A neural network sensitivity analysis in the presence of random fluctuations

Sensitivity of Spiking Neural Networks Due to Input Perturbation

Background: To investigate the behavior of spiking neural networks (SNNs), the sensitivity of input perturbation serves as an effective metric for assessing the influence on the network output. However, existing methods fall short in evaluating the sensitivity of SNNs featuring biologically plausible leaky integrate-and-fire (LIF) neurons due to the intricate neuronal dynamics during the feedforward process. Methods: This paper first defines the sensitivity of a temporal-coded spiking neuron (SN) as the deviation between the perturbed and unperturbed output under a given input perturbation with respect to overall inputs. Then, the sensitivity algorithm of an entire SNN is derived iteratively from the sensitivity of each individual neuron. Instead of using the actual firing time, the desired firing time is employed to derive a more precise analytical expression of the sensitivity. Moreover, the expectation of the membrane potential difference is utilized to quantify the magnitude of the input deviation. Results/Conclusions: The theoretical results achieved with the proposed algorithm are in reasonable agreement with the simulation results obtained with extensive input data. The sensitivity also varies monotonically with changes in other parameters, except for the number of time steps, providing valuable insights for choosing appropriate values to construct the network. Nevertheless, the sensitivity exhibits a piecewise decreasing tendency with respect to the number of time steps, with the length and starting point of each piece contingent upon the specific parameter values of the neuron.

Analysis on the inherent noise tolerance of feedforward network and one noise-resilient structure

In the past few decades, feedforward neural networks have gained much attraction in their hardware implementations. However, when we realize a neural network in analog circuits, the circuit-based model is sensitive to hardware nonidealities. The nonidealities, such as random offset voltage drifts and thermal noise, may lead to variation in hidden neurons and further affect neural behaviors. This paper considers that time-varying noise exists at the input of hidden neurons, with zero-mean Gaussian distribution. First, we derive lower and upper bounds on the mean square error loss to estimate the inherent noise tolerance of a noise-free trained feedforward network. Then, the lower bound is extended for any non-Gaussian noise cases based on the Gaussian mixture model concept. The upper bound is generalized for any non-zero-mean noise case. As the noise could degrade the neural performance, a new network architecture is designed to suppress the noise effect. This noise-resilient design does not require any training process. We also discuss its limitation and give a closed-form expression to describe the noise tolerance when the limitation is exceeded.

Pre-dispersive near-infrared light sensing in non-destructively classifying the brix of intact pineapples

Exported fresh intact pineapples must fulfill the minimum internal quality requirement of 12 degree brix. Even though near-infrared (NIR) spectroscopic approaches are promising to non-destructively and rapidly assess the internal quality of intact pineapples, these approaches involve expensive and complex NIR spectroscopic instrumentation. Thus, this research evaluates the performance of a proposed pre-dispersive NIR light sensing approach in non-destructively classifying the Brix of pineapples using K-fold cross-validation, holdout validation, and sensitive analysis. First, the proposed pre-dispersive NIR sensing device that consisted of a light sensing element and five NIR light emitting diodes with peak wavelengths of 780, 850, 870, 910, and 940 nm, respectively, was developed. After that, the diffuse reflectance NIR light of intact pineapples was non-destructively acquired using the developed NIR sensing device before their Brix values were conventionally measured using a digital refractometer. Next, an artificial neural network (ANN) was trained and optimized to classify the Brix values of pineapples using the acquired NIR light. The results of the sensitivity analysis showed that either one wavelength that was near to the water absorbance or chlorophyll band was redundant in the classification. The performance of the trained ANN was tested using new pineapples with the optimal classification accuracy of 80.56%. This indicates that the proposed pre-dispersive NIR light sensing approach coupled with the ANN is promising to be an alternative to non-destructively classifying the internal quality of fruits.