1,721,080 research outputs found
Explaining Image Classifiers Generating Exemplars and Counter-Exemplars from Latent Representations
No full textWe present an approach to explain the decisions of black box image classifiers through synthetic exemplar and counter-exemplar learnt in the latent feature space. Our explanation method exploits the latent representations learned through an adversarial autoencoder for generating a synthetic neighborhood of the image for which an explanation is required. A decision tree is trained on a set of images represented in the latent space, and its decision rules are used to generate exemplar images showing how the original image can be modified to stay within its class. Counterfactual rules are used to generate counter-exemplars showing how the original image can “morph” into another class. The explanation also comprehends a saliency map highlighting the areas that contribute to its classification, and areas that push it into another class. A wide and deep experimental evaluation proves that the proposed method outperforms existing explainers in terms of fidelity, relevance, coherence, and stability, besides providing the most useful and interpretable explanations
Black Box Explanation by Learning Image Exemplars in the Latent Feature Space
Get PDFWe present an approach to explain the decisions of black box models for image classification. While using the black box to label images, our explanation method exploits the latent feature space learned through an adversarial autoencoder. The proposed method first generates exemplar images in the latent feature space and learns a decision tree classifier. Then, it selects and decodes exemplars respecting local decision rules. Finally, it visualizes them in a manner that shows to the user how the exemplars can be modified to either stay within their class, or to become counter-factuals by "morphing" into another class. Since we focus on black box decision systems for image classification, the explanation obtained from the exemplars also provides a saliency map highlighting the areas of the image that contribute to its classification, and areas of the image that push it into another class. We present the results of an experimental evaluation on three datasets and two black box models. Besides providing the most useful and interpretable explanations, we show that the proposed method outperforms existing explainers in terms of fidelity, relevance, coherence, and stability
Learning explainable concepts in the presence of a qualitative model.
No full textThis thesis addresses the problem of learning concept descriptions that are interpretable, or explainable. Explainability is understood as the ability to justify the learned concept in terms of the existing background knowledge. The starting point for the work was an existing system that would induce only fully explainable rules. The system performed well when the model used during induction was complete and correct. In practice, however, models are likely to be imperfect, i.e. incomplete and incorrect. We report here a new approach that achieves explainability with imperfect models. The basis of the system is the standard inductive search driven by an accuracy-oriented heuristic, biased towards rule explainability. The bias is abandoned when there is heuristic evidence that a significant loss of accuracy results from constraining the search to explainable rules only. The users can express their relative preference for accuracy vs. explainability. Experiments with the system indicate that, even with a partially incomplete and/or incorrect model, insisting on explainability results in only a small loss of accuracy. We also show how the new approach described can repair a faulty model using evidence derived from data during induction
Discovering Tourist Attractions of Cities Using Flickr and OpenStreetMap Data
No full textTourism is a growing industry which needs accurate management and planning. Photography and tourism are inseparable; Photographs play the role of tourists’ footprints during their visit to a touristic city. Nowadays, the large deployment of mobile devices and digital cameras has led to a massive increase in the volume of records of where people have been and when they were there. In this paper, we introduce a new method to automatically discover the touristic attractions of every single city with the use of two open-source platforms, Flickr and OpenStreetMap. We applied techniques to convert raw metadata of geotagged photos downloaded from Flickr to information about popular Points of Interest with the help of additional information retrieved from OpenStreetMap
Structural identification of unintelligible documents.
No full textThis thesis is about the identification of unintelligible documents using machine learning techniques. An unintelligible document is a document that is not necessarily expressed in a natural language. We have developed a three level approach including: the representation of a base alphabet, the textual feature extraction and the induction of classification models using machine learning techniques. For the representation of the base alphabet, we have introduced a clustering technique that reduces the size of the alphabet and increases the density of the training documents. For the textual feature extraction, we present four techniques adapted for unintelligible documents. Among the presented feature extraction techniques, we have developed two techniques based on the data compression concept. We have explored two approaches to accomplish the induction of an identification model: a rule-based system called RIPPER and an approach based on grammatical inference. In order to combine different complementary techniques, we have developed an object-oriented framework that encourages code reuse. The presented techniques have been tested in an experiment on three different domains: the identification of e-mail documents, the identification of BASE64 encoded documents and the identification of GIF images versus BMP images
Feature engineering for a symbolic approach to text classification.
No full textMost text classification research to date has used the standard "bag of words" model for text representation inherited from the word-based indexing techniques used in information retrieval research. There have been a number of past attempts to find better representations, but very few positive results have been found. Most of this previous work, however, has concentrated on retrieval rather than classification tasks, and none has involved symbolic learning algorithms. This thesis investigates a number of feature engineering methods for text classification in the context of a symbolic rule-based learning algorithm. The focus is on changing the standard "bag of words" representation of text by incorporating some shallow linguistic processing techniques. Several new representations of text are explored in the hopes that they will allow the learner to find points of high information gain that were not present in the original set of words. Representations based on both semantic and syntactic linguistic knowledge are defined and evaluated using the RIPPER rule-learning system. Two major corpora are used for evaluation: a standard, widely-used corpus of news stories, and a new corpus of folk song lyrics. The results of the experiments are mostly negative. Although in some cases the new representations are at least as good as the bag of words, the improvements in quantitative performance that were hoped for do not materialize. However, the results are not entirely discouraging. The syntactically defined representations may enable the learner to produce simpler and more comprehensible hypotheses, and the semantically defined representations do produce some real performance gains on smaller classification tasks that for various reasons fail to scale up to larger tasks. Some ideas are offered as to why the new representations fail to produce better results, and some suggestions are made for continuing the research in future
SAFIR, analogy-based system for database definition and query reuse.
No full textThis thesis presents a system which employs analogical reasoning to define and manipulate databases. The objective is to model the behavior of naive users who often adapt database relations and queries from documentation examples to the problem at hand. The user has at her disposal a manual that contains examples of database designs with a variety of database queries in some exemplary milieu called the base domain. The manual explains the example queries in simple terms and gives their implementation in a database query language such as SQL. Our user is not a database expert. She can neither define the database nor can she compose query implementations from scratch. At best she can construct simple queries in the base domain by copying them verbatim from the manual. Our system works in two steps: first it defines a target database by building incrementally analogies represented by pairs of object correspondences, base-target. These analogies are built by mapping objects, relationships, and semantic constructs from the base domain into the target domain. The process is then interactive and incremental (i.e. the user can accept or reject the inferred analogies). The second step of the approach consists to derive target domain query's implementation from its specification by reusing queries developed in the base domain. We rely on a weak form of derivational analogy--traces of a planner operating in the base domain are used to derive a target domain query's implementation in SQL from its specification. The very simple planner used builds in the base domain plans that result in particularly simple, prototypical queries. Queries are specified in a high level subset of natural language
Extending AdaBoost:Varying the Base Learners and Modifying the Weight Calculation
No full textAdaBoost has been considered one of the best classifiers ever
developed, but two important problems have not yet been addressed. The
first is the dependency on the ``weak" learner, and the second is the
failure to maintain the performance of learners with small error rates
(i.e. ``strong" learners). To solve the first problem, this work
proposes using a different learner in each iteration - known as AdaBoost
Dynamic (AD) - thereby ensuring that the performance of the algorithm
is almost equal to that of the best ``weak" learner executed with AdaBoost.M1. The work
then further modifies the procedure to vary the learner in each
iteration, in order to locate the learner with the smallest error rate
in its training data. This is done using the same weight calculation
as in the original AdaBoost; this version is known as AdaBoost Dynamic
with Exponential Loss (AB-EL). The results were poor, because AdaBoost
does not perform well with strong learners, so, in this sense, the work
confirmed previous works' results. To determine how to improve the
performance, the weight calculation is modified to use the sigmoid function
with algorithm output being the derivative of the same sigmoid function,
rather than the logistic
regression weight calculation originally used by AdaBoost; this
version is known as AdaBoost Dynamic with Logistic Loss (AB-DL). This
work presents the convergence proof that binomial weight calculation
works, and that this approach improves the results for the strong
learner, both theoretically and empirically. AB-DL also has some
disadvantages, like the search for the ``best" classifier and that
this search reduces the diversity among the classifiers. In order
to attack these issues, another algorithm is proposed that combines
AD ``weak" leaner execution policy with a small modification of AB-DL's
weight calculation, called AdaBoost Dynamic with Added Cost (AD-AC).
AD-AC also has a theoretical upper bound error, and the algorithm
offers a small accuracy improvement when compared with AB-DL, and traditional
AdaBoost approaches. Lastly, this work also adapts AD-AC's weight calculation
approach to deal with data stream problem, where classifiers must deal with
very large data sets (in the order of millions of instances), and limited
memory availability
Identification of attribute interactions and generation of globally relevant continuous features in machine learning
No full textDatasets found in real world applications of machine learning are often characterized by low-level attributes with important interactions among them. Such interactions may increase the complexity of the learning task by limiting the usefulness of the attributes to dispersed regions of the representation space. In such cases, we say that the attributes are locally relevant. To obtain adequate performance with locally relevant attributes, the learning algorithm must be able to analyse the interacting attributes simultaneously and fit an appropriate model for the type of interactions observed. This is a complex task that surpasses the ability of most existing machine learning systems. This research proposes a solution to this problem by extending the initial representation with new globally relevant features. The new features make explicit the important information that was previously hidden by the initial interactions, thus reducing the complexity of the learning task.
This dissertation first proposes an idealized study of the potential benefits of globally relevant features assuming perfect knowledge of the interactions among the initial attributes. This study involves synthetic data and a variety of machine learning systems. Recognizing that not all interactions produce a negative effect on performance, the dissertation introduces a novel technique named Relevance graphs to identify the interactions that negatively affect the performance of existing learning systems. The tool of interactive relevance graphs addresses another important need by providing the user with an opportunity to participate in the construction of a new representation that cancels the effects of the negative attribute interactions. The dissertation extends the concept of relevance graphs by introducing a series of algorithms for the automatic discovery of appropriate transformations. We use the named GLOREF (GLObally RElevant Features) to designate the approach that integrates these algorithms. The dissertation fully describes the GLOREF approach along with an extensive empirical evaluation with both synthetic and UCI datasets. This evaluation shows that the features produced by the GLOREF approach significantly improve the accuracy with both synthetic and real-world data
- …
