Improving the efficiency of Transformer-based language pre-training is an important task in NLP, especially for the self-attention module, which is computationally expensive. In this paper, we propose a simple but effective solution, called \emph{LazyFormer}, which computes the self-attention distribution infrequently. LazyFormer composes of multiple lazy blocks, each of which contains multiple Transformer layers. In each lazy block, the self-attention distribution is only computed once in the first layer and then is reused in all upper layers. In this way, the cost of computation could be largely saved. We also provide several training tricks for LazyFormer. Extensive experiments demonstrate the effectiveness of the proposed method.

In a February 19 speech at the Munich Security Conference, delivered virtually from the White House, President Joe Biden declared, "We must shape the rules that will govern the advance of technologies and the norms of behavior in cyberspace, artificial intelligence, biotechnology, so they are used to lift people up, not used to pin them down." A few weeks earlier, during an address at the State Department's Truman Building, the president said, "Diplomacy is back at the center of our foreign policy." The Trump administration's undermining of years of work on internet diplomacy makes technology an ever more vital (and challenging) element of renewed US engagement abroad. Digital issues are no longer extricable from "traditional" foreign policy issues across trade, human rights, and security. And as the new White House starts to navigate these waters, one idea in particular has become a sort of bumper sticker for an overarching strategy: Unite democracies on technology. As the Chinese and Russian governments become more technologically assertive and undermine human rights, and as democracies grapple with how to appropriately implement rules and regulations for the likes of artificial intelligence systems, this work is essential.

A pizza box sized solar panel in orbit is producing enough electricity to power an iPad, according to a succesful test of the technology by the US Navy. The Photovoltaic Radiofrequency Antenna Module (PRAM) was launched in May 2020 attached to a drone that loops around the Earth every 90 minutes and is designed to harness light from the sun to convert to electricity. The 12x12 inch panel is an early experiment for a technology that could one day harness solar radiation from the sun and beam it to anywhere on the Earth. It is designed to make the best use of light in space, which doesn't have to pass through the atmosphere where it loses energy before reaching the ground. The Pentagon one day envisages an array of panels in space that could send power to even the most remote parts of the planet and create a new global power grid.

They had spent many months honing an algorithm designed to steer an unmanned hot air balloon all the way from Puerto Rico to Peru. The balloon, controlled by its machine mind, kept veering off course. Salvatore Candido of Google's now-defunct Project Loon venture, which aimed to bring internet access to remote areas via the balloons, couldn't explain the craft's trajectory. His colleagues manually took control of the system and put it back on track. It was only later that they realised what was happening.

Learning composable policies for environments with complex rules and tasks is a challenging problem. We introduce a hierarchical reinforcement learning framework called the Logical Options Framework (LOF) that learns policies that are satisfying, optimal, and composable. LOF efficiently learns policies that satisfy tasks by representing the task as an automaton and integrating it into learning and planning. We provide and prove conditions under which LOF will learn satisfying, optimal policies. And lastly, we show how LOF's learned policies can be composed to satisfy unseen tasks with only 10-50 retraining steps. We evaluate LOF on four tasks in discrete and continuous domains, including a 3D pick-and-place environment.

With the advent of digital technology, it is more common that committed crimes or legal disputes involve some form of speech recording where the identity of a speaker is questioned [1]. In face of this situation, the field of forensic speaker identification has been looking to shed light on the problem by quantifying how much a speech recording belongs to a particular person in relation to a population. In this work, we explore the use of speech embeddings obtained by training a CNN using the triplet loss. In particular, we focus on the Spanish language which has not been extensively studies. We propose extracting the embeddings from speech spectrograms samples, then explore several configurations of such spectrograms, and finally, quantify the embeddings quality. We also show some limitations of our data setting which is predominantly composed by male speakers. At the end, we propose two approaches to calculate the Likelihood Radio given out speech embeddings and we show that triplet loss is a good alternative to create speech embeddings for forensic speaker identification.

Several recent works have demonstrated the use of variational autoencoders (VAEs) for both generating levels in the style of existing games as well as blending levels across different games. Additionally, quality-diversity (QD) algorithms have also become popular for generating varied game content by using evolution to explore a search space while focusing on both variety and quality. In order to reap the benefits of both these approaches, we present a level generation and game blending approach that combines the use of VAEs and QD algorithms. Specifically, we train VAEs on game levels and then run the MAP-Elites QD algorithm using the learned latent space of the VAE as the search space. The latent space captures the properties of the games whose levels we want to generate and blend, while MAP-Elites searches this latent space to find a diverse set of levels optimizing a given objective such as playability. We test our method using models for 5 different platformer games as well as a blended domain spanning 3 of these games. Our results show that using MAP-Elites in conjunction with VAEs enables the generation of a diverse set of playable levels not just for each individual game but also for the blended domain while illuminating game-specific regions of the blended latent space.

Natural Language Processing algorithms have made incredible progress recently, but they still struggle when applied to out-of-distribution examples. In this paper, we address a very challenging and previously underexplored version of this domain adaptation problem. In our setup an algorithm is trained on several source domains, and then applied to examples from an unseen domain that is unknown at training time. Particularly, no examples, labeled or unlabeled, or any other knowledge about the target domain are available to the algorithm at training time. We present PADA: A Prompt-based Autoregressive Domain Adaptation algorithm, based on the T5 model. Given a test example, PADA first generates a unique prompt and then, conditioned on this prompt, labels the example with respect to the NLP task. The prompt is a sequence of unrestricted length, consisting of pre-defined Domain Related Features (DRFs) that characterize each of the source domains. Intuitively, the prompt is a unique signature that maps the test example to the semantic space spanned by the source domains. In experiments with two tasks: Rumour Detection and Multi-Genre Natural Language Inference (MNLI), for a total of 10 multi-source adaptation scenarios, PADA strongly outperforms state-of-the-art approaches and additional strong baselines.

The 1980s saw the evolution of computing as it transitioned from industry to homes through the boom of personal computers. In the field of data management, the Relational Database industry was developing rapidly (Oracle, Sybase, IBM, among others). Object-oriented abstractions were developed as a new form of representational independence. The Internet changed the way people communicated and exchanged information.

Support vector machine (SVM) is one of the most popular classification algorithms in the machine learning literature. We demonstrate that SVM can be used to balance covariates and estimate average causal effects under the unconfoundedness assumption. Specifically, we adapt the SVM classifier as a kernel-based weighting procedure that minimizes the maximum mean discrepancy between the treatment and control groups while simultaneously maximizing effective sample size. We also show that SVM is a continuous relaxation of the quadratic integer program for computing the largest balanced subset, establishing its direct relation to the cardinality matching method. Another important feature of SVM is that the regularization parameter controls the trade-off between covariate balance and effective sample size. As a result, the existing SVM path algorithm can be used to compute the balance-sample size frontier. We characterize the bias of causal effect estimation arising from this trade-off, connecting the proposed SVM procedure to the existing kernel balancing methods. Finally, we conduct simulation and empirical studies to evaluate the performance of the proposed methodology and find that SVM is competitive with the state-of-the-art covariate balancing methods.