Adaptive Pooling Operators for Weakly Labeled Sound Event Detection

By Brian McFee, Justin Salamon, Juan Pablo Bello. 2018

161 citations recorded by the source corpus.

Sound event detection (SED) methods are tasked with labeling segments of audio recordings by the presence of active sound sources. SED is typically posed as a supervised machine learning problem, requiring strong annotations for the presence or absence of each sound source at every time instant within the recording. However, strong annotations of this type are both labor- and cost-intensive for human annotators to produce, which limits the practical scalability of SED methods. In this paper, we treat SED as a multiple instance learning (MIL) problem, where training labels are static over a short excerpt, indicating the presence or absence of sound sources but not their temporal locality. The models, however, must still produce temporally dynamic predictions, which must be aggregated (pooled) when comparing against static labels during training. To facilitate this aggregation, we develop a family of adaptive pooling operators—referred to as autopool—which smoothly interpolate between common pooling operators, such as min-, max-, or average-pooling, and automatically adapt to the characteristics of the sound sources in question. We evaluate the proposed pooling operators on three datasets, and demonstrate that in each case, the proposed methods outperform nonadaptive pooling operators for static prediction, and nearly match the performance of models trained with strong, dynamic annotations. The proposed method is evaluated in conjunction with convolutional neural networks, but can be readily applied to any differentiable model for time-series label prediction. While this paper focuses on SED applications, the proposed methods are general, and could be applied widely to MIL problems in any domain.

Subfields: Computer Vision and Pattern Recognition, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Signal Processing

Citations in the corpus

FSD50K: An Open Dataset of Human-Labeled Sound Events (2020, background): “Weakly supervised learning has demonstrated effectiveness to learn sound event recognizers, both for classification and detection [42].”
Active Learning for Sound Event Detection (2020, methodology): “Previous weakly supervised learning studies [6], [7], [18], [21], [22] use pooling functions to aggregate frame-level class probabilities into segment-level.”
Limitations of Weak Labels for Embedding and Tagging (2020, unknown): “The use of weakly labeled data to train a sound event detection system, which outputs labels together with their time localization in a segment of audio, has been studied in recent years [15–17].”
Evaluation of CNN-based Automatic Music Tagging Models (2020, background): “In most cases, we do not have time precise instance-level annotations because the precise labeling can be laborsome, therefore a music tag associated with a song is simply applied to all music excerpts (instances) of the song during training [12].”
SONYC-UST-V2: An Urban Sound Tagging Dataset with Spatiotemporal Context (2020, methodology): “For the baseline model 3, we use a multi-label multi-layer perceptron model, using a single hidden layer of size 128 (with ReLU non-linearities), and using AutoPool [16] to aggregate frame level predictions.”
Connecting the Dots between Audio and Text without Parallel Data through Visual Knowledge Transfer (2021, background): “Others use weaker forms of supervision for tagging (Kumar and Raj, 2017; Kong et al., 2018) and localization (McFee et al., 2018; Kim and Pardo, 2019).”
Sound Event Detection in Domestic Environments with Weakly Labeled Data and Soundscape Synthesis (2019, unknown): “Sound event detection (SED) systems trained using weak labels have have seen signiﬁcant interest [6, 9, 10, 11, 12] in the research community, as they address some of the challenges involved in developing models that require strongly labeled data for training.”
Large Scale Audiovisual Learning of Sounds with Weakly Labeled Data (2020, methodology): “While mean and max pooling have been shown to be effective in handling weakly labeled data [Kumar, 2018], more involved methods such as adaptive pooling [McFee et al., 2018] or attention based pooling [Wang et al., 2019b; Kong et al., 2019] have also been proposed.”
Sound Event Detection Using Point-Labeled Data (2019, methodology): “Deep neural networks have been successfully applied to recent SED systems [4, 5, 6] and are the current state-of-the-art.”
Torchaudio-Squim: Reference-Less Speech Quality and Intelligibility Measures in Torchaudio (2023, background): “As outlined in [27], the auto-pooling operator can automatically adapt to the characteristics of representations via a learnable , where x 1 , . . . , x L is a sequence of features.”
The Benefit of Temporally-Strong Labels in Audio Event Classification (2021, methodology): “Other works use strategies including pooling [9] and attention [12] to train on weak labels while still producing scores at fine time resolutions.”
SeCoST: Sequential Co-Supervision for Weakly Labeled Audio Event Detection (2019, background): “Several works have shown promising results on weakly labeled AED [7], [8], [9], [10].”