Search

The goal of explainable Artificial Intelligence (XAI) is to generate human-interpretable explanations, but there are no computationally precise theories of how humans interpret AI generated explanations. The lack of theory means that validation of XAI must be done empirically, on a case-by-case basis, which prevents systematic theory-building in XAI. We propose a psychological theory of how humans draw conclusions from saliency maps, the most common form of XAI explanation, which for the first time allows for precise prediction of explainee inference conditioned on explanation. Our theory posits that absent explanation humans expect the AI to make similar decisions to themselves, and that they interpret an explanation by comparison to the explanations they themselves would give. Comparison is formalized via Shepard's universal law of generalization in a similarity space, a classic theory from cognitive science. A pre-registered user study on AI image classifications with saliency map explanations demonstrate that our theory quantitatively matches participants' predictions of the AI., Comment: 15 pages, 3 figures, ICML 2022

Theory-of-Mind inference is natural for humans but poses significant computational challenges. The core difficulty can be traced back to the exponential growth of paths to consider in planning given a mental state. In this paper we tackle this problem in a search-and-rescue task implemented in Minecraft. Our goal is to infer differences in knowledge from participants' continuous-time trajectory. By abstracting the spatio-temporal state space and the reward function together, we surface natural decision points, on which we compare the participants' behavior to myopic rational agents of varying knowledgeability. Collectively, the abstraction and rational agent analysis yield successful inference of participants' knowledge states and reveal distinct patterns of their exploratory behavior.

Adversarial images highlight how vulnerable modern image classifiers are to perturbations outside of their training set. Human oversight might mitigate this weakness, but depends on humans understanding the AI well enough to predict when it is likely to make a mistake. In previous work we have found that humans tend to assume that the AI's decision process mirrors their own. Here we evaluate if methods from explainable AI can disrupt this assumption to help participants predict AI classifications for adversarial and standard images. We find that both saliency maps and examples facilitate catching AI errors, but their effects are not additive, and saliency maps are more effective than examples.

Deep Gaussian Processes (DGPs) combine the expressiveness of Deep Neural Networks (DNNs) with quantified uncertainty of Gaussian Processes (GPs). Expressive power and intractable inference both result from the non-Gaussian distribution over composition functions. We propose interpretable DGP based on approximating DGP as a GP by calculating the exact moments, which additionally identify the heavy-tailed nature of some DGP distributions. Consequently, our approach admits interpretation as both NNs with specified activation functions and as a variational approximation to DGP. We identify the expressivity parameter of DGP and find non-local and non-stationary correlation from DGP composition. We provide general recipes for deriving the effective kernels for DGP of two, three, or infinitely many layers, composed of homogeneous or heterogeneous kernels. Results illustrate the expressiveness of our effective kernels through samples from the prior and inference on simulated and real data and demonstrate advantages of interpretability by analysis of analytic forms, and draw relations and equivalences across kernels., Preprint with 12 pages and 3 figures. The updated version (Oct 9 2019) consider the second and fourth moments, inspecting the heavy-tailed nature of DGP distribution, justifying the validity of approximating DGP as GP. A connection with the expressivity parameter in Poole et al NIPS paper is also added. New reference and 4th moments of SC[]

DNA replication in higher organisms starts at many places across the genome and throughout S (synthesis) phase. In order to understand replication in eukaryotes, one needs to know not only how the replicative machinery functions on the molecular level but also how the machinery is organized genome wide to ensure complete duplication. Over the past fifteen years, advances in technology have allowed researchers to perform genome-wide experiments that probe the state of replication in many organisms. These datasets make possible quantitative modelling of the replication process. The kinetics of DNA replication is formally analogous to a physical phase-transformation process. In replication, the DNA is transformed from a "non-replicated'' phase to a "replicated'' phase, just as freezing water is transformed from a liquid phase to a solid phase. Using this analogy, we map the replication process onto a stochastic nucleation-and-growth model introduced in statistical physics to describe first-order phase transitions. Extending the model, we develop a mathematical framework that is flexible enough to describe the kinetics of replication in eukaryotes. We present three applications of our theory: 1) We apply the theory to a recent dataset on budding yeast to extract its genome-wide replication program. Based on this study, we give the first proposal to explain how the temporal aspect of the replication program can be controlled mechanistically. 2) We address the "random-completion problem,'' which asks how replication-completion times can be controlled when replication starts at random places and times. We find that the strategy adopted in frog embryos to solve the problem also nearly minimizes the use of certain replicative machinery. 3) We study possible ways to extract information from a popular technique used to probe replication in multicellular eukaryotes, ranging from worms to humans. We show preliminary results that can be extended to real experiments in the near future.