Correctly detecting the semantic type of data columns is crucial for data science tasks such as automated data cleaning, schema matching, and data discovery. Existing data preparation and analysis systems rely on dictionary lookups and regular expression matching to detect semantic types. However, these matching-based approaches often are not robust to dirty data and only detect a limited number of types. We introduce Sherlock, a multi-input deep neural network for detecting semantic types. We train Sherlock on $686,765$ data columns retrieved from the VizNet corpus by matching $78$ semantic types from DBpedia to column headers. We characterize each matched column with $1,588$ features describing the statistical properties, character distributions, word embeddings, and paragraph vectors of column values. Sherlock achieves a support-weighted F$_1$ score of $0.89$, exceeding that of machine learning baselines, dictionary and regular expression benchmarks, and the consensus of crowdsourced annotations.

This paper addresses the problem of assessing the variability of predictions from deep neural networks. There is a growing literature on using and improving the predictive accuracy of deep networks, but a concomitant improvement in the quantification of their uncertainty is lacking. We provide a prediction interval network (PI-Network) which is a transparent, tractable modification of the standard predictive loss used to train deep networks. The PI-Network outputs three values instead of a single point estimate and optimizes a loss function inspired by quantile regression. We go beyond merely motivating the construction of these networks and provide two prediction interval methods with provable, finite sample coverage guarantees without any assumptions on the underlying distribution from which our data is drawn. We only require that the observations are independent and identically distributed. Furthermore, our intervals adapt to heteroskedasticity and asymmetry in the conditional distribution of the response given the covariates. The first method leverages the conformal inference framework and provides average coverage. The second method provides a new, stronger guarantee by conditioning on the observed data. Lastly, our loss function does not compromise the predictive accuracy of the network like other prediction interval methods. We demonstrate the ease of use of the PI-Network as well as its improvements over other methods on both simulated and real data. As the PI-Network can be used with a host of deep learning methods with only minor modifications, its use should become standard practice, much like reporting standard errors along with mean estimates.

In this work we propose Hebbian-descent as a biologically plausible learning rule for hetero-associative as well as auto-associative learning in single layer artificial neural networks. It can be used as a replacement for gradient descent as well as Hebbian learning, in particular in online learning, as it inherits their advantages while not suffering from their disadvantages. We discuss the drawbacks of Hebbian learning as having problems with correlated input data and not profiting from seeing training patterns several times. For gradient descent we identify the derivative of the activation function as problematic especially in online learning. Hebbian-descent addresses these problems by getting rid of the activation function's derivative and by centering, i.e. keeping the neural activities mean free, leading to a biologically plausible update rule that is provably convergent, does not suffer from the vanishing error term problem, can deal with correlated data, profits from seeing patterns several times, and enables successful online learning when centering is used. We discuss its relationship to Hebbian learning, contrastive learning, and gradient decent and show that in case of a strictly positive derivative of the activation function Hebbian-descent leads to the same update rule as gradient descent but for a different loss function. In this case Hebbian-descent inherits the convergence properties of gradient descent, but we also show empirically that it converges when the derivative of the activation function is only non-negative, such as for the step function for example. Furthermore, in case of the mean squared error loss Hebbian-descent can be understood as the difference between two Hebb-learning steps, which in case of an invertible and integrable activation function actually optimizes a generalized linear model. ...

We consider the use of look-up tables (LUT) to speed up and simplify the hardware implementation of a deep learning network for inferencing after weights have been successfully trained. The use of LUT replaces the matrix multiply and add operations with a small number of LUTs and addition operations resulting in a multiplier-less implementation. We compare the different tradeoffs of this approach in terms of accuracy versus LUT size and the number of operations.

Deep-learning based classification algorithms have been shown to be susceptible to adversarial attacks: minor changes to the input of classifiers can dramatically change their outputs, while being imperceptible to humans. In this paper, we present a simple hypothesis about a feature compression property of artificial intelligence (AI) classifiers and present theoretical arguments to show that this hypothesis successfully accounts for the observed fragility of AI classifiers to small adversarial perturbations. Drawing on ideas from information and coding theory, we propose a general class of defenses for detecting classifier errors caused by abnormally small input perturbations. We further show theoretical guarantees for the performance of this detection method. We present experimental results with (a) a voice recognition system, and (b) a digit recognition system using the MNIST database, to demonstrate the effectiveness of the proposed defense methods. The ideas in this paper are motivated by a simple analogy between AI classifiers and the standard Shannon model of a communication system.

This work investigates fundamental questions related to locating and defining features in convolutional neural networks (CNN). The theoretical investigations guided by the locality principle show that the relevance of locations within a representation decreases with distance from the center. This is aligned with empirical findings across multiple architectures such as VGG, ResNet, Inception, DenseNet and MobileNet. To leverage our insights, we introduce Locality-promoting Regularization (LOCO-REG). It yields accuracy gains across multiple architectures and datasets.

In deep neural nets, lower level embedding layers account for a large portion of the total number of parameters. Tikhonov regularization, graph-based regularization, and hard parameter sharing are approaches that introduce explicit biases into training in a hope to reduce statistical complexity. Alternatively, we propose stochastically shared embeddings (SSE), a data-driven approach to regularizing embedding layers, which stochastically transitions between embeddings during stochastic gradient descent (SGD). Because SSE integrates seamlessly with existing SGD algorithms, it can be used with only minor modifications when training large scale neural networks. We develop two versions of SSE: SSE-Graph using knowledge graphs of embeddings; SSE-SE using no prior information. We provide theoretical guarantees for our method and show its empirical effectiveness on 6 distinct tasks, from simple neural networks with one hidden layer in recommender systems, to the transformer and BERT in natural languages. We find that when used along with widely-used regularization methods such as weight decay and dropout, our proposed SSE can further reduce overfitting, which often leads to more favorable generalization results.

Previous work shows that adversarially robust generalization requires larger sample complexity, and the same dataset, e.g., CIFAR-10, which enables good standard accuracy may not suffice to train robust models. Since collecting new training data could be costly, we instead focus on inducing locally dense sample distribution, i.e., high sample density in the feature space which could lead to locally sufficient samples for robust learning. We first formally show that the softmax cross-entropy (SCE) loss and its variants induce inappropriate sample density distributions in the feature space, which inspires us to design appropriate training objectives. Specifically, we propose the Max-Mahalanobis center (MMC) loss to create high-density regions for better robustness. It encourages the learned features to gather around the preset class centers with optimal inter-class dispersion. Comparing to the SCE loss and its variants, we empirically demonstrate that applying the MMC loss can significantly improve robustness even under strong adaptive attacks, while keeping state-of-the-art accuracy on clean inputs with little extra computation.

Composition of elementary skills into complex behaviors to solve challenging problems is one of the key elements toward building intelligent machines. To date, there has been plenty of work on learning new policies or skills but almost no focus on composing them to perform complex decision-making. In this paper, we propose a policy ensemble composition framework that takes the robot's primitive policies and learns to compose them concurrently or sequentially through reinforcement learning. We evaluate our method in problems where traditional approaches either fail or exhibit high sample complexity to find a solution. We show that our method not only solves the problems that require both task and motion planning but also exhibits high data efficiency, which is currently one of the main limitations of reinforcement learning.

Attention-based deep neural networks (DNNs) that emphasize the informative information in a local receptive field of an input image have successfully boosted the performance of deep learning in various challenging problems. In this paper, we propose a Dense-and-Implicit-Attention (DIA) unit that can be applied universally to different network architectures and enhance their generalization capacity by repeatedly fusing the information throughout different network layers. The communication of information between different layers is carried out via a modified Long Short Term Memory (LSTM) module within the DIA unit that is in parallel with the DNN. The sharing DIA unit links multi-scale features from different depth levels of the network implicitly and densely. Experiments on benchmark datasets show that the DIA unit is capable of emphasizing channel-wise feature interrelation and leads to significant improvement of image classification accuracy. We further empirically show that the DIA unit is a nonlocal normalization tool that enhances the Batch Normalization. The code is released at https://github.com/gbup-group/DIANet.