This paper considers the decentralized optimization problem, which has applications in large scale machine learning, sensor networks, and control theory. We propose a novel algorithm that can achieve near optimal communication complexity, matching the known lower bound up to a logarithmic factor of the condition number of the problem. Our theoretical results give affirmative answers to the open problem on whether there exists an algorithm that can achieve a communication complexity (nearly) matching the lower bound depending on the global condition number instead of the local one. Moreover, the proposed algorithm achieves the optimal computation complexity matching the lower bound up to universal constants. Furthermore, to achieve a linear convergence rate, our algorithm \emph{doesn't} require the individual functions to be (strongly) convex. Our method relies on a novel combination of known techniques including Nesterov's accelerated gradient descent, multi-consensus and gradient-tracking. The analysis is new, and may be applied to other related problems. Empirical studies demonstrate the effectiveness of our method for machine learning applications.

Airline Crew Pairing Optimization (CPO) aims at generating a set of legal flight sequences (crew pairings), to cover an airline's flight schedule, at minimum cost. It is usually performed using Column Generation (CG), a mathematical programming technique for guided search-space exploration. CG exploits the interdependencies between the current and the preceding CG-iteration for generating new variables (pairings) during the optimization-search. However, with the unprecedented scale and complexity of the emergent flight networks, it has become imperative to learn higher-order interdependencies among the flight-connection graphs, and utilize those to enhance the efficacy of the CPO. In first of its kind and what marks a significant departure from the state-of-the-art, this paper proposes a novel adaptation of the Variational Graph Auto-Encoder for learning plausible combinatorial patterns among the flight-connection data obtained through the search-space exploration by an Airline Crew Pairing Optimizer, AirCROP (developed by the authors and validated by the research consortium's industrial sponsor, GE Aviation). The resulting flight-connection predictions are combined on-the-fly using a novel heuristic to generate new pairings for the optimizer. The utility of the proposed approach is demonstrated on large-scale (over 4200 flights), real-world, complex flight-networks of US-based airlines, characterized by multiple hub-and-spoke subnetworks and several crew bases.

In the literature, there is a strong interest to identify and define activation functions which can improve neural network performance. In recent years there is a renovated interest of the scientific community in investigating activation functions which can be trained during the learning process, usually referred as trainable, learnable or adaptable activation functions. They appear to lead to better network performance. Diverse and heterogeneous models of trainable activation function have been proposed in the literature. In this paper, we present a survey of these models. Starting from a discussion on the use of the term "activation function" in literature, we propose a taxonomy of trainable activation functions, highlight common and distinctive proprieties of recent and past models, and discuss on main advantages and limitations of this type of approach. We show that many of the proposed approaches are equivalent to add neuron layers which use fixed activation functions (nontrainable activation functions) and some simple local rule constrains the corresponding weight layers.

The article explores different deep convolutional neural network architectures trained and tested on posteroanterior chest X-rays of 327 patients who are healthy (152 patients), diagnosed with COVID-19 (125), and other types of pneumonia (48). In particular, this paper looks at the deep convolutional neural networks VGG16 and VGG19, InceptionResNetV2 and InceptionV3, as well as Xception, all followed by a flat multi-layer perceptron and a final 30% drop-out. The paper has found that the best performing network is VGG16 with a final $30$% drop-out trained over 3 classes (COVID-19, No Finding, Other Pneumonia). It has an internal cross-validated accuracy of $93.9(\pm3.4)$%, a COVID-19 sensitivity of $87.7(-1.9,+2)$%, and a No Finding sensitivity of $96.8(\pm0.8)$%. The respective external cross-validated values are $84.1(\pm13.5)$%, $87.7(-1.9,2)$%, and $96.8(\pm0.8)$%. The model optimizer was Adam with a 1e-4 learning rate, and categorical cross-entropy loss. It is hoped that, once this research will be put to practice in hospitals, healthcare professionals will be able in the medium to long-term to diagnosing through machine learning tools possible pneumonia, and if detected, whether it is linked to a COVID-19 infection, allowing the detection of new possible COVID-19 foyers after the end of possible "stop-and-go" lockdowns as expected by until a vaccine is found and widespread. Furthermore, in the short-term, it is hoped practitioners can compare the diagnosis from the deep convolutional neural networks with possible RT-PCR testing results, and if clashing, a Computed Tomography could be performed as they are more accurate in showing COVID-19 pneumonia.

In order to facilitate the production of codes using WiSARD-based models, LabZero developed an ML library C++/Python called wisardpkg. This library is an MIT-licensed open-source package hosted on GitHub under the license.

What might sound like the beginning of a joke has become an attractive prospect for many cognitive scientists: the use of deep neural network models (DNNs) as models of human behavior in perceptual and cognitive tasks. Although DNNs have taken over machine learning, attempts to use them as models of human behavior are still in the early stages. Can they become a versatile model class in the cognitive scientist's toolbox? We first argue why DNNs have the potential to be interesting models of human behavior. We then discuss how that potential can be more fully realized. On the one hand, we argue that the cycle of training, testing, and revising DNNs needs to be revisited through the lens of the cognitive scientist's goals. Specifically, we argue that methods for assessing the goodness of fit between DNN models and human behavior have to date been impoverished. On the other hand, cognitive science might have to start using more complex tasks (including richer stimulus spaces), but doing so might be beneficial for DNN-independent reasons as well. Finally, we highlight avenues where traditional cognitive process models and DNNs may show productive synergy.

Building equitable and inclusive NLP technologies demands consideration of whether and how social attitudes are represented in ML models. In particular, representations encoded in models often inadvertently perpetuate undesirable social biases from the data on which they are trained. In this paper, we present evidence of such undesirable biases towards mentions of disability in two different English language models: toxicity prediction and sentiment analysis. Next, we demonstrate that the neural embeddings that are the critical first step in most NLP pipelines similarly contain undesirable biases towards mentions of disability. We end by highlighting topical biases in the discourse about disability which may contribute to the observed model biases; for instance, gun violence, homelessness, and drug addiction are over-represented in texts discussing mental illness.

In this paper, we consider the recent trend of evaluating progress on reinforcement learning technology by using text-based environments and games as evaluation environments. This reliance on text brings advances in natural language processing into the ambit of these agents, with a recurring thread being the use of external knowledge to mimic and better human-level performance. We present one such instantiation of agents that use commonsense knowledge from ConceptNet to show promising performance on two text-based environments.

Transformer-based QA models use input-wide self-attention -- i.e. across both the question and the input passage -- at all layers, causing them to be slow and memory-intensive. It turns out that we can get by without input-wide self-attention at all layers, especially in the lower layers. We introduce DeFormer, a decomposed transformer, which substitutes the full self-attention with question-wide and passage-wide self-attentions in the lower layers. This allows for question-independent processing of the input text representations, which in turn enables pre-computing passage representations reducing runtime compute drastically. Furthermore, because DeFormer is largely similar to the original model, we can initialize DeFormer with the pre-training weights of a standard transformer, and directly fine-tune on the target QA dataset. We show DeFormer versions of BERT and XLNet can be used to speed up QA by over 4.3x and with simple distillation-based losses they incur only a 1% drop in accuracy. We open source the code at https://github.com/StonyBrookNLP/deformer.

Question answering (QA) tasks have been posed using a variety of formats, such as extractive span selection, multiple choice, etc. This has led to format-specialized models, and even to an implicit division in the QA community. We argue that such boundaries are artificial and perhaps unnecessary, given the reasoning abilities we seek to teach are not governed by the format. As evidence, we use the latest advances in language modeling to build a single pre-trained QA model, UnifiedQA, that performs surprisingly well across 17 QA datasets spanning 4 diverse formats. UnifiedQA performs on par with 9 different models that were trained on individual datasets themselves. Even when faced with 12 unseen datasets of observed formats, UnifiedQA performs surprisingly well, showing strong generalization from its out-of-format training data. Finally, simply fine-tuning this pre-trained QA model into specialized models results in a new state of the art on 6 datasets, establishing UnifiedQA as a strong starting point for building QA systems.