Uri Shalit

Large Scale Online Learning of Image Similarity through Ranking

Learning a measure of similarity between pairs of objects is a fundamental problem in machine learning. Pairwise similarity plays a crucial role in classification algorithms like nearest neighbors, and is practically important for applications like searching for images that are similar to a given image or finding videos that are relevant to a given video. In these tasks, users look for objects that are both visually similar and semantically related to a given object. Unfortunately, current approaches for learning semantic similarity are limited to small scale datasets, because their complexity grows quadratically with the sample size, and because they impose costly positivity constraints on the learned similarity functions. To address real-world large-scale AI problem, like learning similarity over all images on the web, we need to develop new algorithms that scale to many samples, many classes, and many features. The current abstract presents OASIS, an Online Algorithm for Scalable Image Similarity learning that learns a bilinear similarity measure over sparse representations. OASIS is an online dual approach using the passive-aggressive family of learning algorithms with a large margin criterion and an efficient hinge loss cost. Our experiments show that OASIS is both fast and accurate at a wide range of scales: for a dataset with thousands of images, it achieves better results than existing state-of-the-art methods, while being an order of magnitude faster. Comparing OASIS with different symmetric variants, provides unexpected insights into the effect of symmetry on the quality of the similarity. For large, web scale, datasets, OASIS can be trained on more than two million images from 150K text queries within two days on a single CPU. Human evaluations showed that 35% of the ten top images ranked by OASIS were semantically relevant to a query image. This suggests that query-independent similarity could be accurately learned even for large-scale datasets that could not be handled before.

Descending Systems Translate Transient Cortical Commands into a Sustained Muscle Activation Signal

Controlling motor actions requires online adjustments of time-varying parameters. Although numerous studies have attempted to identify the parameters coded in different motor sites, the relationships between the temporal profile of neuronal responses and the dynamics of motor behavior remain poorly understood in particular because motor parameters such as force and movement direction often change over time. We studied time-dependent coding of cortical and spinal neurons in primates performing an isometric wrist task with an active hold period, which made it possible to segregate motor behavior into its phasic and sustained components. Here, we show that cortical neurons transiently code motor-related parameters when actively acquiring a goal, whereas spinal interneurons provide persistent information regarding maintained torque level and posture. Moreover, motor cortical neurons differed substantially from spinal neurons with regard to the evolvement of parameter-specific coding over the course of a trial. These results suggest that the motor cortex and spinal cord use different control policies: Cortical neurons produce transient motor commands governing ensuing actions, whereas spinal neurons exhibit sustained coding of ongoing motor states. Hence, motor structures downstream to M1 need to integrate cortical commands to produce state-dependent spinal firing.

Computation in spinal circuitry: lessons from behaving primates.

Performing voluntary motor actions requires the translation of motor commands into a specific set of muscle activation. While it is assumed that this process is carried out via cooperative interactions between supraspinal and spinal neurons, the unique contribution of each of these areas to the process is still unknown. Many studies have focused on the neuronal representation of the motor command, mostly in the motor cortex. Nonetheless, to execute these commands there must be a mechanism that can translate this representation into a sustained drive to the spinal motoneurons (MNs). Here we review different candidate mechanisms for activating MNs and their possible role in voluntary movements. We discuss recent studies which directly estimate the contribution of segmental INs to the transmission of cortical command to MNs, both in terms of functional connectivity and as a computational link. Finally, we suggest a conceptual framework in which the cortical motor command is processed simultaneously via MNs and INs. In this model, the motor cortex provides a transient signal which is important for initiating new patterns of recruited muscles, whereas the INs translate this command into a sustained, amplified and muscle-based signal which is necessary to maintain ongoing muscle activity.

FuncISH: learning a functional representation of neural ISH images

Motivation: High-spatial resolution imaging datasets of mammalian brains have recently become available in unprecedented amounts. Images now reveal highly complex patterns of gene expression varying on multiple scales. The challenge in analyzing these images is both in extracting the patterns that are most relevant functionally and in providing a meaningful representation that allows neuroscientists to interpret the extracted patterns. Results: Here, we present FuncISH—a method to learn functional representations of neural in situ hybridization (ISH) images. We represent images using a histogram of local descriptors in several scales, and we use this representation to learn detectors of functional (GO) categories for every image. As a result, each image is represented as a point in a low-dimensional space whose axes correspond to meaningful functional annotations. The resulting representations define similarities between ISH images that can be easily explained by functional categories. We applied our method to the genomic set of mouse neural ISH images available at the Allen Brain Atlas, finding that most neural biological processes can be inferred from spatial expression patterns with high accuracy. Using functional representations, we predict several gene interaction properties, such as protein–protein interactions and cell-type specificity, more accurately than competing methods based on global correlations. We used FuncISH to identify similar expression patterns of GABAergic neuronal markers that were not previously identified and to infer new gene function based on image–image similarities. Contact: noalis@gmail.com Supplementary information: Supplementary data are available at Bioinformatics online.