Sampling (music) | EPFL Graph Search

In sound and music, sampling is the reuse of a portion (or sample) of a sound recording in another recording. Samples may comprise elements such as rhythm, melody, speech, sound effects or longer portions of music, and may be layered, equalized, sped up or slowed down, repitched, looped, or otherwise manipulated. They are usually integrated using electronic music instruments (samplers) or software such as digital audio workstations. A process similar to sampling originated in the 1940s with musique concrète, experimental music created by splicing and looping tape. The mid-20th century saw the introduction of keyboard instruments that played sounds recorded on tape, such as the Mellotron. The term sampling was coined in the late 1970s by the creators of the Fairlight CMI, a synthesizer with the ability to record and playback short sounds. As technology improved, cheaper standalone samplers with more memory emerged, such as the E-mu Emulator, Akai S950 and Akai MPC. Sampling is a foun

Multiscale light-sheet organoid imaging framework

Andrea Boni, Raphaël Ortiz, Petr Strnad

Live imaging of organoid growth remains a challenge: it requires long-term imaging of several samples simultaneously and dedicated analysis pipelines. Here the authors report an experimental and image processing framework to turn long-term light-sheet imaging of intestinal organoids into digital organoids. Organoids provide an accessible in vitro system to mimic the dynamics of tissue regeneration and development. However, long-term live-imaging of organoids remains challenging. Here we present an experimental and image-processing framework capable of turning long-term light-sheet imaging of intestinal organoids into digital organoids. The framework combines specific imaging optimization combined with data processing via deep learning techniques to segment single organoids, their lumen, cells and nuclei in 3D over long periods of time. By linking lineage trees with corresponding 3D segmentation meshes for each organoid, the extracted information is visualized using a web-based "Digital Organoid Viewer" tool allowing combined understanding of the multivariate and multiscale data. We also show backtracking of cells of interest, providing detailed information about their history within entire organoid contexts. Furthermore, we show cytokinesis failure of regenerative cells and that these cells never reside in the intestinal crypt, hinting at a tissue scale control on cellular fidelity.

NATURE PORTFOLIO2022

Stop Wasting my FLOPS: Improving the Efficiency of Deep Learning Models

Angelos Katharopoulos

Deep neural networks have completely revolutionized the field of machinelearning by achieving state-of-the-art results on various tasks ranging fromcomputer vision to protein folding. However, their application is hindered bytheir large computational and memory requirements. In this thesis, we proposemethods for improving the efficiency of deep neural networks.Firstly, we tackle the sample inefficiency of neural network training with animportance sampling algorithm suitable for deep neural networks. This algorithmallows us to focus computation on datapoints that are going to provide usefulgradients for training our models and ignore the ones that will have negligiblegradients. We show that our algorithm can improve the performance of variousneural networks when compared to uniform sampling under a fixed computationalbudget.Secondly, we design a model that is suitable for processing large input imageswith a fraction of the computational and memory requirements of traditionalapproaches. We achieve this by sampling from a data-dependent attentiondistribution in order to only process a portion of the input in highresolution. We demonstrate that our model can learn both the attention and thefeatures in an end-to-end fashion using only single image-wise labels forsupervision.Subsequently, we shift our attention to transformer architectures and introducea kernelized formulation for self-attention that reduces its quadraticcomplexity to linear with respect to the input sequence's length. Furthermore,we uncover the relationship between autoregressive transformers and recurrentneural networks and show that our formulation enables up to 3 orders ofmagnitude faster autoregressive inference.Finally, we develop clustered, attention a method that can approximate softmaxtransformers with reduced computation. This is achieved by grouping elements ofthe input using clustering. We showcase that our formulation provides a bettertrade-off between performance and computation in comparison to the originaltransformer architecture. In addition, we demonstrate that clustered attentioncan approximate pretrained transformer models without any fine-tuning and withminimal loss in performance.

EPFL2022

Stop Wasting my FLOPS: Improving the Efficiency of Deep Learning Models

Angelos Katharopoulos

EPFL2022