1.: Illustration of an Unlabeled Graph Consisting of 5 Nodes and 6 Edges. 

Contexts in source publication

Context 1

... order to enable the user to comprehend the retrieval systems result finding process, a dedicated mapping view is integrated into the WebUI (see fig. 5.20). This full screen view allows for comparing the entered floor plan to a selected search result. Both floor plans are rendered in the background, and an overlay allows to see the semantic fingerprints of both floor plans. The mapping (displayed in red arrows) shows the mapping as generated by the AP. All of the mentioned elements can be switched on and of by dedicated buttons, the type of fingerprint can be altered by a slider. An also switchable mapping info box displays how the AS has found the mapping. ...

Context 2

... its simplest form, an RNN scans a the input data in one direction (most intuitively forwards in time or along an space-related axis). In the following, the computation steps are referred to as time steps, although the scanned data are not necessarily related to a physical time domain (e.g. when considering a image data). Consequently, at each time step, only information up to that time step is available to the network. For some applications, information from future time steps would be helpful to solve a problem. As a makeshift, the target value of the RNN could be shifted back in time by n time steps, resulting in the availability of the n future time steps in every point of time. As pointed out in [49], this strategy only works for a limited amount of time steps. In order to solve this problem (and hence allow for an availability of the input information of all time steps in all computation steps) the bidirectional recurrent network structure has been proposed (see figure 2.11) for an overview). In this structure, there is an RNN for each direction of time. The outputs of both RNN have to be merged by a third neural network, for example an MLP (one layer is usually ...

Context 3

... neural networks (ANNs) are a biologically motivated approach in the field of artificial intelligence (AI). There is a wide range of different kinds of ANNs described in literature, some of them are pointed out in more detail below. Generally, ANNs mimic the behavior of neurons of biological organisms. More precisely, the interaction between these units is of most interest here. An ANN usually has a fixed number of inputs and outputs (often described as a vector) and neural units in between. In the neural network, units are connected to each other by so-called weighted connections. The weights of these connections influence how the neural units communicate. These weighted connections are initiated with random values when creating the ANN and adapted during a training process in order to a imprint a desired behavior to the neural network. After such a training process, the ANN can be deployed in a productive situation, where it is used to inference from input. In the following, a trained neural networks is referred to as model. Neurons are information processing cells that exist in a wide range of biological organ- isms. Neurons possess several projections that can roughly be divided into axons and dendrites. Both connect a neuron to other neurons. While dendrites mainly carry the receiver devices of a neuron, axons propagate signals and the axon terminals transmit the cell state signal to other neurons. These inter-cell connections are referred to as synapses. Roughly speaking, the cell state can be described as either activated or in- activated. Whether or not the cell is activated is determined by the incoming signals, where each incoming signal can contribute either to raising the likelihood of an active cell state (activation) or lowering this probability (inhibition). Perceptrons mimic the information processing capabilities of neurons (see fig.2.3). The output of each perceptron is calculated by an activation function applied to the weighted sum of inputs (r is the number of input connections to the ...

Context 4

... The weights of an ANN are usually initialized with random values. During the training process, these weights are altered so that the forward function of the ANN approaches the desired function. One commonly used algorithm to accomplish training is the so-called backpropagation [29] (BP) algorithm. At this point, BP training is shown for MLPs. In order to apply BP, an input is propagated forward through the MLP (see figure 2.5). During forward propagation, for every perceptron inside the MLP the activation a and derivative d of the activation function at the same argument are calculated (a n,k is the activation of the k-th perceptron in the n-th layer, K n is the number of perceptrons in layer n): a n,k = f activation ...

Context 5

... chapter introduces the technical and scientific foundations that are used in the thesis at hand. For each scientific branch that is utilized in this thesis, there is a section in this chapter. The first section introduces the mathematical concept of graphs that are employed in the floor plan retrieval function. Then, artificial neural networks are introduced which are needed in the creativity engine for generating design proposals. Finally, human computer interaction is briefly introduced as a scientific basis for the design and more importantly the analysis of the WebUI as a user interface. Graphs [21] are a mathematical construct (see fig. 2.1 for an example illustration). The idea behind a graph is to model the relation between a set of objects. The objects in such a model are referred to as nodes and the relations are referred to as ...

Context 6

... multi layer perceptron (MLP) consists of a stack of layers of perceptron units, in which the outputs of every layer are fed into the subsequent layer (see fig. 2.4). Hence, there is a full interconnection between the preceptron units of each layer, while the perceptron units of each layer are not connected to each other. The number perceptrons in each layer as well as the number of layers themselves are arbitrary, and appropriate values for these so-called hyperparameters are usually determined by trial-and-error experiments. Likewise, the number of MLP net inputs and output perceptrons is usually determined by the application. Every connection is equipped with a so-called weight determining the influence an outgoing perceptron has to the ingoing. Since two consecutive are assumed to be fully connected, their connections are usually described by a weight matrix. The inputs of the MLP are usually expected to be ranging between zero and ...

Context 7

... answers filled in the questionnaires serves as a primary source of processed infor- mation. The questionnaire was designed to measure the different dimensions of the ISO usability definition (see Section 2.3.1). However, since the questionnaire information only reflects the opinions of the users, this data only approximates the usability dimensions. Generally, this experiment is a comparative study in which two software prototypes were compared to their traditional working method manners; thereby checking the software prototype's user acceptance. The hypothesis of the study was: "The user is able to express his thoughts with the computer tools as good as with the traditional tools." Results The accumulated responses of the participants to the closed questions are shown in fig. 6.1 and fig. 6.2. The effectiveness of the WebUI (i.e. to what degree the participants were able to fulfill their task) was measured indirectly by the questions depicted in fig. 6.1c and fig. 6.2e. The responses to these questions showed that the majority of participants were able to fulfill their tasks at least to a certain degree. The freestyle drawing prototype "Touchtect" from the TU Munich performed slightly better, as shown in fig. 6.1d and fig. 6.2d. The efficiency of the WebUI was vaguely assessed by asking the participants about the perceived time they needed to complete the task (see fig. 6.2c). It is measured more precisely by asking the participants how difficult they perceived the usage of the WebUI ( fig. 6.1a), how exhaused they were by using the WebUI ( fig. 6.1b) and how hindered they were by using the WebUI ( fig. 6.2a). Generally, the results vary widely, but Touchtect performes slightly better ( fig. ...

Context 8

... answers filled in the questionnaires serves as a primary source of processed infor- mation. The questionnaire was designed to measure the different dimensions of the ISO usability definition (see Section 2.3.1). However, since the questionnaire information only reflects the opinions of the users, this data only approximates the usability dimensions. Generally, this experiment is a comparative study in which two software prototypes were compared to their traditional working method manners; thereby checking the software prototype's user acceptance. The hypothesis of the study was: "The user is able to express his thoughts with the computer tools as good as with the traditional tools." Results The accumulated responses of the participants to the closed questions are shown in fig. 6.1 and fig. 6.2. The effectiveness of the WebUI (i.e. to what degree the participants were able to fulfill their task) was measured indirectly by the questions depicted in fig. 6.1c and fig. 6.2e. The responses to these questions showed that the majority of participants were able to fulfill their tasks at least to a certain degree. The freestyle drawing prototype "Touchtect" from the TU Munich performed slightly better, as shown in fig. 6.1d and fig. 6.2d. The efficiency of the WebUI was vaguely assessed by asking the participants about the perceived time they needed to complete the task (see fig. 6.2c). It is measured more precisely by asking the participants how difficult they perceived the usage of the WebUI ( fig. 6.1a), how exhaused they were by using the WebUI ( fig. 6.1b) and how hindered they were by using the WebUI ( fig. 6.2a). Generally, the results vary widely, but Touchtect performes slightly better ( fig. ...

Context 9

... subsection describes how a sequence of feature vectors makes up the description of a floor plan. An illustration for such a sequence for a sample floor plan can be found in fig. A.2. A tag is the smallest unit of a feature vector sequence. Currently, there are three different kinds of tag types (see fig. 5.24 for an overview). Every tag starts with a feature vector that carries no information but the tag types ...

Context 10

... participants are working independently from each other in a dedicated environ- ment and had no strict time limit. The only other person present is a supervisor who answers questions about using the prototype. While working on the tasks, the partici- pants are videotaped. After completing the task, the participants are asked to fill out an questionnaire (see fig. C.2, fig. C.3, fig. C.4, fig. C.5 and fig. C.6 for the original ...

Context 11

... error δ n,k is now used to calculate the weight update ∆w n,k,i , which is added to the existing weights: ∆w n,k,i = −µ · δ n,k · a n−1,i (2.12) MLPs always map an input vector of fixed length to an output vector of fixed length. That makes is difficult to handle input of varying size. To overcome this problem, recurrent neural networks have been introduced that are scanning over an input of variable size in a series or time steps. In order to incorporate knowledge from previous time steps, at least one hidden layer needs interconnections to it previous activation (see figure ...

Context 12

... matching deals with the generation of mappings between two graphs (see figure 2.2). Formally, in the course of the thesis at hand, a subgraph matching algorithm is an algorithm that creates injective maps m : N α → N β from the nodes of graph α to the nodes of a graph β. The map has to preserve the graph structure of α in ...