-
A Unified Approach Towards Active Learning and Out-of-Distribution Detection
Authors:
Sebastian Schmidt,
Leonard Schenk,
Leo Schwinn,
Stephan Günnemann
Abstract:
When applying deep learning models in open-world scenarios, active learning (AL) strategies are crucial for identifying label candidates from a nearly infinite amount of unlabeled data. In this context, robust out-of-distribution (OOD) detection mechanisms are essential for handling data outside the target distribution of the application. However, current works investigate both problems separately…
▽ More
When applying deep learning models in open-world scenarios, active learning (AL) strategies are crucial for identifying label candidates from a nearly infinite amount of unlabeled data. In this context, robust out-of-distribution (OOD) detection mechanisms are essential for handling data outside the target distribution of the application. However, current works investigate both problems separately. In this work, we introduce SISOM as the first unified solution for both AL and OOD detection. By leveraging feature space distance metrics SISOM combines the strengths of the currently independent tasks to solve both effectively. We conduct extensive experiments showing the problems arising when migrating between both tasks. In these evaluations SISOM underlined its effectiveness by achieving first place in two of the widely used OpenOOD benchmarks and second place in the remaining one. In AL, SISOM outperforms others and delivers top-1 performance in three benchmarks
△ Less
Submitted 25 May, 2024; v1 submitted 18 May, 2024;
originally announced May 2024.
-
Partial correlation graphs for continuous-parameter time series
Authors:
Vicky Fasen-Hartmann,
Lea Schenk
Abstract:
In this paper, we establish the partial correlation graph for multivariate continuous-time stochastic processes, assuming only that the underlying process is stationary and mean-square continuous with expectation zero and spectral density function. In the partial correlation graph, the vertices are the components of the process and the undirected edges represent partial correlations between the ve…
▽ More
In this paper, we establish the partial correlation graph for multivariate continuous-time stochastic processes, assuming only that the underlying process is stationary and mean-square continuous with expectation zero and spectral density function. In the partial correlation graph, the vertices are the components of the process and the undirected edges represent partial correlations between the vertices. To define this graph, we therefore first introduce the partial correlation relation for continuous-time processes and provide several equivalent characterisations. In particular, we establish that the partial correlation relation defines a graphoid. The partial correlation graph additionally satisfies the usual Markov properties and the edges can be determined very easily via the inverse of the spectral density function. Throughout the paper, we compare and relate the partial correlation graph to the mixed (local) causality graph of Fasen-Hartmann and Schenk (2023a). Finally, as an example, we explicitly characterise and interpret the edges in the partial correlation graph for the popular multivariate continuous-time AR (MCAR) processes.
△ Less
Submitted 30 January, 2024;
originally announced January 2024.
-
Mixed causality graphs for continuous-time state space models and orthogonal projections
Authors:
Vicky Fasen-Hartmann,
Lea Schenk
Abstract:
In this paper we derive (local) causality graphs for the popular continuous-time state space models, including in particular multivariate continuous-time ARMA (MCARMA) processes. In these (local) causality graphs, vertices represent the components of the process, directed edges between the vertices indicate causal influences, and undirected edges indicate contemporaneous uncorrelatednesses between…
▽ More
In this paper we derive (local) causality graphs for the popular continuous-time state space models, including in particular multivariate continuous-time ARMA (MCARMA) processes. In these (local) causality graphs, vertices represent the components of the process, directed edges between the vertices indicate causal influences, and undirected edges indicate contemporaneous uncorrelatednesses between the component processes. We present sufficient criteria for state space models to satisfy the assumptions of Fasen-Hartmann and Schenk (2023) so that the (local) causality graphs are well defined and various causal Markov properties hold. Both directed and undirected edges in these graphs are characterised by orthogonal projections on well-defined linear spaces. To compute these orthogonal projections, we use the unique controller canonical form of a state space model, which exists under mild assumptions, to recover the input process from the output process. We are then able to derive some alternative representations of the output process and its highest derivative. Finally, we apply these representations to calculate the necessary orthogonal projections, which culminate in the characterisations of the edges in the (local) causality graph. These characterisations are interpretatively meaningful and are given by the parameters of the controller canonical form and the covariance matrix of the driving Lévy process.
△ Less
Submitted 8 November, 2023;
originally announced November 2023.
-
Mixed causality graphs for continuous-time stationary processes
Authors:
Vicky Fasen-Hartmann,
Lea Schenk
Abstract:
In this paper, we introduce different concepts of Granger non-causality and contemporaneous uncorrelation for stationary continuous-time processes to model the different dependencies between the component series of multivariate time series models. Several equivalent characterisations for the different definitions are given, in particular by linear projections. We then define two mixed graphs based…
▽ More
In this paper, we introduce different concepts of Granger non-causality and contemporaneous uncorrelation for stationary continuous-time processes to model the different dependencies between the component series of multivariate time series models. Several equivalent characterisations for the different definitions are given, in particular by linear projections. We then define two mixed graphs based on different definitions of causality and contemporaneous uncorrelation, the (mixed) causality graph and the local (mixed) causality graph, to visualise and to analyse the different dependencies in stationary continuous-time processes. In these graphs, the components of the process are represented by vertices, directed edges between the vertices indicate causal influences and undirected edges indicate contemporaneous uncorrelation between the component processes. Further, we introduce various notions of causal Markov properties in analogy to Eichler (2012), which relate the different dependence structures of subprocesses, and we derive sufficient criteria for the (local) causality graph to satisfy them. Finally, as an example, the popular multivariate continuous-time AR (MCAR) processes satisfy our assumptions. For MCAR processes we show that the (local) causality graphs can be characterised explicitly by the model parameters.
△ Less
Submitted 17 August, 2023;
originally announced August 2023.
