Audio computing in the wild: frameworks for big data and small computers

Machine Learning;Signal Processing;Blind Source Separation;Big Data;Deep Learning;Neural Networks

Kim, Minje

关键词: Machine Learning;    Signal Processing;    Blind Source Separation;    Big Data;    Deep Learning;    Neural Networks;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/90573/KIM-DISSERTATION-2016.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

This dissertation presents some machine learning algorithms that are designed to process as much data as needed while spending the least possible amount of resources, such as time, energy, and memory. Examples of those applications, but not limited to, can be a large-scale multimedia information retrieval system where both queries and the items in the database are noisy signals; collaborative audio enhancement from hundreds of user-created clips of a music concert; an event detection system running in a small device that has to process various sensor signals in real time; a lightweight custom chipset for speech enhancement on hand-held devices; instant music analysis engine running on smartphone apps. In all those applications, efficient machine learning algorithms are supposed to achieve not only a good performance, but also a great resource-efficiency. We start from some efficient dictionary-based single-channel source separation algorithms. We can train this kind of source-specific dictionaries by using some matrix factorization or topic modeling, whose elements form a representative set of spectra for the particular source. During the test time, the system estimates the contribution of the participating dictionary items for an unknown mixture spectrum. In this way we can estimate the activation of each source separately, and then recover the source of interest by using that particular source's reconstruction. There are some efficiency issues during this procedure. First off, searching for the optimal dictionary size is time consuming. Although for some very common types of sources, e.g. English speech, we know the optimal rank of the model by trial and error, it is hard to know in advance as to what is the optimal number of dictionary elements for the unknown sources, which are usually modeled during the test time in the semi-supervised separation scenarios. On top of that, when it comes to the non-stationary unknown sources, we had better maintain a dictionary that adapts its size and contents to the change of the source's nature. In this online semi-supervised separation scenario, a mechanism that can efficiently learn the optimal rank is helpful. To this end, a deflation method is proposed for modeling this unknown source with a nonnegative dictionary whose size is optimal. Since it has to be done during the test time, the deflation method that incrementally adds up new dictionary items shows better efficiency than a corresponding na\"ive approach where we simply try a bunch of different models. We have another efficiency issue when we are to use a large dictionary for better separation. It has been known that considering the manifold of the training data can help enhance the performance for the separation. This is because of the symptom that the usual manifold-ignorant convex combination models, such as from low-rank matrix decomposition or topic modeling, tend to result in ambiguous regions in the source-specific subspace defined by the dictionary items as the bases. For example, in those ambiguous regions, the original data samples cannot reside. Although some source separation techniques that respect data manifold could increase the performance, they call for more memory and computational resources due to the fact that the models call for larger dictionaries and involve sparse coding during the test time. This limitation led the development of hashing-based encoding of the audio spectra, so that some computationally heavy routines, such as nearest neighbor searches for sparse coding, can be performed in a cheaper bit-wise fashion. Matching audio signals can be challenging as well, especially if the signals are noisy and the matching task involves a big amount of signals. If it is an information retrieval application, for example, the bigger size of the data leads to a longer response time. On top of that, if the signals are defective, we have to perform the enhancement or separation job in the first place before matching, or we might need a matching mechanism that is robust to all those different kinds of artifacts. Likewise, the noisy nature of signals can add an additional complexity to the system. In this dissertation we will also see some compact integer (and eventually binary) representations for those matching systems. One of the possible compact representations would be a hashing-based matching method, where we can employ a particular kind of hash functions to preserve the similarity among original signals in the hash code domain. We will see that a variant of Winner Take All hashing can provide Hamming distance from noise-robust binary features, and that matching using the hash codes works well for some keyword spotting tasks. From the fact that some landmark hashes (e.g. local maxima from non-maximum suppression on the magnitudes of a mel-scaled spectrogram) can also robustly represent the time-frequency domain signal efficiently, a matrix decomposition algorithm is also proposed to take those irregular sparse matrices as input. Based on the assumption that the number of landmarks is a lot smaller than the number of all the time-frequency coefficients, we can think of this matching algorithm efficient if it operates entirely on the landmark representation. On the contrary to the usual landmark matching schemes, where matching is defined rigorously, we see the audio matching problem as soft matching where we find a similar constellation of landmarks to the query. In order to perform this soft matching job, the landmark positions are smoothed by a fixed-width Gaussian caps, with which the matching job is reduced down to calculating the amount of overlaps in-between those Gaussians. The Gaussian-based density approximation is also useful when we perform decomposition on this landmark representation, because otherwise the landmarks are usually too sparse to perform an ordinary matrix factorization algorithm, which are originally for a dense input matrix. We also expand this concept to the matrix deconvolution problem as well, where we see the input landmark representation of a source as a two-dimensional convolution between a source pattern and its corresponding sparse activations. If there are more than one source, as a noisy signal, we can think of this problem as factor deconvolution where the mixture is the combination of all the source-specific convolutions.The dissertation also covers Collaborative Audio Enhancement (CAE) algorithms that aim to recover the dominant source at a sound scene (e.g. music signals of a concert rather than the noise from the crowd) from multiple low-quality recordings (e.g. Youtube video clips uploaded by the audience). CAE can be seen as crowdsourcing a recording job, which needs a substantial amount of denoising effort afterward, because the user-created recordings might have been contaminated with various artifacts. In the sense that the recordings are from not-synchronized heterogenous sensors, we can also think of CAE as big ad-hoc sensor array processing. In CAE, each recording is assumed to be uniquely corrupted by a specific frequency response of the microphone, an aggressive audio coding algorithm, interference, band-pass filtering, clipping, etc. To consolidate all these recordings and come up with an enhanced audio, Probabilistic Latent Component Sharing (PLCS) has been proposed as a method of simultaneous probabilistic topic modeling on synchronized input signals. In PLCS, some of the parameters are fixed to be same during and after the learning process to capture common audio content, while the rest of the parameters are for the unwanted recording-specific interference and artifacts. We can speed up PLCS by incorporating a hashing-based nearest neighbor search so that at every EM iteration PLCS can be applied only to a small number of recordings that are closest to the current source estimation. Experiments on a small simulated CAE setup shows that the proposed PLCS can improve the sound quality from variously contaminated recordings. The nearest neighbor search technique during PLCS provides sensible speed-up at larger scaled experiments (up to 1000 recordings).Finally, to describe an extremely optimized deep learning deployment system, Bitwise Neural Networks (BNN) will be also discussed. In the proposed BNN, all the input, hidden, and output nodes are binaries (+1 and -1), and so are all the weights and bias. Consequently, the operations on them during the test time are defined with Boolean algebra, too. BNNs are spatially and computationally efficient in implementations, since (a) we represent a real-valued sample or parameter with a bit (b) the multiplication and addition correspond to bitwise XNOR and bit-counting, respectively. Therefore, BNNs can be used to implement a deep learning system in a resource-constrained environment, so that we can deploy a deep learning system on small devices without using up the power, memory, CPU clocks, etc. The training procedure for BNNs is based on a straightforward extension of backpropagation, which is characterized by the use of the quantization noise injection scheme, and the initialization strategy that learns a weight-compressed real-valued network only for the initialization purpose.Some preliminary results on the MNIST dataset and speech denoising demonstrate that a straightforward extension of backpropagation can successfully train BNNs whose performance is comparable while necessitating vastly fewer computational resources.