Long-term learning behavior in a recurrent neural network for sound recognition

Abstract

Abstract-In this paper, the long-term learning properties of an artificial neural network model, designed for sound recognition and computational auditory scene analysis in general, are investigated. The model is designed to run for long periods of time (weeks to months) on low-cost hardware, used in a noise monitoring network, and builds upon previous work by the same authors. It consists of three neural layers, connected to each other by feedforward and feedback excitatory connections. It is shown that the different mechanisms that drive auditory attention emerge naturally from the way in which neural activation and intra-layer inhibitory connections are implemented in the model. Training of the artifical neural network is done following the Hebb principle, dictating that "Cells that fire together, wire together", with some important modifications, compared to standard Hebbian learning. As the model is designed to be on-line for extended periods of time, also learning mechanisms need to be adapted to this. The learning needs to be strongly attention-and saliency-driven, in order not to waste available memory space for sounds that are of no interest to the human listener. The model also implements plasticity, in order to deal with new or changing input over time, without catastrophically forgetting what it already learned. On top of that, it is shown that also the implementation of shortterm memory plays an important role in the long-term learning properties of the model. The above properties are investigated and demonstrated by training on real urban sound recordings.

Machine Listening for Park Soundscape Quality Assessment

“…The neural network builds on previous work by the same authors, and many of its mechanics are the same as described in detail in [21] [22], but for clarity the essential elements and differences are described in this paragraph. The network consists of a first layer, called the input layer of 768 neurons as mentioned before.…”

“…The output layer has excitatory feedback connections to the middle layer, with a time delay of one timestep, making the excitation pattern of the middle layer dependent on both the current input layer activation and the output layer activation on the previous timestep. Excitation of a neuron is calculated as the sum of the exciting inputs weighed by their respective neural connection weights, after which a normalization and saturation procedure is applied, as described in [22]. Final activation of the neuron is then calculated by means of a biologically inspired competitive selection procedure as will be explained in more detail below.…”

“…During training, these weights are then modified by both the learning and forgetting mechanisms, while limiting their values to the [0, 1] interval. For detailed analysis and mathematical details of the implementation of these mechanisms, we refer to [22].…”

“…• Bottom-up, saliency-driven attention is implemented by the choice of input features to the model, encoding intensity, spectral and temporal contrast similarly to the features for sound saliency calculation [20] [26]. The intra-layer normalization and saturation (implemented exactly as in [22]), combined with the competitive selection procedure (more details provided below) are mathematically very similar to the method used in these references, and thus the activation in the input layer reflects saliency values as calculated there. As connection weights to further layers are initially all around their base levels, activation in these layers will also reflect sali-ency.…”

Smart-Object-Based Reasoning System for Indoor Acoustic Profiling of Elderly Inhabitants

“…The same technique was recently used in a wearable device to detect sound from a heartbeat [40] and to monitor a person's health. In hardware, Boes et al [41] investigated NN learning models on low-cost devices to detect volatile events by learning signalling patterns. In their study, the authors discovered that better results could be achieved by providing greater flexibility and accommodation (in the learning model) but without losing prior knowledge.…”