Transformer architectures have become the building blocks for many state-of-the-art natural language processing (NLP) models. While transformers are certainly powerful, researchers' understanding of how they actually work remains limited. This is problematic due to the lack of transparency and the possibility of biases being inherited via training data and algorithms, which could cause models to produce unfair or incorrect predictions. In the paper Transformer Visualization via Dictionary Learning: Contextualized Embedding as a Linear Superposition of Transformer Factors, a Yann LeCun team from Facebook AI Research, UC Berkeley and New York University leverages dictionary learning techniques to provide detailed visualizations of transformer representations and insights into the semantic structures -- such as word-level disambiguation, sentence-level pattern formation, and long-range dependencies -- that are captured by transformers. Previous attempts to visualize and analyze this "black box" issue in transformers include direct visualization and, more recently, "probing tasks" designed to interpret transformer models.

Model explainability is one of the most important problems in machine learning today. It's often the case that certain "black box" models such as deep neural networks are deployed to production and are running critical systems from everything in your workplace security cameras to your smartphone. It's a scary thought that not even the developers of these algorithms understand why exactly the algorithms make the decisions they do -- or even worse, how to prevent an adversary from exploiting them. While there are many challenges facing the designer of a "black box" algorithm, it's not completely hopeless. There are actually many different ways to illuminate the decisions a model makes.

The number of connected Internet of Things (IoT) devices within cyber-physical infrastructure systems grows at an increasing rate. This poses significant device management and security challenges to current IoT networks. Among several approaches to cope with these challenges, data-based methods rooted in deep learning (DL) are receiving an increased interest. In this paper, motivated by the upcoming surge of 5G IoT connectivity in industrial environments, we propose to integrate a DL-based anomaly detection (AD) as a service into the 3GPP mobile cellular IoT architecture. The proposed architecture embeds autoencoder based anomaly detection modules both at the IoT devices (ADM-EDGE) and in the mobile core network (ADM-FOG), thereby balancing between the system responsiveness and accuracy. We design, integrate, demonstrate and evaluate a testbed that implements the above service in a real-world deployment integrated within the 3GPP Narrow-Band IoT (NB-IoT) mobile operator network.

In this work, we propose a novel missile guidance algorithm that combines deep learning based trajectory prediction with nonlinear model predictive control. Although missile guidance and threat interception is a well-studied problem, existing algorithms' performance degrades significantly when the target is pulling high acceleration attack maneuvers while rapidly changing its direction. We argue that since most threats execute similar attack maneuvers, these nonlinear trajectory patterns can be processed with modern machine learning methods to build high accuracy trajectory prediction algorithms. We train a long short-term memory network (LSTM) based on a class of simulated structured agile attack patterns, then combine this predictor with quadratic programming based nonlinear model predictive control (NMPC). Our method, named nonlinear model based predictive control with target acceleration predictions (NMPC-TAP), significantly outperforms compared approaches in terms of miss distance, for the scenarios where the target/threat is executing agile maneuvers.

We consider the problem of estimating an object's physical properties such as mass, friction, and elasticity directly from video sequences. Such a system identification problem is fundamentally ill-posed due to the loss of information during image formation. Current solutions require precise 3D labels which are labor-intensive to gather, and infeasible to create for many systems such as deformable solids or cloth. We present gradSim, a framework that overcomes the dependence on 3D supervision by leveraging differentiable multiphysics simulation and differentiable rendering to jointly model the evolution of scene dynamics and image formation. This novel combination enables backpropagation from pixels in a video sequence through to the underlying physical attributes that generated them. Moreover, our unified computation graph -- spanning from the dynamics and through the rendering process -- enables learning in challenging visuomotor control tasks, without relying on state-based (3D) supervision, while obtaining performance competitive to or better than techniques that rely on precise 3D labels.

Learning the dynamics of a physical system wherein an autonomous agent operates is an important task. Often these systems present apparent geometric structures. For instance, the trajectories of a robotic manipulator can be broken down into a collection of its transitional and rotational motions, fully characterized by the corresponding Lie groups and Lie algebras. In this work, we take advantage of these structures to build effective dynamical models that are amenable to sample-based learning. We hypothesize that learning the dynamics on a Lie algebra vector space is more effective than learning a direct state transition model. To verify this hypothesis, we introduce the Group Enhanced Model (GEM). GEMs significantly outperform conventional transition models on tasks of long-term prediction, planning, and model-based reinforcement learning across a diverse suite of standard continuous-control environments, including Walker, Hopper, Reacher, Half-Cheetah, Inverted Pendulums, Ant, and Humanoid. Furthermore, plugging GEM into existing state of the art systems enhances their performance, which we demonstrate on the PETS system. This work sheds light on a connection between learning of dynamics and Lie group properties, which opens doors for new research directions and practical applications along this direction. Our code is publicly available at: https://tinyurl.com/GEMMBRL.

During a geosteering operation the well path is intentionally adjusted in response to the new data acquired while drilling. To achieve consistent high-quality decisions, especially when drilling in complex environments, decision support systems can help cope with high volumes of data and interpretation complexities. They can assimilate the real-time measurements into a probabilistic earth model and use the updated model for decision recommendations. Recently, machine learning (ML) techniques have enabled a wide range of methods that redistribute computational cost from on-line to off-line calculations. In this paper, we introduce two ML techniques into the geosteering decision support framework. Firstly, a complex earth model representation is generated using a Generative Adversarial Network (GAN). Secondly, a commercial extra-deep electromagnetic simulator is represented using a Forward Deep Neural Network (FDNN). The numerical experiments demonstrate that the combination of the GAN and the FDNN in an ensemble randomized maximum likelihood data assimilation scheme provides real-time estimates of complex geological uncertainty. This yields reduction of geological uncertainty ahead of the drill-bit from the measurements gathered behind and around the well bore.

In the past decade, convolutional neural networks (CNNs) have been widely adopted as the main building block for end-to-end audio classification models, which aim to learn a direct mapping from audio spectrograms to corresponding labels. To better capture long-range global context, a recent trend is to add a self-attention mechanism on top of the CNN, forming a CNN-attention hybrid model. However, it is unclear whether the reliance on a CNN is necessary, and if neural networks purely based on attention are sufficient to obtain good performance in audio classification. In this paper, we answer the question by introducing the Audio Spectrogram Transformer (AST), the first convolution-free, purely attention-based model for audio classification. We evaluate AST on various audio classification benchmarks, where it achieves new state-of-the-art results of 0.485 mAP on AudioSet, 95.6% accuracy on ESC-50, and 98.1% accuracy on Speech Commands V2.

In recent years, deep learning has become prevalent to solve applications from multiple domains. Convolutional Neural Networks (CNNs) particularly have demonstrated state of the art performance for the task of image classification. However, the decisions made by these networks are not transparent and cannot be directly interpreted by a human. Several approaches have been proposed to explain to understand the reasoning behind a prediction made by a network. In this paper, we propose a topology of grouping these methods based on their assumptions and implementations. We focus primarily on white box methods that leverage the information of the internal architecture of a network to explain its decision. Given the task of image classification and a trained CNN, this work aims to provide a comprehensive and detailed overview of a set of methods that can be used to create explanation maps for a particular image, that assign an importance score to each pixel of the image based on its contribution to the decision of the network. We also propose a further classification of the white box methods based on their implementations to enable better comparisons and help researchers find methods best suited for different scenarios.

QAT refers to the quantization of activations to binary, bipolar, or multi-bit values during neural network training. On the other hand, if deep neural networks (DNNs) [1]-[31], ultra-low-latency realization a set of values can only assume non-negative numbers, it relies on of these models for applications with stringent, sub-microsecond the parameterized clipping activation (PACT) [9] function to quantize latency requirements continues to be an unresolved, challenging activations. Field-programmable gate array (FPGA)-based DNN accelerators FCP applies fanin constraints to individual filters/neurons such that are gaining traction as a serious contender to replace graphics the number of inputs to each filter/neuron is small enough to make processing unit/central processing unit-based platforms considering a realization based on input enumeration (as described in NullaNet their performance, flexibility, and energy efficiency. In this work, FCP is either based on the alternating [32], LUTNet (2019) [33], and LogicNets (2020) [34] are among accelerators direction method of multipliers [35] or gradual pruning [11]. Finally, functions of different filters/neurons are represented using This paper presents NullaNet Tiny, an across-the-stack design truth tables which are then fed to the logic minimization module.