Search

One of the most well-established applications of machine learning is in deciding what content to show website visitors. When observation data comes from high-velocity, user-generated data streams, machine learning methods perform a balancing act between model complexity, training time, and computational costs. Furthermore, when model freshness is critical, the training of models becomes time-constrained. Parallelized batch offline training, although horizontally scalable, is often not time-considerate or cost-effective. In this paper, we propose Lambda Learner, a new framework for training models by incremental updates in response to mini-batches from data streams. We show that the resulting model of our framework closely estimates a periodically updated model trained on offline data and outperforms it when model updates are time-sensitive. We provide theoretical proof that the incremental learning updates improve the loss-function over a stale batch model. We present a large-scale deployment on the sponsored content platform for a large social network, serving hundreds of millions of users across different channels (e.g., desktop, mobile). We address challenges and complexities from both algorithms and infrastructure perspectives, illustrate the system details for computation, storage, stream processing training data, and open-source the system.

We present the evolution of a large-scale content recommendation platform for LinkedIn Learning, serving 645M+ LinkedIn users across several different channels (e.g., desktop, mobile). We address challenges and complexities from both algorithms and infrastructure perspectives. We describe the progression from unsupervised models that exploit member similarity with course content, to supervised learning models leveraging member interactions with courses, and finally to hyper-personalized mixed-effects models with several million coefficients. For all the experiments, we include metric lifts achieved via online A/B tests and illustrate the trade-offs between computation and storage requirements.

Different versions of a composite apparent polar wander (APW) path of variably selected global poles assembled and averaged in North American coordinates using plate reconstructions show either a smooth progression or a large (∼30°) gap in mean paleopoles in the Late Jurassic, between about 160 and 145 Ma. In an effort to further examine this issue, we sampled accessible outcrops/subcrops of kimberlites associated with high-precision U-Pb perovskite ages in the Timiskaming area of Ontario, Canada. The 154.9 ± 1.1 Ma Peddie kimberlite yields a stable normal polarity magnetization that is coaxial within less than 5° of the reverse polarity magnetization of the 157.5 ± 1.2 Ma Triple B kimberlite. The combined ∼156 Ma Triple B and Peddie pole (75.5°N, 189.5°E, A95 = 2.8°) lies about midway between igneous poles from North America nearest in age (169 Ma Moat volcanics and the 146 Ma Ithaca kimberlites), showing that the polar motion was at a relatively steady yet rapid (∼1.5°/Myr) pace. A similar large rapid polar swing has been recognized in the Middle to Late Jurassic APW path for Adria-Africa and Iran-Eurasia, suggesting a major mass redistribution. One possibility is that slab breakoff and subduction reversal along the western margin of the Americas triggered an episode of true polar wander.

Unmanned airborne vehicles (UAVs) are increasingly being used for a variety of commercial and research applications. The small (wingspan ~3 meters), fully autonomous aircraft are well suited for use in remote areas or dangerous settings. UAVs that are launched and recovered on land are being used for atmospheric chemistry and surface imaging studies and for aeromagnetic surveys [Curry et al., 2004; Ramanathan et al., 2007; Funaki et al., 2007]. We report here on the first deployment of UAVs launched from a marine research vessel. The UAVs mapped fluctuations in the magnetic field in a remote area of the southwestern Pacific Ocean.

Marine magnetic anomalies provide the framework for the geomagnetic polarity timescale for the Late Jurassic to Recent (since 160 Ma). Magnetostratigraphic records confirm that the polarity reversal sequence interpreted from magnetic anomalies is complete to a resolution of better than 30 ky. In addition to this record of polarity reversals, magnetic anomalies also appear to preserve information on geomagnetic intensity fluctuations. The correspondence of coherent near-bottom anomaly variations with independent estimates of field intensity provides strong evidence that geomagnetic intensity modulates the magnetization of the ocean crust. Indeed, many short wavelength anomaly variations in sea-surface magnetic profiles over fast-spreading ridges are likely attributable to geomagnetic intensity variations. Although longer-term geomagnetic field behavior may also be reflected in anomaly amplitudes, documenting such a signal requires a better understanding of time-dependent changes in the magnetic source (e.g., from low-temperature alteration) that may also affect magnetic anomalies. The extrusive layer, with an average remanence of ∼ 5 A m−1, is the largest contributor to magnetic anomalies. However, enhanced sampling of oceanic gabbros (average remanence ∼ 1 A m−1) and, to a lesser extent, dikes (average remanence ∼ 2 A m−1) reveals that these deeper (and thicker) layers likely generate anomalies comparable to those from the lavas. Lava accumulation at intermediate- and fast-spreading ridges typically occurs over a narrow (1–3 km) region and dike emplacement is even more narrowly confined, resulting in a relatively high fidelity record of geomagnetic field behavior. The slow cooling of the gabbroic layer, however, results in gently dipping polarity boundaries that significantly affect the skewness of the resulting anomalies, which is also a sensitive measure of net rotations of the source layer(s). The magnetizations of the dikes and gabbros are characterized by high stability and are not expected to significantly change with time, although there are insufficient data to confirm this. The lavas, however, typically show evidence of low-temperature alteration, which has been long regarded as a process that progressively reduces the magnetization (and degrades the geomagnetic signal) in the extrusive layer and reduces the amplitude of magnetic anomalies. Sufficient data have become available to examine this conventional wisdom. There is a substantial (∼ 4x) reduction in magnetization from on-axis samples to immediately off-axis drillsites (∼ 0.5 My), but little further change in half-dozen or so deep crustal sites to ∼ 160 Ma. High paleointensity that characterizes the last few thousand years may contribute significantly to the high on-axis magnetization. The task of evaluating changes in remanence of the extrusive layer is made more difficult by substantial cooling-rate-dependent changes in magnetic properties and the systematic variation in remanence with iron content (magnetic telechemistry). The commonly cited magnetic anomaly amplitude envelope is in fact not systematically observed – the Central Anomaly is elevated at slow-spreading ridges but is not as prominent at faster spreading rates. Nonetheless, magnetic anomaly amplitudes are consistent with magnetization change is poorly constrained. Direct determinations of the degree of low-temperature oxidation reveal the presence of highly oxidized titanomagnetite in samples less than 1 My old, suggesting a short (∼ 105 years) time constant though the effects of low-temperature oxidation are quite heterogeneous. While low-temperature oxidation does have some affect on lava magnetization and anomaly amplitudes, there is increasing evidence that marine magnetic anomalies are capable of recording a broad spectrum of geomagnetic field behavior, ranging from millennial-scale paleointensity variations to polarity reversals to apparent polar wander to, more speculatively, long-term changes in average field strength. Several emerging tools and approaches – autonomous vehicles, oriented samples, absolute paleointensity of near-ridge lavas, and measurements of the vector anomalous field – are therefore likely to significantly advance our understanding of the geomagnetic signal recorded in the oceanic crust, as well as our ability to utilize this information in addressing outstanding problems in crustal accretion processes.

Placing accurate age constraints on near-axis lava flows has become increasingly important given the structural and volcanic complexity of the neovolcanic zone at fast spreading ridges. Geomagnetic paleointensity of submarine basaltic glass (SBG) holds promise for placing quantitative age constraints on near-axis flows. In one of the first extensive tests of paleointensity as a dating tool or temporal marker we present the results of over 550 successful SBG paleointensity estimates from 189 near-axis (

Cooling rate effects on the intensity of thermoremanent magnetization (TRM) have been well documented in ceramics. In that case, laboratory cooling is generally more rapid than the initial cooling, leading to an overestimate of the paleofield by 5-10% in Thellier-type paleointensity experiments. The reverse scenario, however, has never been tested. We examine the effects of cooling rate on paleointensity estimates from rapidly quenched submarine basaltic glass (SBG) samples from 13 sites at 17°30'-18°30'S on the East Pacific Rise. Absolute cooling rates determined by relaxation geospeedometry at five of these sites range from ~10 to ~330°C min⁻¹ at the glass transition (~650°C). Over the dominant range of remanence blocking temperatures (~200-400°C), the natural cooling rates are approximately equal to or slightly slower than the laboratory cooling rates during the Thellier experiment. These results suggest that while the cooling rate effect might introduce some within-site scatter, it should not result in a systematic bias in paleointensity from SBG. Paleointensity estimates from the 15 sites range from ~29 to 59μT, with an average standard error of ~1μT. Comparison with models of geomagnetic field intensity variations at the site indicate the youngest group of samples is very recent (indistinguishable from present-day) and the oldest is at least 500, and probably several thousand, years old. These age estimates are consistent with available radiometric ages and geologic observations.