Logistic Circuits 
This paper proposes a brand new classification mannequin referred to as logistic circuits. On MNIST and Trend datasets, our studying algorithm outperforms neural networks which have an order of magnitude extra parameters. But, logistic circuits have a definite origin in symbolic AI, forming a discriminative counterpart to probabilistic-logical circuits reminiscent of ACs, SPNs, and PSDDs. We present that parameter studying for logistic circuits is convex optimization, and {that a} easy native search algorithm can induce sturdy mannequin constructions from information. …
Needle-Haystack (NH)
Photographs from social media can replicate numerous viewpoints, heated arguments, and expressions of creativity — including new complexity to look duties. Researchers engaged on Content material-Primarily based Picture Retrieval (CBIR) have historically tuned their search algorithms to match filtered outcomes with person search intent. Nevertheless, we are actually bombarded with composite photos of unknown origin, authenticity, and even which means. With such uncertainty, customers could not have an preliminary thought of what the outcomes of a search question ought to appear like. As an example, hidden individuals, spliced objects, and subtly altered scenes might be tough for a person to detect initially in a meme picture, however could contribute considerably to its composition. We suggest a brand new framework for picture retrieval that fashions object-level areas utilizing picture keypoints retrieved from a picture index, that are then used to precisely weight small contributing objects throughout the outcomes, with out the necessity for expensive object detection steps. We name this technique Needle-Haystack (NH) scoring, and it’s optimized for quick matrix operations on CPUs. We present that this technique not solely performs comparably to state-of-the-art strategies in basic CBIR issues, but in addition outperforms them in fine-grained object- and instance-level retrieval on the Oxford 5K, Paris 6K, Google-Landmarks, and NIST MFC2018 datasets, in addition to meme-style imagery from Reddit. …
AttentionXML
Excessive multi-label textual content classification (XMTC) is a job for tagging every given textual content with essentially the most related a number of labels from a particularly large-scale label set. This job might be discovered in lots of functions, reminiscent of product categorization,internet web page tagging, information annotation and so forth. Many strategies have been proposed to date for fixing XMTC, whereas a lot of the current strategies use conventional bag-of-words (BOW) illustration, ignoring phrase context in addition to deep semantic info. XML-CNN, a state-of-the-art deep learning-based technique, makes use of convolutional neural community (CNN) with dynamic pooling to course of the textual content, going past the BOW-based appraoches however failing to seize 1) the long-distance dependency amongst phrases and a couple of) completely different ranges of significance of a phrase for every label. We suggest a brand new deep learning-based technique, AttentionXML, which makes use of bidirectional lengthy short-term reminiscence (LSTM) and a multi-label consideration mechanism for fixing the above 1st and 2nd issues, respectively. We empirically in contrast AttentionXML with different six state-of-the-art strategies over 5 benchmark datasets. AttentionXML outperformed all competing strategies underneath all experimental settings besides solely a few circumstances. As well as, a consensus ensemble of AttentionXML with the second finest technique, Parabel, might additional enhance the efficiency over all 5 benchmark datasets. …
PerceptNet
In an effort to design haptic icons or construct a haptic vocabulary, we require a set of simply distinguishable haptic alerts to keep away from perceptual ambiguity, which in flip requires a method to precisely estimate the perceptual (dis)similarity of such alerts. On this work, we current a novel technique to be taught such a perceptual metric based mostly on information from human research. Our technique is predicated on a deep neural community that initiatives alerts to an embedding area the place the pure Euclidean distance precisely fashions the diploma of dissimilarity between two alerts. The community is educated solely on non-numerical comparisons of triplets of alerts, utilizing a novel triplet loss that considers each varieties of triplets which can be straightforward to order (inequality constraints), in addition to these which can be unorderable/ambiguous (equality constraints). In contrast to prior MDS-based non-parametric approaches, our technique might be educated on a partial set of comparisons and may embed new haptic alerts with out retraining the mannequin from scratch. Intensive experimental evaluations present that our technique is considerably simpler at modeling perceptual dissimilarity than alternate options. …
