This paper summarizes the music demixing (MDX) track of the Sound Demixing Challenge (SDX’23). We provide a summary of the challenge setup and introduce the task of robust music source separation (MSS), i.e., training MSS models in the presence of errors in the training data. We propose a formalization of the errors that can occur in the design of a training dataset for MSS systems and introduce two new datasets that simulate such errors: SDXDB23_LabelNoise and SDXDB23_Bleeding.1 We describe the methods that achieved the highest scores in the competition. Moreover, we present a direct comparison with the previous edition of the challenge (the Music Demixing Challenge 2021): the best performing system achieved an improvement of over 1.6dB in signal-to-distortion ratio over the winner of the previous competition, when evaluated on MDXDB21. Besides relying on the signal-to-distortion ratio as objective metric, we also performed a listening test with renowned producers and musicians to study the perceptual quality of the systems and report here the results. Finally, we provide our insights into the organization of the competition and our prospects for future editions.

While spatial proteomics by fluorescence imaging has quickly become an essential discovery tool for researchers, fast and scalable methods to classify and embed single-cell protein distributions in such images are lacking. Here, we present the design and analysis of the results from the competition Human Protein Atlas – Single-Cell Classification hosted on the Kaggle platform. This represents a crowd-sourced competition to develop machine learning models trained on limited annotations to label single-cell protein patterns in fluorescent images. The particular challenges of this competition include class imbalance, weak labels and multi-label classification, prompting competitors to apply a wide range of approaches in their solutions. The winning models serve as the first subcellular omics tools that can annotate single-cell locations, extract single-cell features and capture cellular dynamics.

Modern computational tasks are often required to not only guarantee predefined accuracy, but get the result fast. Optimizing calculations using floating point numbers, as opposed to integers, is a non-trivial task. For this reason, there is a need to explore new ways to improve such operations. This paper presents analysis and comparison of various floating point formats - float, posit and bfloat. One of the promising areas in which the problem of using such values can be considered to be the most acute is neural networks. That is why we pay special attention to algorithms of linear algebra and artificial intelligence to assess efficiency of new data types in this area. The research results showed that software implementations of posit16 and posit32 have high accuracy, but they are not particularly fast; on the other hand, bfloat16 is not much different from float32 in accuracy, but significantly surpasses it in performance for large amounts of data and complex machine learning algorithms. Thus, posit16 can be used in systems with less stringent performance requirements, as well as in conditions of limited computer memory; and also in cases when bfloat16 cannot provide required accuracy. As for bfloat16, it can speed up systems based on the IEEE 754 standard, but it cannot solve all the problems of conventional floating point arithmetic. Thus, although posits and bfloats are not a full fledged replacement for float, they provide (under certain conditions) advantages that can be useful for implementation of machine learning algorithms.

This paper summarizes the cinematic demixing (CDX) track of the Sound Demixing Challenge 2023 (SDX’23). We provide a comprehensive summary of the challenge setup, detailing the structure of the competition and the datasets used. Especially, we detail CDXDB23, a new hidden dataset constructed from real movies that was used to rank the submissions. The paper also offers insights into the most successful approaches employed by participants. Compared to the cocktail-fork baseline, the best-performing system trained exclusively on the simulated Divide and Remaster (DnR) dataset achieved an improvement of 1.8 dB in SDR, whereas the top-performing system on the open leaderboard, where any data could be used for training, saw a significant improvement of 5.7 dB. A major source of this improvement was making the simulated data better match real cinematic audio, which we further investigate in detail.

Modern mobile neural networks with a reduced number of weights and parameters do a good job with image classification tasks, but even they may be too complex to be implemented in an FPGA for video processing tasks. The article proposes neural network architecture for the practical task of recognizing images from a camera, which has several advantages in terms of speed. This is achieved by reducing the number of weights, moving from a floating-point to a fixed-point arithmetic, and due to a number of hardware-level optimizations associated with storing weights in blocks, a shift register, and an adjustable number of convolutional blocks that work in parallel. The article also proposed methods for adapting the existing data set for solving a different task. As the experiments showed, the proposed neural network copes well with real-time video processing even on the cheap FPGAs.