O slideshow foi denunciado.

Rsqrd AI: Incorporating Priors with Feature Attribution on Text Classification

0

Compartilhar

1 de 13
1 de 13

Rsqrd AI: Incorporating Priors with Feature Attribution on Text Classification

0

Compartilhar

Baixar para ler offline

Descrição

In this talk, Frederick Liu from Google talks about feature attribution on text classification.

Presented on 07/17/2019

**These slides are from a talk given at Rsqrd AI. Learn more at rsqrdai.org**

Transcrição

  1. 1. Proprietary + Confidential Frederick Liu 7/17/19 @ Robust AI Incorporating priors with feature attribution on text classification
  2. 2. Proprietary + ConfidentialProprietary + Confidential Machine learning .6 .2 .1 .3 .7 .1 .8 .7 .5.3 .1 .4 .9 .2 .0 .6.8 .2.1.6Toxic … … … … … … … Neutral … … … … … … Toxic … … … … … … … Toxic … … … … … … Neutral … … … … … … Training Inference Gay pride is in June. .6 .2 .1 .3 .7 .1 .8 .7 .5.3 .1 .4 .9 .2 .0 .6.8 .2.1.6 95% Toxic
  3. 3. Proprietary + ConfidentialProprietary + Confidential Machine learning + Explainability .6 .2 .1 .3 .7 .1 .8 .7 .5.3 .1 .4 .9 .2 .0 .6.8 .2.1.6Toxic … … … … … … … Neutral … … … … … … Toxic … … … … … … … Toxic … … … … … … Neutral … … … … … … Training Inference Gay pride is in June. .6 .2 .1 .3 .7 .1 .8 .7 .5.3 .1 .4 .9 .2 .0 .6.8 .2.1.6 95% Toxic Gay Pride is in June 90% 1% 1% 1% 2%
  4. 4. Proprietary + ConfidentialProprietary + Confidential Machine learning + Regularization Toxic … … … … … … … Neutral … … … … … … Toxic … … … … … … … Toxic … … … … … … Neutral … … … … … … Training Inference Gay pride is in June. .5 .2 .1 .3 .5 .1 .5 .5 .5.3 .1 .4 .5 .2 .0 .5.5 .2.1.5 85% Toxic .5 .2 .1 .3 .5 .1 .5 .5 .5.3 .1 .4 .5 .2 .0 .5.5 .2.1.5
  5. 5. Proprietary + ConfidentialProprietary + Confidential Machine learning + Regularization + Explainability Toxic … … … … … … … Neutral … … … … … … Toxic … … … … … … … Toxic … … … … … … Neutral … … … … … … Training Inference Gay pride is in June. 15% Toxic Gay Pride is in June He is an impolite gay 0% .7 .2 .1 .3 .7 .1 .7 .5.3.4 .9 .2 .1 .6.8 .20. 1 .1 .2 .5 .7 .2 .1 .3 .7 .1 .7 .5.3.4 .9 .2 .1 .6.8 .20. 1 .1 .2 .5 + person
  6. 6. Proprietary + ConfidentialProprietary + Confidential Regularizing + Explainability → Controllability .6 .2 .1 .3 .7 .1 .8 .7 .5.3 .1 .4 .9 .2 .0 .6.8 .2 .1.6 Explanation
  7. 7. Proprietary + ConfidentialProprietary + Confidential Regularizing + Explainability → Controllability .6 .2 .1 .3 .7 .1 .3 .7 .5.3 .9 .4 .9 .2 .0 .6.8 .4 .1.7 Explanation More Red! Less Green!
  8. 8. Proprietary + ConfidentialProprietary + Confidential Explainability - Integrated Gradients Link to paper - https://arxiv.org/pdf/1703.01365.pdf
  9. 9. Proprietary + ConfidentialProprietary + Confidential Explainability + Regularization
  10. 10. Proprietary + ConfidentialProprietary + Confidential Results - Classification Metric
  11. 11. Proprietary + ConfidentialProprietary + Confidential Results - Fairness Metric
  12. 12. Proprietary + ConfidentialProprietary + Confidential Results - Shift in embedding
  13. 13. Proprietary + Confidential Thank You Link to paper - https://arxiv.org/pdf/1906.08286.pdf Sign up if you want to know more: bit.ly/model-interpret-interest

Descrição

In this talk, Frederick Liu from Google talks about feature attribution on text classification.

Presented on 07/17/2019

**These slides are from a talk given at Rsqrd AI. Learn more at rsqrdai.org**

Transcrição

  1. 1. Proprietary + Confidential Frederick Liu 7/17/19 @ Robust AI Incorporating priors with feature attribution on text classification
  2. 2. Proprietary + ConfidentialProprietary + Confidential Machine learning .6 .2 .1 .3 .7 .1 .8 .7 .5.3 .1 .4 .9 .2 .0 .6.8 .2.1.6Toxic … … … … … … … Neutral … … … … … … Toxic … … … … … … … Toxic … … … … … … Neutral … … … … … … Training Inference Gay pride is in June. .6 .2 .1 .3 .7 .1 .8 .7 .5.3 .1 .4 .9 .2 .0 .6.8 .2.1.6 95% Toxic
  3. 3. Proprietary + ConfidentialProprietary + Confidential Machine learning + Explainability .6 .2 .1 .3 .7 .1 .8 .7 .5.3 .1 .4 .9 .2 .0 .6.8 .2.1.6Toxic … … … … … … … Neutral … … … … … … Toxic … … … … … … … Toxic … … … … … … Neutral … … … … … … Training Inference Gay pride is in June. .6 .2 .1 .3 .7 .1 .8 .7 .5.3 .1 .4 .9 .2 .0 .6.8 .2.1.6 95% Toxic Gay Pride is in June 90% 1% 1% 1% 2%
  4. 4. Proprietary + ConfidentialProprietary + Confidential Machine learning + Regularization Toxic … … … … … … … Neutral … … … … … … Toxic … … … … … … … Toxic … … … … … … Neutral … … … … … … Training Inference Gay pride is in June. .5 .2 .1 .3 .5 .1 .5 .5 .5.3 .1 .4 .5 .2 .0 .5.5 .2.1.5 85% Toxic .5 .2 .1 .3 .5 .1 .5 .5 .5.3 .1 .4 .5 .2 .0 .5.5 .2.1.5
  5. 5. Proprietary + ConfidentialProprietary + Confidential Machine learning + Regularization + Explainability Toxic … … … … … … … Neutral … … … … … … Toxic … … … … … … … Toxic … … … … … … Neutral … … … … … … Training Inference Gay pride is in June. 15% Toxic Gay Pride is in June He is an impolite gay 0% .7 .2 .1 .3 .7 .1 .7 .5.3.4 .9 .2 .1 .6.8 .20. 1 .1 .2 .5 .7 .2 .1 .3 .7 .1 .7 .5.3.4 .9 .2 .1 .6.8 .20. 1 .1 .2 .5 + person
  6. 6. Proprietary + ConfidentialProprietary + Confidential Regularizing + Explainability → Controllability .6 .2 .1 .3 .7 .1 .8 .7 .5.3 .1 .4 .9 .2 .0 .6.8 .2 .1.6 Explanation
  7. 7. Proprietary + ConfidentialProprietary + Confidential Regularizing + Explainability → Controllability .6 .2 .1 .3 .7 .1 .3 .7 .5.3 .9 .4 .9 .2 .0 .6.8 .4 .1.7 Explanation More Red! Less Green!
  8. 8. Proprietary + ConfidentialProprietary + Confidential Explainability - Integrated Gradients Link to paper - https://arxiv.org/pdf/1703.01365.pdf
  9. 9. Proprietary + ConfidentialProprietary + Confidential Explainability + Regularization
  10. 10. Proprietary + ConfidentialProprietary + Confidential Results - Classification Metric
  11. 11. Proprietary + ConfidentialProprietary + Confidential Results - Fairness Metric
  12. 12. Proprietary + ConfidentialProprietary + Confidential Results - Shift in embedding
  13. 13. Proprietary + Confidential Thank You Link to paper - https://arxiv.org/pdf/1906.08286.pdf Sign up if you want to know more: bit.ly/model-interpret-interest

Mais Conteúdo rRelacionado

Audiolivros relacionados

Gratuito durante 30 dias do Scribd

Ver tudo

×