Processing math: 100%
+ - 0:00:00
Notes for current slide
Notes for next slide

Current Research Trends

Curso JAE ICMAT 2021

Roi Naveiro y David Ríos Insua

2021-06-28

1 / 18

Modern research (ML and Bayes)

  • NLP.

  • RL.

  • Interpretability and Explainability.

  • Fairness.

  • Causal Inference.

  • Scalable Bayesian Inference.

  • Adversarial ML.

2 / 18

NLP - Arms race

  • Pre-trained language models + fine tuning

  • Wu Dao 1.75 trillion parameters!!
3 / 18

NLP - Research

  • Architecture design.

  • Applications.

    1. Dialog systems.
    2. Emotion recognition.
    3. Computer vision + NLP.s
  • Theoretical results?

  • Non democratic... only accesible to people with data and computational power (google, facebook, ...)

4 / 18

Reinforcenment Learning

  • According to a lot of people, the way to AGI.

  • Multi-agent RL.

  • Many applications.

5 / 18

Interpretability & Explainability

  • Black-box algorithms are not always applicable.

  • We need algorithms to tell us why the make the decisions.

  • Interpretability: use algorithms that are interpretable per se.

  • Explainability: use black box algorithms and ellaborate explanations a posteriori.

    • The do not explain reality!!
  • Rashomon sets

Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206-215.

6 / 18

Fairness - Bias in ML

  • Data is biased algorithms are biased

7 / 18

Fairness - Bias in ML

8 / 18

Research on Fairness

  • Definition.

  • How to guarantee that models are not using, e.g., racial information?

  • Uncovering biases.

  • Interpretability...

9 / 18

Causal Inference

  • Data will provide associations between variables...

  • ... to do science we need to measure causation!

10 / 18

Correlation is not causation

11 / 18

Causal Inference

  • The model is key!

12 / 18

Causal Inference

  • Can we determine cause-effect relations using data?

  • Under some circumstances, and some assumptions, yes.

  • Essentially, we assume that we observe every confounder.

  • Research on how to adjust for every confounder...

  • And on applications.

13 / 18

Scalable Bayesian Inference

14 / 18

Adversarial Machine Learning

  • Automation using ML increases... attempts to cheat ML systems increases.

15 / 18

Adversarial Machine Learning

16 / 18

Adversarial Machine Learning

  • Make ML algorithms robust against likely attacks.

  • Idea: formalize confrontation between algorithm and adversary as a game.

  • Compute Nash equilibrium and use it as robust solution.

  • Common Knowledge?

  • Bayesian games, adversarial risk analysis?

17 / 18

18 / 18

Modern research (ML and Bayes)

  • NLP.

  • RL.

  • Interpretability and Explainability.

  • Fairness.

  • Causal Inference.

  • Scalable Bayesian Inference.

  • Adversarial ML.

2 / 18
Paused

Help

Keyboard shortcuts

, , Pg Up, k Go to previous slide
, , Pg Dn, Space, j Go to next slide
Home Go to first slide
End Go to last slide
Number + Return Go to specific slide
b / m / f Toggle blackout / mirrored / fullscreen mode
c Clone slideshow
p Toggle presenter mode
t Restart the presentation timer
?, h Toggle this help
Esc Back to slideshow