Search for: All records

We present a continuous high-resolution precisely dated multiproxy record of hydroclimate variability at Anjohibe cave in northwestern Madagascar using speleothem AB13. The record spans from ~4484 y BP to ~2863 y BP. Stalagmite δ¹⁸O, δ¹³C and Sr/Ca ratios show very similar changes in hydroclimate. The mechanism controlling Sr/Ca changes, however, from prior calcite precipitation to degree of dolomite dissolution at about 4 ky BP. Our record is also in good agreement with previously published speleothem records from the same area. This agreement and multiproxy consensus indicate that AB13 provides a robust record of hydroclimate variability, including a continuous record of hydroclimate variability across the 4.2 ka event. This 4.2 ka event in Madagascar is marked by two distinct periods of drying between ~3900 y BP to 4300 y BP. A dry 4.2 ka event at this Southern Hemisphere site helps limit possible mechanisms for the event, indicating that a meridional shift to the south in the ITCZ is not responsible for the 4.2 ka event. In addition, the 4.2 ka event does not stand out as a unique dry period in our record. The longest and driest period of the record lasted ~300 years with peak dryness at ~3000 y BP. Our record differs significantly from a speleothem record from Rodrigues Island, located ~1800 km to the east of our study area in Madagascar suggesting different climatological controls on northwest Madagascar and more oceanic sites to the east. 

Audio-visual representation learning aims to develop systems with human-like perception by utilizing correlation between auditory and visual information. However, current models often focus on a limited set of tasks, and generalization abilities of learned representations are unclear. To this end, we propose the AV-SUPERB benchmark that enables general-purpose evaluation of unimodal audio/visual and bimodal fusion representations on 7 datasets covering 5 audio-visual tasks in speech and audio processing. We evaluate 5 recent self-supervised models and show that none of these models generalize to all tasks, emphasizing the need for future study on improving universal model performance. In addition, we show that representations may be improved with intermediate-task fine-tuning and audio event classification with AudioSet serves as a strong intermediate task. We release our benchmark with evaluation code and a model submission platform to encourage further research in audio-visual learning.