Using relative geologic time to constrain CNN-based seismic interpretation and property estimation

Three-dimensional seismic interpretation and property estimation is essential to subsurface mapping and characterization, in which machine learning, particularly supervised convolutional neural network (CNN) has been extensively implemented for improved efficiency and accuracy in the past years. In most seismic applications, however, the amount of available expert annotations is often limited, which raises the risk of overfitting a CNN particularly when only seismic amplitudes are used for learning. In such a case, the trained CNN would have poor generalization capability, causing the interpretation and property results of obvious artifacts, limited lateral consistency and thus restricted application to following interpretation/modeling procedures. This study proposes addressing such an issue by using relative geologic time (RGT), which explicitly preserves the large-scale continuity of seismic patterns, to constrain a seismic interpretation and/or property estimation CNN. Such constrained learning is enforced in twofold: (1) from the perspective of input, the RGT is used as an additional feature channel besides seismic amplitude; and more innovatively (2) the CNN has two output branches, with one for matching the target interpretation or properties and the other for reconstructing the RGT. In addition is the use of multiplicative regularization to facilitate the simultaneous minimization of the target-matching loss and the RGT-reconstruction loss. The performance of such an RGT-constrained CNN is validated by two examples, including facies identification in the Parihaka dataset and property estimation in the F3 Netherlands dataset. Compared to those purely from seismic amplitudes, both the facies and property predictions with using the proposed RGT constraint demonstrate significantly reduced artifacts and improved lateral consistency throughout a seismic survey.

L2 Acquisition of a Complex Stress Pattern: UG-Constrained Learning Paths in Khalkha Mongolian

This paper examines second language (L2) acquisition of stress in Khalkha Mongolian, which is one of the few Default-to-Opposite Edge stress systems of the world, and as such, demonstrates “conflicting directionality” regarding stress assignment, resulting in the leftmost edge of a word being more prominent in certain words and the rightmost edge in certain others. Given the additional fact that the language exhibits Non-finality effects, and that, unlike English, codas are not moraic, its acquisition presents unique difficulties and challenges for English-speaking learners of the language. Many of these challenges potentially lead these learners to make Universal Grammar (UG)-unconstrained (but cognitively reasonable) assumptions about how the phonology of Mongolian works, especially since the learners do not have all the Mongolian data available to them all at once. The learning scenario here, thus, provides unique opportunities to investigate whether L2 phonologies are constrained by the options made available by UG. The findings of a semi-controlled production experiment indicate that although learners do not necessarily converge on the prosodic representations employed by native speakers of the L2 (i.e., footless intonational prominence, at least for the leftmost/default edge ‘stress’), and although certain changes to the grammar are very difficult to implement, such as switching from moraic codas to non-moraic codas, the learners nevertheless demonstrate a stage-like behavior where each step exhibits the parameter settings employed by a natural language, one that is neither like the L2 nor the L1. Conversely, despite the input leading them to do so, learners do not entertain UG-unconstrained prosodic representations, such as End-Rule-Middle or End-Rule-Variable; End-Rule is set either to Right or Left, as is expected in a system constrained by the options made available by UG. We conclude that the hypothesis space for interlanguage phonologies is determined by UG.

Optimal auctions through deep learning

Designing an incentive compatible auction that maximizes expected revenue is an intricate task. The single-item case was resolved in a seminal piece of work by Myerson in 1981. Even after 30--40 years of intense research, the problem remains unsolved for settings with two or more items. We overview recent research results that show how tools from deep learning are shaping up to become a powerful tool for the automated design of near-optimal auctions auctions. In this approach, an auction is modeled as a multilayer neural network, with optimal auction design framed as a constrained learning problem that can be addressed with standard machine learning pipelines. Through this approach, it is possible to recover to a high degree of accuracy essentially all known analytically derived solutions for multi-item settings and obtain novel mechanisms for settings in which the optimal mechanism is unknown.

A Survey of On-Device Machine Learning

The predominant paradigm for using machine learning models on a device is to train a model in the cloud and perform inference using the trained model on the device. However, with increasing numbers of smart devices and improved hardware, there is interest in performing model training on the device. Given this surge in interest, a comprehensive survey of the field from a device-agnostic perspective sets the stage for both understanding the state of the art and for identifying open challenges and future avenues of research. However, on-device learning is an expansive field with connections to a large number of related topics in AI and machine learning (including online learning, model adaptation, one/few-shot learning, etc.). Hence, covering such a large number of topics in a single survey is impractical. This survey finds a middle ground by reformulating the problem of on-device learning as resource constrained learning where the resources are compute and memory. This reformulation allows tools, techniques, and algorithms from a wide variety of research areas to be compared equitably. In addition to summarizing the state of the art, the survey also identifies a number of challenges and next steps for both the algorithmic and theoretical aspects of on-device learning.