Few-shot learning has become essential for producing models that generalize from few examples. In this work, we identify metric scaling and task conditioning are important to improve the performance of few-shot algorithms. Our analysis reveals simple completely changes nature algorithm parameter updates. Metric provides improvements up 14% in accuracy certain metrics on mini-Imagenet 5-way 5-shot classification task. We further propose a effective way learner sample set, resulting...

We focus on solving the univariate times series point forecasting problem using deep learning. propose a neural architecture based backward and forward residual links very stack of fully-connected layers. The has number desirable properties, being interpretable, applicable without modification to wide array target domains, fast train. test proposed several well-known datasets, including M3, M4 TOURISM competition datasets containing time from diverse domains. demonstrate state-of-the-art...

Recent progress in neural forecasting accelerated improvements the performance of large-scale systems. Yet, long-horizon remains a very difficult task. Two common challenges afflicting task are volatility predictions and their computational complexity. We introduce NHITS, model which addresses both by incorporating novel hierarchical interpolation multi-rate data sampling techniques. These techniques enable proposed method to assemble its sequentially, emphasizing components with different...

Deep learning has thrived by training on large-scale datasets. However, in robotics applications sample efficiency is critical. We propose a novel adaptive masked proxies method that constructs the final segmentation layer weights from few labelled samples. It utilizes multiresolution average pooling base embeddings with label to act as positive proxy for new class, while fusing it previously learned class signatures. Our evaluated PASCAL-5 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML"...

Metric-based meta-learning techniques have successfully been applied to few-shot classification problems. In this paper, we propose leverage cross-modal information enhance metric-based learning methods. Visual and semantic feature spaces different structures by definition. For certain concepts, visual features might be richer more discriminative than text ones. While for others, the inverse true. Moreover, when support from is limited in image classification, representations (learned...

Forecasting of multivariate time-series is an important problem that has applications in traffic management, cellular network configuration, and quantitative finance. A special case the arises when there a graph available captures relationships between time-series. In this paper we propose novel learning architecture achieves performance competitive with or better than best existing algorithms, without requiring knowledge graph. The key element our proposed learnable fully connected hard...

In this paper, we demonstrate, both theoretically and by numerical examples, that adding a local prediction component to the update rule can significantly improve convergence rate of distributed averaging algorithms. We focus on case where predictor is linear combination node's two previous values (i.e., memory taps), our computes usual weighted received from neighbouring nodes. derive optimal mixing parameter for combining with neighbors' values, carry out theoretical analysis improvement...

Can meta-learning discover generic ways of processing time series (TS) from a diverse dataset so as to greatly improve generalization on new TS coming different datasets? This work provides positive evidence this using broad framework which we show subsumes many existing algorithms. Our theoretical analysis suggests that residual connections act adaptation mechanism, generating subset task-specific parameters based given input, thus gradually expanding the expressive power architecture...

This paper proposes an approach to accelerate local, linear iterative network algorithms asymptotically achieving distributed average consensus. We focus on the class of in which each node initializes its ldquostate valuerdquo local measurement and then at iteration algorithm, updates this state value by adding a weighted sum own neighbors' values. Provided weight matrix satisfies certain convergence conditions, values converge measurements, but is generally slow, impeding practical...

Significant progress has been made recently in developing few-shot object segmentation methods. Learning is shown to be successful settings, using pixel-level, scribbles and bounding box supervision. This paper takes another approach, i.e., only requiring image-level label for segmentation. We propose a novel multi-modal interaction module that utilizes co-attention mechanism both visual word embedding. Our model labels achieves 4.8% improvement over previously proposed It also outperforms...

We present a distributed particle filtering algorithm for target tracking in sensor networks. Several existing algorithms rely on the establishment and maintenance of spanning path or tree. This is challenging networks with dynamic topologies induced by mobile nodes changing wireless conditions; are vulnerable to link node failure. More recent employ consensus improve robustness but they adopt suboptimal fusion rules leading significant deterioration performance. In our algorithm, run local...

We examine the requirements and available methods software to provide (or imitate) uniform random numbers in parallel computing environments. In this context, for great majority of applications, independent streams are required, each being computed on a single processing element at time. Sometimes, thousands or even millions such needed. explain how they can be produced managed. devote particular attention multiple GPU devices.

We propose, develop, and compare new stochastic models for the daily arrival rate in a call center. Following standard practice, day is divided into time periods of equal length (e.g., 15 or 30 minutes), assumed random but constant each period, arrivals are from Poisson process, conditional on rate. The period taken as deterministic base (or expected rate) multiplied by busyness factor having mean 1. Models which factors independent across periods, common applies to all have been studied...

This paper presents <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">greedy gossip with eavesdropping</i> (GGE), a novel randomized algorithm for distributed computation of the average consensus problem. In algorithms, nodes in network randomly communicate their neighbors and exchange information iteratively. The algorithms are simple decentralized, making them attractive wireless applications. general, robust to unreliable conditions time varying...

We show that the task of synthesizing human motion conditioned on a set key frames can be solved more accurately and effectively if deep learning based interpolator operates in delta mode using spherical linear as baseline. empirically demonstrate strength our approach publicly available datasets achieving state-of-the-art performance. further generalize these results by showing <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML"...

This paper proposes a novel framework for delay-tolerant particle filtering that is computationally efficient and has limited memory requirements. Within this the informativeness of delayed (out-of-sequence) measurement (OOSM) estimated using lightweight procedure uninformative measurements are immediately discarded. The requires identification threshold separates informative from uninformative; selection task formulated as constrained optimization problem, where goal to minimize tracking...

Recent progress in neural forecasting accelerated improvements the performance of large-scale systems. Yet, long-horizon remains a very difficult task. Two common challenges afflicting task are volatility predictions and their computational complexity. We introduce N-HiTS, model which addresses both by incorporating novel hierarchical interpolation multi-rate data sampling techniques. These techniques enable proposed method to assemble its sequentially, emphasizing components with different...

Deep learning has thrived by training on large-scale datasets. However, in robotics applications sample efficiency is critical. We propose a novel adaptive masked proxies method that constructs the final segmentation layer weights from few labelled samples. It utilizes multi-resolution average pooling base embeddings with label to act as positive proxy for new class, while fusing it previously learned class signatures. Our evaluated PASCAL-$5^i$ dataset and outperforms state-of-the-art...