Kambhampati, 2020; Etzioni and DeCario, 2020; Vaishya et al., 2020; Wynants and colleagues, 2020; Artificial Intelligence (AI) technologies have long Srivastava, 2020]. Early in the pandemic, authors been positioned as a tool to provide crucial datadriven like [Kambhampati, 2020; Etzioni and DeCario, 2020; decision support to people. In this survey Vaishya et al., 2020] highlighted various scenarios where paper, we look at how AI in general, and collaboration AI could help in tackling COVID19 as well as some of the assistants (CAs or chatbots for short) in particular, potential pitfalls. The AI efforts were helped by different have been used during a true global exigency types of data being freely made available, calls for open - the COVID-19 pandemic. The key observation collaboration [Woodward, 2020] and a sense of urgency. is that chatbots missed their Apollo moment In Table 1, a sample of AI's potential application during when they could have really provided contextual, COVID-19 is shown. They range from decisions to foster personalized, reliable decision support at scale that understanding of the disease and its impact to helping take the state-of-the-art makes possible. We review the actions for individuals, groups and the society at large.

Graph Neural Networks (GNNs), a generalization of deep neural networks on graph data have been widely used in various domains, ranging from drug discovery to recommender systems. However, GNNs on such applications are limited when there are few available samples. Meta-learning has been an important framework to address the lack of samples in machine learning, and in recent years, the researchers have started to apply meta-learning to GNNs. In this work, we provide a comprehensive survey of different meta-learning approaches involving GNNs on various graph problems showing the power of using these two approaches together. We categorize the literature based on proposed architectures, shared representations, and applications. Finally, we discuss several exciting future research directions and open problems.

In recent years, there has been an explosion of AI research on counterfactual explanations as a solution to the problem of eXplainable AI (XAI). These explanations seem to offer technical, psychological and legal benefits over other explanation techniques. We survey 100 distinct counterfactual explanation methods reported in the literature. This survey addresses the extent to which these methods have been adequately evaluated, both psychologically and computationally, and quantifies the shortfalls occurring. For instance, only 21% of these methods have been user tested. Five key deficits in the evaluation of these methods are detailed and a roadmap, with standardised benchmark evaluations, is proposed to resolve the issues arising; issues, that currently effectively block scientific progress in this field.

Zero-shot learning (ZSL) which aims at predicting classes that have never appeared during the training using external knowledge (a.k.a. side information) has been widely investigated. In this paper we present a literature review towards ZSL in the perspective of external knowledge, where we categorize the external knowledge, review their methods and compare different external knowledge. With the literature review, we further discuss and outlook the role of symbolic knowledge in addressing ZSL and other machine learning sample shortage issues.

Variable importance measures are the main tools to analyze the black-box mechanism of random forests. Although the Mean Decrease Accuracy (MDA) is widely accepted as the most efficient variable importance measure for random forests, little is known about its theoretical properties. In fact, the exact MDA definition varies across the main random forest software. In this article, our objective is to rigorously analyze the behavior of the main MDA implementations. Consequently, we mathematically formalize the various implemented MDA algorithms, and then establish their limits when the sample size increases. In particular, we break down these limits in three components: the first two are related to Sobol indices, which are well-defined measures of a variable contribution to the output variance, widely used in the sensitivity analysis field, as opposed to the third term, whose value increases with dependence within input variables. Thus, we theoretically demonstrate that the MDA does not target the right quantity when inputs are dependent, a fact that has already been noticed experimentally. To address this issue, we define a new importance measure for random forests, the Sobol-MDA, which fixes the flaws of the original MDA. We prove the consistency of the Sobol-MDA and show its good empirical performance through experiments on both simulated and real data. An open source implementation in R and C++ is available online.

Human-made physical random number generators (RNGs) can be traced back 5000 years or more. Early examples such as knucklebones, two-sided throwsticks, or dice have been found in the Middle East, India, and China. RNGs were used for fortune telling and games of chance, with the oldest known board games of similar age as those of the number generators. Today, RNGs are vital for services and state-of-the-art technologies such as cryptographically secured communication, blockchain technologies, and quantum key distribution. Moreover, RNGs are needed in machine learning and scientific applications such as Monte Carlo numerical methods. On page 948 of this issue, Kim et al. ([ 1 ][1]) demonstrate an ultrafast RNG based on a broad-area laser with a multispot beam that is analogous to generating random numbers by using many dice at once. Random numbers are often generated by using a software algorithm running on a computer, called “pseudo”-random because the sequence eventually repeats. Moreover, relations among the numbers can exist that reveal that the numbers are not uniformly random. Hence, true RNGs (TRNGs) are of great interest, providing random numbers based on physical measurements that involve some noisy or stochastic process. All TRNGs have some nonidealities, such as generating zeroes more frequently than ones for a binary-output device, which must be mitigated by carefully engineering the device and postprocessing the data to improve the randomness quality ([ 2 ][2]). ![Figure][3] Creating bits with laser intensity Ultrafast random bits are generated from a broad-area laser with a bow-tie cavity. In order to generate the random bits, first intensities separated by ∼6 ps are subtracted from each other for the same positions on the detector. This creates bits of either ones or zeroes, which then undergo the exclusive-OR (XOR) logic operation with bits of another spot separated by half the width of the aperture. The XOR operation produces a one if the two inputs are different or a zero if the two inputs are the same. The broad-area laser allows for many different positions on the detector to be used simultaneously, allowing for fast generation of bits. GRAPHIC: N.DESAI/ SCIENCE Some applications require generating random numbers at very high rates, such as encrypting data in cloud-computing data centers, high-speed communication networks, or massive simulations. Photonic devices are a natural fit for these applications because of their potential for high-speed operation, compact size for chip-scale devices, and low power consumption. Recently, Marangon et al. ([ 3 ][4]) developed a TRNG that is based on interfering two different lasers on a beam splitter and detecting the resulting powers that emanate from its two output ports. The randomness comes about from quantum fluctuations in a laser due to a process known as spontaneous emission of photons. This process randomizes the phase of the light emitted by each laser, and this phase variation is converted to an intensity variation through the interference effect. Measuring which output port of the interferometer has the higher or lower intensity can be used to generate a one or a zero, respectively, at random. A compact device can be realized by generating random numbers in real time at a rate of 8 Gb/s for days at a time, passing tests that are used to assess the quality of the bit stream. The bottleneck in reaching higher speeds is that the lasers are single-mode and only generate a Gaussian beam–like spot at a single frequency. Kim et al. overcome this bottleneck by using a broad-area laser that simultaneously emits a plethora of modes, resulting in a multispot beam. The patterns undergo a complex dance, writhing and growing bright and dim because of phase and amplitude variation of the light within the laser (see the figure). For a good TRNG, engineering the broad-area laser cavity is especially necessary so that spatial and temporal correlations are minimized. The authors do so, which is a major achievement. Common broad-area lasers are known to exhibit irregular intensity pulsations in space and time because of the nonlinear interaction of light and the laser medium ([ 4 ][5]). Such instabilities result in correlations of their emission with characteristic spatial and temporal scales and are exceedingly difficult to avoid. This situation has plagued attempts to apply broad-area lasers more widely. Bittner et al. ([ 5 ][6]) showed that they could largely suppress the onset of spatiotemporal instabilities by using a cavity with a D shape, inspired by chaotic billiards; the balls on a D-shaped billiard table follow chaotic trajectories ([ 6 ][7]). Kim et al. introduce another approach based on adapting the shape of the cavity. After performing extensive numerical modeling, the authors chose a bow-tie shape and precisely microfabricated a laser chip. The authors managed to boost the number of modes, avoiding their locking, and thereby substantially reduced the spatial and temporal correlation scales to 1.5 µm and 2.8 ps, respectively. Another advantage of using the spatial degree of freedom of the special laser design is avoiding the two separate lasers and interference on an auxiliary beam splitter ([ 3 ][4]). Random numbers can be “harvested” from the complex emitted pattern by measuring the intensity at 254 spatial positions on ultrafast time scales (on the order of 1 ps) by using a special high-speed camera. This strategy is truly an ultrafast Demeter meeting chance. Through this effort, they achieved a random bit generation rate of 250 Tb/s, which is much more than an order of magnitude greater than previous efforts. A full technical implementation of such an ultrafast TRNG still faces several challenges that need to be overcome. The high-speed camera could only capture data over a limited time (∼2 ns), so they had to collect and concatenate multiple records to generate the more than 109 random numbers needed for the various statistical tests of randomness. Replacing the camera with a multitude of integrated photodetectors is yet to be achieved. Also, the required postprocessing of the measured intensities to ensure randomness is, at such speed, a task for the future. Looking beyond, the innovative approach to tailor the spatial and temporal emission properties of broad-area lasers and manipulating the nonlinear interaction of light with the laser medium opens other applications that require many degrees of freedom. Several machine-learning approaches are based on a random mapping of low-dimensional input data onto a high-dimensional-state space, which might be accomplished by injecting a data-encoded beam into a tailored laser. Hence, broad-area lasers may become attractive photonic integrated circuits for ultrafast information processing ([ 7 ][8], [ 8 ][9]). 1. [↵][10]1. K. Kim et al ., Science 371, 948 (2021). [OpenUrl][11][CrossRef][12] 2. [↵][13]1. J. D. Hart et al ., Appl. Phys. Lett. Photonics 2, 090901 (2017). [OpenUrl][14] 3. [↵][15]1. D. G. Marangon et al ., J. Lightwave Technol. 36, 3778 (2018). [OpenUrl][16] 4. [↵][17]1. I. Fischer, 2. O. Hess, 3. W. Elsäßer, 4. E. Göbel , Europhys. Lett. 35, 579 (1996). [OpenUrl][18][CrossRef][19] 5. [↵][20]1. S. Bittner et al ., Science 361, 1225 (2018). [OpenUrl][21][Abstract/FREE Full Text][22] 6. [↵][23]1. H. Cao, 2. J. Wiersig , Rev. Mod. Phys. 87, 61 (2015). [OpenUrl][24][CrossRef][25][PubMed][26] 7. [↵][27]1. P. R. Prucnal, 2. B. J. Shastri , Neuromorphic Photonics (CRC Press, 2017). 8. [↵][28]1. D. Brunner, 2. M. C. Soriano, 3. G. Van der Sande , Eds., Photonic Reservoir Computing (De Gruyter, 2019). [1]: #ref-1 [2]: #ref-2 [3]: pending:yes [4]: #ref-3 [5]: #ref-4 [6]: #ref-5 [7]: #ref-6 [8]: #ref-7 [9]: #ref-8 [10]: #xref-ref-1-1 "View reference 1 in text" [11]: {openurl}?query=rft.jtitle%253DScience%26rft.stitle%253DScience%26rft.aulast%253DKim%26rft.auinit1%253DK.%26rft.volume%253D371%26rft.issue%253D6532%26rft.spage%253D948%26rft.epage%253D952%26rft.atitle%253DMassively%2Bparallel%2Bultrafast%2Brandom%2Bbit%2Bgeneration%2Bwith%2Ba%2Bchip-scale%2Blaser%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.abc2666%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [12]: /lookup/external-ref?access_num=10.1126/science.abc2666&link_type=DOI [13]: #xref-ref-2-1 "View reference 2 in text" [14]: {openurl}?query=rft.jtitle%253DAppl.%2BPhys.%2BLett.%2BPhotonics%26rft.volume%253D2%26rft.spage%253D090901%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [15]: #xref-ref-3-1 "View reference 3 in text" [16]: {openurl}?query=rft.jtitle%253DJ.%2BLightwave%2BTechnol.%26rft.volume%253D36%26rft.spage%253D3778%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [17]: #xref-ref-4-1 "View reference 4 in text" [18]: {openurl}?query=rft.jtitle%253DEurophys.%2BLett.%26rft.volume%253D35%26rft.spage%253D579%26rft_id%253Dinfo%253Adoi%252F10.1209%252Fepl%252Fi1996-00154-7%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [19]: /lookup/external-ref?access_num=10.1209/epl/i1996-00154-7&link_type=DOI [20]: #xref-ref-5-1 "View reference 5 in text" [21]: {openurl}?query=rft.jtitle%253DScience%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.aas9437%26rft_id%253Dinfo%253Apmid%252F30115744%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [22]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6Mzoic2NpIjtzOjU6InJlc2lkIjtzOjEzOiIzNjEvNjQwOC8xMjI1IjtzOjQ6ImF0b20iO3M6MjI6Ii9zY2kvMzcxLzY1MzIvODg5LmF0b20iO31zOjg6ImZyYWdtZW50IjtzOjA6IiI7fQ== [23]: #xref-ref-6-1 "View reference 6 in text" [24]: {openurl}?query=rft.jtitle%253DRev.%2BMod.%2BPhys.%26rft.volume%253D87%26rft.spage%253D61%26rft_id%253Dinfo%253Adoi%252F10.1103%252FRevModPhys.87.61%26rft_id%253Dinfo%253Apmid%252F25739324%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [25]: /lookup/external-ref?access_num=10.1103/RevModPhys.87.61&link_type=DOI [26]: /lookup/external-ref?access_num=25739324&link_type=MED&atom=%2Fsci%2F371%2F6532%2F889.atom [27]: #xref-ref-7-1 "View reference 7 in text" [28]: #xref-ref-8-1 "View reference 8 in text"

This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Another prominent example in this regard came from DeepMind's publication of the possible protein structures associated with the COVID-19 virus (SARS-CoV-2) using their AlphaFold system. For example, our process of vetting results in the Global Burden of Disease Study [2] included the visual inspection of thousands of plots showing data together with model estimates. Our experience developing methods for computer certification of verbal autopsy has bolstered our belief that using an explainable approach, even with a reduction in accuracy, can be superior. Qualified practitioners are in short supply. There is increasing awareness that health … enhancing the ability to see and navigate in a procedure. Going beyond the conventional long-haul process, AI techniques are increasingly being applied to accelerate the fundamental processes of early-stage candidate selection and mechanism discovery. This could be the biggest impact of AI tools as it can potentially transform the quality of life for billions of people around the world. These technologies are also being used in the following ways: Preventing crime: AI and machine learning help authorities track and manage the huge amount of data generated by public surveillance devices, and analyze that data in real time for anomalies and threats.

Modern natural language processing (NLP) methods employ self-supervised pretraining objectives such as masked language modeling to boost the performance of various application tasks. These pretraining methods are frequently extended with recurrence, adversarial or linguistic property masking, and more recently with contrastive learning objectives. Contrastive self-supervised training objectives enabled recent successes in image representation pretraining by learning to contrast input-input pairs of augmented images as either similar or dissimilar. However, in NLP, automated creation of text input augmentations is still very challenging because a single token can invert the meaning of a sentence. For this reason, some contrastive NLP pretraining methods contrast over input-label pairs, rather than over input-input pairs, using methods from Metric Learning and Energy Based Models. In this survey, we summarize recent self-supervised and supervised contrastive NLP pretraining methods and describe where they are used to improve language modeling, few or zero-shot learning, pretraining data-efficiency and specific NLP end-tasks. We introduce key contrastive learning concepts with lessons learned from prior research and structure works by applications and cross-field relations. Finally, we point to open challenges and future directions for contrastive NLP to encourage bringing contrastive NLP pretraining closer to recent successes in image representation pretraining.

The widespread adoption of black-box models in Artificial Intelligence has enhanced the need for explanation methods to reveal how these obscure models reach specific decisions. Retrieving explanations is fundamental to unveil possible biases and to resolve practical or ethical issues. Nowadays, the literature is full of methods with different explanations. We provide a categorization of explanation methods based on the type of explanation returned. We present the most recent and widely used explainers, and we show a visual comparison among explanations and a quantitative benchmarking.

Making decisions in complex environments is a key challenge in artificial intelligence (AI). Situations involving multiple decision makers are particularly complex, leading to computational intractability of principled solution methods. A body of work in AI has tried to mitigate this problem by trying to distill interaction to its essence: how does the policy of one agent influence another agent? If we can find more compact representations of such influence, this can help us deal with the complexity, for instance by searching the space of influences rather than the space of policies. However, so far these notions of influence have been restricted in their applicability to special cases of interaction. In this paper we formalize influence-based abstraction (IBA), which facilitates the elimination of latent state factors without any loss in value, for a very general class of problems described as factored partially observable stochastic games (fPOSGs). On the one hand, this generalizes existing descriptions of influence, and thus can serve as the foundation for improvements in scalability and other insights in decision making in complex multiagent settings. On the other hand, since the presence of other agents can be seen as a generalization of single agent settings, our formulation of IBA also provides a sufficient statistic for decision making under abstraction for a single agent. We also give a detailed discussion of the relations to such previous works, identifying new insights and interpretations of these approaches. In these ways, this paper deepens our understanding of abstraction in a wide range of sequential decision making settings, providing the basis for new approaches and algorithms for a large class of problems.