Trustworthy AI in Julia
The Alan Turing Institute
2024-05-08
Introduction
Economist by training, previously Bank of England, currently 3rd year PhD in Trustworthy AI @ TU Delft.
Taija
Taija is the organization that hosts software geared towards Trustworthy Artificial Intelligence in Julia.
Counterfactual Explanations
CounterfactualExplanations.jl: A package for Counterfactual Explanations and Algorithmic Recourse in Julia.
Background
\[ \begin{aligned} \min_{\mathbf{Z}^\prime \in \mathcal{Z}^L} \{ {\text{yloss}(M_{\theta}(f(\mathbf{Z}^\prime)),\mathbf{y}^+)} + \lambda {\text{cost}(f(\mathbf{Z}^\prime)) } \} \end{aligned} \]
Counterfactual Explanations (CE) explain how inputs into a model need to change for it to produce different outputs.
📜 Altmeyer, Deursen, et al. (2023) @ JuliaCon 2022.
Pick your Poison
All of these counterfactuals are valid explanations for the model’s prediction.
Which one would you pick?
Figure 2: Turning a 9 into a 7: Counterfactual explanations for an image classifier produced using Wachter (Wachter, Mittelstadt, and Russell 2017), Schut (Schut et al. 2021) and REVISE (Joshi et al. 2019).
ECCCos from the Black-Box
📜 Altmeyer, Farmanbar, et al. (2023) @ AAAI 2024
ECCCo objective1:
\[ \begin{aligned} & \min_{\mathbf{Z}^\prime \in \mathcal{Z}^L} \{ {L_{\text{clf}}(f(\mathbf{Z}^\prime);M_{\theta},\mathbf{y}^+)}+ \lambda_1 {\text{cost}(f(\mathbf{Z}^\prime)) } \\ &+ \lambda_2 \mathcal{E}_{\theta}(f(\mathbf{Z}^\prime)|\mathbf{y}^+) + \lambda_3 \Omega(C_{\theta}(f(\mathbf{Z}^\prime);\alpha)) \} \end{aligned} \]
Faithful Counterfactuals
Conformal Prediction
ConformalPrediction.jl: A package for Predictive Uncertainty Quantification through Conformal Prediction for Machine Learning models trained in MLJ.
Background
Intuitively, CP works under the premise of turning heuristic notions of uncertainty into rigorous uncertainty estimates through repeated sampling or the use of dedicated calibration data.
Conformal Prediction in action: prediction intervals at varying coverage rates. As coverage grows, so does the width of the prediction interval.
Examples
- A simple usage example.
- Conformalizing an image classifier here.
- How to build a conformal chatbot here.
ConformalPrediction.jl meets SymbolicRegression.jl.
Laplace Redux
LaplaceRedux.jl: A package for Effortless Bayesian Deep Learning through Laplace Approximation for Flux.jl neural networks.
Background
We want BMA for neural networks underspecified by the data (Wilson 2020),
\[ p(y|x,\mathcal{D}) = \int p(y|x,\theta)p(\theta|\mathcal{D})d\theta \tag{1}\]
where \(p(y|x,\theta)\) is the likelihood, \(p(\theta|\mathcal{D})\) is the posterior and \(\mathcal{D}\) is the training data.
- Problem 😢: Intractable posterior \(p(\theta|\mathcal{D})\).
- Convention 🤔: Use MAP estimate \(p(y|x,\mathcal{D}) \approx p(y|x,\hat\theta)\) with \(\hat\theta=\arg\max_{\theta}p(\theta|\mathcal{D})\).
- Idea 💡: Taylor approximation at the mode amounts to multivariate Gaussian centered around MAP.
Examples
- A basic usage example.
- A binary classification example with prior tuning through empirical Bayes.
- Counterfactual explanations with Laplace Redux here.
Joint Energy Models
JointEnergyModels.jl: A package for Joint Energy Models and Energy-Based Models in Julia.
Background
Joint Energy Models (JEM) are a class of energy-based models that learn a joint distribution over inputs and outputs (Grathwohl et al. 2020).
Training JEMs.
Package
Questions?