Listen to the Waves: Using a Neuronal Model of the Human Auditory System to Predict Ocean Waves

TL;DR

The paper addresses real-time, phase-resolved ocean-wave forecasting in shallow water by introducing audESN, a neuroscience-informed echo-state network whose reservoir mirrors auditory cortex architecture and whose input is FFT-preprocessed. Compared to a standard randomly connected ESN, audESN with a core-belt-parabelt, tonotopically organized reservoir and online readout learning delivers consistently lower RMS prediction errors, especially under harsher sea states, while using a compact network (~128 units). The approach demonstrates real-time, wave-by-wave forecasting with a two-wave-period horizon and robust performance across reservoir parameters, suggesting viability for embedded deployment and nearshore energy applications. The work also highlights the value of integrating neuroscience principles into ML architectures to enhance efficiency and predictive quality for complex time-series tasks.

Abstract

Artificial neural networks (ANNs) have evolved from the 1940s primitive models of brain function to become tools for artificial intelligence. They comprise many units, artificial neurons, interlinked through weighted connections. ANNs are trained to perform tasks through learning rules that modify the connection weights. With these rules being in the focus of research, ANNs have become a branch of machine learning developing independently from neuroscience. Although likely required for the development of truly intelligent machines, the integration of neuroscience into ANNs has remained a neglected proposition. Here, we demonstrate that designing an ANN along biological principles results in drastically improved task performance. As a challenging real-world problem, we choose real-time ocean-wave prediction which is essential for various maritime operations. Motivated by the similarity of ocean waves measured at a single location to sound waves arriving at the eardrum, we redesign an echo state network to resemble the brain's auditory system. This yields a powerful predictive tool which is computationally lean, robust with respect to network parameters, and works efficiently across a wide range of sea states. Our results demonstrate the advantages of integrating neuroscience with machine learning and offer a tool for use in the production of green energy from ocean waves.

Figures (6)

• Figure 1: Modeling the auditory cortex. (A) In the auditory system, sound waves are transduced in the cochlea into electric signals. These propagate up the auditory pathway to the auditory cortex via the cochlear nucleus and the superior olivary nucleus in the brainstem as well as the inferior colliculus (IC) and the medial geniculate nucleus (MGN) in thalamus as part of the midbrain. (B) Schematic view of the primate auditory cortex 16 and the two midbrain components. The input from the ascending auditory pathway, i.e. from IC and thalamus, arrives in the core area comprising three tonotopically organized fields. The core fields are surrounded by eight belt fields and two parabelt fields. The black and blue arrows indicate reciprocal connections between fields, with high or low density of connections, respectively. (C) The structure of the auditory cortex in (B) is represented as a connection matrix, which can be used to simulate the functioning of auditory cortex. In this model, each field comprises multiple units, and each dot in the matrix representation shows a connection between two units. The superposition of two matrices indicating excitatory-to-excitatory connections and excitatory-to-inhibitory connections are shown 17 (for details, see appendix A). The color bar on the right represents the strength of the connections between any two columns: positive values (in red) stand for excitatory, negative values (in blue) for inhibitory connections (adapted from Hajizadeh et al. 17).
• Figure 2: Simplified model of auditory cortex. (A) Schematic representation of the core-belt structure of the AC model used in the wave prediction algorithm. The single core field receives the input signal which is then distributed to three surrounding belt fields. The arrows indicate the connections between the core and belt fields. (B) The connection matrices corresponding to the four fields shown in (A) are shown (for details, see appendix A). In the case shown here, there are $16$ columns per field. Each matrix element represents the strength of the connection between two columns. The left matrix shows the superposition of the matrices indicating excitatory-to-excitatory connection ($W_{ee}$, red) and excitatory-to-inhibitory connections ($W_{ie}$, blue). The matrix on the right shows the two (overlaid) diagonal matrices representing inhibitory-to-inhibitory connections ($W_{ii}$) and inhibitory-to-excitatory ($W_{ei}$) connections.
• Figure 3: Structure of the reservoir. The blue waveform represents the incoming signal. On the left, the reservoir of a standard ESN (stdESN) features random connections, and the input feeds directly into all units (only five input connections shown). On the right, the simplified version of the reservoir of the auditory ESN (audESN) described in Fig. 2 is shown. The incoming signal to the audESN is first pre-processed by a Fast Fourier Transform (FFT). Red triangles and curves represent excitatory units and connections. Inhibitory units and connections are in blue. Output units are not shown.
• Figure 4: Prediction performance of the networks. This example shows how stdESN (top row) and audESN (bottom row) track the wave data (black curve) during training ($< 900$ s, green curve) and prediction ($> 900$ s, red curve). The network training was updated online at every other wave trough, as indicated by the vertical lines. Clearly, audESN provides much better predictions than stdESN which is evident in the two enlarged time windows shown on the right.
• Figure 5: Prediction performance. (A) Example of the prediction performance of $16$ networks. The figure shows, for each network type, the RMS error between simulated and predicted waves over $1.5$ h for five sea states whose significant wave height (first row) and peak wave period (second row) are given at the bottom of the figure. For each sea state, tests were carried out for 125 combinations of the three network parameters: leak rate $\alpha$, spectral radius $\rho$, and maximum bias $\beta$ (for details, see appendix A). The four-digit codes on the abscissa denote the four binary network construction choices: neuron model, architecture, input domain, and network size (see section 3 and appendix A). (B) The median RMS errors of stdESN (‘0001’, open circles) and audESN (‘1111’, filled circles) with the corresponding confidence intervals (CIs) are shown. Among all sea states, the lowest RMS errors are achieved with audESN. Importantly, audESN also shows the best performance stability against variations in sea state and parameter choice. Note that the CIs for all five sea states are so small that they are covered by the size of the median symbol.