오르면 작성하겠습니다....