ISSP 2030 Spring AI Forum Week 12

09 Apr 2021 - Arun Balajiee

Date: 04/09/2021

Speaker: Bryan Andrews

Title: Score-Based Causal Discovery

In this talk Andrews talked about discovering causal relationships using Probabilistic Conditional Independence. Specifically, with graphical representation of nodes as the different factors that could affect one another, and the edges in the graph to represent their relations, the method he discussed was to visualize and identify causal relationships. For automated causal discovery, each edge in the causal graph is assumed to satisfy two properties - markov property and faithfulness. Further, the output inferred from the graph has two components – the markov equivalence class and report invariant features. True casual relations are the edges in the graph that repeat across different markov equivalence classes. Using conditional parameteric constraints, the graph can be represented as a plane in 3 dimensions. Using Directed Acyclic Graphs and Bayesian Networks (Graphical Markov Models)

An issue with this method of causal discovery is causal “sufficiency” (identifying only directly identifiable relations and not the latent/confounding relations) as well as selection bias. To enchance some shortcomings of these graphs along with cases where false relations are not learnt, he discussed the use of Maximal Ancestral Graphs (MAGs) to identify causal and confounding relationships. Finally, Andrews discussed the feasibility of this method by applying it to an environmental and clinical dataset which identify the relationships between harmful gases such as carbon monoxide, sulphur, ozone, nitrogen dioxide with cardiovascular and respiratory diseases.

The takeaways from his talk were the use of applications of discrete mathematics in data mining of causal relations using Bayesian networks with markov probabilistic models. USing these causal relations one can be applied in non-experimental data, which contains a lot of observational datapoints and in studies where randomized control trials is not possible where these relations can act as gold standard

Speaker: Daniel Steinberg

Title: Characterizing the hidden layer representation impact of FGSM adversarial attacks

In his talk, Steinberg introduced the ideas of using adversarial examples for neural networks to learn for image classification tasks – specifically convolutional neural networks and use the principle of Fast Gradient Sign Method (FGSM) to use as an optimization of the learning loss and improve the network to handle noises in the images. Specifically the research questions were two fold - understanding the effects of image morphing (adding noise to the image so it is misclassified by the neural network) using FGSM to create perturbations to change hidden layer representations of image. The second research question was to consider patterns of activity across input image and the effect of image perturbations on the representation of the image in the hidden layers.

To address the first research question, the team identified the change in image represetion in the hidden layers of ResNet-18. They observe the changes in the confusion matrix for the classification task based on image perturbations. They also used violin plots to notice the change in the one-hot representation of the output with the change in the input image to visualize the image representation acorss the hidden layers of the hidden layers. Further, they used UMAP function to project the representations of the hidden layers onto a 2D space.

For the second research question to understand if the neural network can identify an image perturbation, they use an output unit for the image classifier that identifies if the image is adversial or not with 76% accuracy or 0.84 AUC score

The takeaways from the talk was that neural networks are very sensitive to small changes in input by noise and to detect this noise is a challenging task. Being able to detect this and embed the information as a gradient function that can be used in backpropagation is a another challenging tasks. Steinberg addresses these two problems in his work on adversarial input data for neural networks specifcally in the case of CNNs for image classification tasks. As a scope for future work, Steinberg looks to explore the localization of image perturbations (where in the image there is noise), ways to visualize these pertubations are represented by the hidden layers and methods to compare adversarial and non-adversarial examples using ablation studies. Further, there is always a trade-off between the performance of the network vs handling the adversarial examples for input.