Learning Nonlinear Visual Processing from Natural Images

The paradigm of computational vision hypothesizes that any visual function -- such as the recognition of your grandparent -- can be replicated by computational processing of the visual input. What are these computations that the brain performs? What should or could they be? Working on the latter question, this dissertation takes the statistical approach, where the suitable computations are attempted to be learned from the natural visual data itself. In particular, we empirically study the computational processing that emerges from the statistical properties of the visual world and the constraints and objectives specified for the learning process. This thesis consists of an introduction and 7 peer-reviewed publications, where the purpose of the introduction is to illustrate the area of study to a reader who is not familiar with computational vision research. In the scope of the introduction, we will briefly overview the primary challenges to visual processing, as well as recall some of the current opinions on visual processing in the early visual systems of animals. Next, we describe the methodology we have used in our research, and discuss the presented results. We have included some additional remarks, speculations and conclusions to this discussion that were not featured in the original publications. We present the following results in the publications of this thesis. First, we empirically demonstrate that luminance and contrast are strongly dependent in natural images, contradicting previous theories suggesting that luminance and contrast were processed separately in natural systems due to their independence in the visual data. Second, we show that simple cell -like receptive fields of the primary visual cortex can be learned in the nonlinear contrast domain by maximization of independence. Further, we provide first-time reports of the emergence of conjunctive (corner-detecting) and subtractive (opponent orientation) processing due to nonlinear projection pursuit with simple objective functions related to sparseness and response energy optimization. Then, we show that attempting to extract independent components of nonlinear histogram statistics of a biologically plausible representation leads to projection directions that appear to differentiate between visual contexts. Such processing might be applicable for priming, \ie the selection and tuning of later visual processing. We continue by showing that a different kind of thresholded low-frequency priming can be learned and used to make object detection faster with little loss in accuracy. Finally, we show that in a computational object detection setting, nonlinearly gain-controlled visual features of medium complexity can be acquired sequentially as images are encountered and discarded. We present two online algorithms to perform this feature selection, and propose the idea that for artificial systems, some processing mechanisms could be selectable from the environment without optimizing the mechanisms themselves. In summary, this thesis explores learning visual processing on several levels. The learning can be understood as interplay of input data, model structures, learning objectives, and estimation algorithms. The presented work adds to the growing body of evidence showing that statistical methods can be used to acquire intuitively meaningful visual processing mechanisms. The work also presents some predictions and ideas regarding biological visual processing.

Clustering Users of Online Content Service

Online content services can greatly benefit from personalisation features that enable delivery of content that is suited to each user's specific interests. This thesis presents a system that applies text analysis and user modeling techniques in an online news service for the purpose of personalisation and user interest analysis. The system creates a detailed thematic profile for each content item and observes user's actions towards content items to learn user's preferences. A handcrafted taxonomy of concepts, or ontology, is used in profile formation to extract relevant concepts from the text. User preference learning is automatic and there is no need for explicit preference settings or ratings from the user. Learned user profiles are segmented into interest groups using clustering techniques with the objective of providing a source of information for the service provider. Some theoretical background for chosen techniques is presented while the main focus is in finding practical solutions to some of the current information needs, which are not optimally served with traditional techniques.

Investigating online reading with eye tracking and EEG : The influence of text format, reading task and parafoveal stimuli on reading processes

Research on reading has been successful in revealing how attention guides eye movements when people read single sentences or text paragraphs in simplified and strictly controlled experimental conditions. However, less is known about reading processes in more naturalistic and applied settings, such as reading Web pages. This thesis investigates online reading processes by recording participants eye movements. The thesis consists of four experimental studies that examine how location of stimuli presented outside the currently fixated region (Study I and III), text format (Study II), animation and abrupt onset of online advertisements (Study III), and phase of an online information search task (Study IV) affect written language processing. Furthermore, the studies investigate how the goal of the reading task affects attention allocation during reading by comparing reading for comprehension with free browsing, and by varying the difficulty of an information search task. The results show that text format affects the reading process, that is, vertical text (word/line) is read at a slower rate than a standard horizontal text, and the mean fixation durations are longer for vertical text than for horizontal text. Furthermore, animated online ads and abrupt ad onsets capture online readers attention and direct their gaze toward the ads, and distract the reading process. Compared to a reading-for-comprehension task, online ads are attended to more in a free browsing task. Moreover, in both tasks abrupt ad onsets result in rather immediate fixations toward the ads. This effect is enhanced when the ad is presented in the proximity of the text being read. In addition, the reading processes vary when Web users proceed in online information search tasks, for example when they are searching for a specific keyword, looking for an answer to a question, or trying to find a subjectively most interesting topic. A scanning type of behavior is typical at the beginning of the tasks, after which participants tend to switch to a more careful reading state before finishing the tasks in the states referred to as decision states. Furthermore, the results also provided evidence that left-to-right readers extract more parafoveal information to the right of the fixated word than to the left, suggesting that learning biases attentional orienting towards the reading direction.