Unsupervised Learning of Depth and Camera Pose with Feature Map Warping

Estimating the depth of image and egomotion of agent are important for autonomous and robot in understanding the surrounding environment and avoiding collision. Most existing unsupervised methods estimate depth and camera egomotion by minimizing photometric error between adjacent frames. However, the photometric consistency sometimes does not meet the real situation, such as brightness change, moving objects and occlusion. To reduce the influence of brightness change, we propose a feature pyramid matching loss (FPML) which captures the trainable feature error between a current and the adjacent frames and therefore it is more robust than photometric error. In addition, we propose the occlusion-aware mask (OAM) network which can indicate occlusion according to change of masks to improve estimation accuracy of depth and camera pose. The experimental results verify that the proposed unsupervised approach is highly competitive against the state-of-the-art methods, both qualitatively and quantitatively. Specifically, our method reduces absolute relative error (Abs Rel) by 0.017–0.088.

Features Reweighting and Selection in ligand-based Virtual Screening for Molecular Similarity Searching Based on Deep Belief Networks

Virtual screening (VS) is defined as the use of a compilation of computational procedures to grade, score and/or sort several chemical formations. The purpose of VS is to identify the molecules holding the greatest prior probabilities of activity. Many of the conventional similarity methods assume that molecular features that do not relate to the biological activity carry the same weight as the important ones. For this reason, the researchers on this paper investigated that some features are being more important than others through the chemist structure diagrams and the weight for each fragment should be taken into consideration by giving more weight to those fragments that are more important. In this paper, a deep learning method specifically known as Deep Belief Networks (DBN) has been used to reweight the molecule features and based on this new weigh, the reconstruction feature error has been calculated for all the features. Based on the reconstruction feature error values, Principal Component Analysis (PCA) has been used for the dimension’s reduction and only few hundreds of features have been selected based on the less error rate. The main aim of this research is to show an improvement of the similarity searching performance based on the selected features those have less error rate. The results derived through the DBN were compared with those derived through other similarity methods, such as the Tanimoto coefficient and the quantum-based methods. This comparison revealed the performance of the DBN with the structurally heterogeneous data sets (DS1 and DS3) to be superior to the performances of all the other techniques.

An Acoustical Study of the Fricative /s/ in the Speech of Palestinian-speaking Broca's Asphasics – Preliminary Findings

This study was conducted with four Palestinian-speaking Broca's aphasics and four normal speakers to examine the production of the fricative /s/ and to analyze the difference in acoustic patterns between the two groups. The acoustic analysis revealed that the Broca's aphasics were able to maintain the phonetic distinction between voiced and voiceless fricatives and did not exhibit voicing or devoicing errors. This result is inconsistent with findings from other languages. The spectral peak of the alveolar /s/ was lower for the aphasic subjects than for the control speakers. The Broca's aphasics displayed a longer duration for /s/ than the control subjects did. The /s/ spectra for the Broca's aphasics were "flatter" than the /s/ spectra for the control speakers and were characterized by a relatively low-frequency energy peak. Substitution errors predominated. In most cases, a one-distinctive-feature error occurred that primarily involved the place of articulation. Generally, the findings of the study suggest that the errors exhibited by Broca's aphasics reflect articulatory movements and articulatory implementation deficits rather than a selection of inappropriate speech segments.

Selective Attention and Error Processing in an Illusory Conjunction Task

We recorded event-related potentials in an illusory conjunction task, in which subjects were cued on each trial to search for a particular colored letter in a subsequently presented test array, consisting of three different letters in three different colors. In a proportion of trials the target letter was present and in other trials none of the relevant features were present. In still other trials one of the features (color or letter identity) were present or both features were present but not combined in the same display element. When relevant features were present this resulted in an early posterior selection negativity (SN) and a frontal selection positivity (FSP). When a target was presented, this resulted in a FSP that was enhanced after 250 ms as compared to when both relevant features were present but not combined in the same display element. This suggests that this effect reflects an extra process of attending to both features bound to the same object. There were no differences between the ERPs in feature error and conjunction error trials, contrary to the idea that these two types of errors are due to different (perceptual and attentional) mechanisms. The P300 in conjunction error trials was much reduced relative to the P300 in correct target detection trials. A similar, error-related negativity-like component was visible in the response-locked averages in correct target detection trials, in feature error trials, and in conjunction error trials. Dipole modeling of this component resulted in a source in a deep medial-frontal location. These results suggested that this type of task induces a high level of response conflict, in which decision-related processes may play a major role.

Masked and Filtered Simulation of Hearing Loss

Hearing threshold configuration of the impaired ear was simulated via two different methods in the normal ear of each of 6 subjects with unilateral sensorineural hearing loss. The two hearing loss simulation methods were (a) frequency-specific attenuation (filtering), and (b) masking by shaped masking noise. Identification responses to 20 consonant-vowel nonsense syllables were obtained from the normal ear of each subject in the two hearing loss simulation conditions, as well as from the impaired ear. Speech data were scored both for overall percent correct consonant recognition and for error patterns on a set Of eight consonant features. The computed correlation between gross error difference scores and a similarity metric for feature error pattern was essentially zero, indicating that these may be relatively independent measures of speech recognition, and thus may reflect different aspects Of subjects' behavior. For 3 subjects, both masking and filtering successfully simulated the effects of sensorineural hearing loss on consonant feature error patterns. For 1 subject, only filtering produced feature error patterns similar to those of the impaired ear. For the remaining subjects, neither simulation was successful.

Articulation Instruction Based on Distinctive Features Theory

An analysis of a child’s articulation performance from a distinctive features point of view shows that several error phonemes can be accounted for by only one feature error. More efficient instruction can be obtained by choosing a distinctive feature, as opposed to a single phoneme, as the remediation target. Measurement steps for the selection of target phonemes for treatment from a distinctive feature viewpoint are recommended.