Successes and critical failures of neural networks in capturing human-like speech recognition

What is the impact of discrete memristor on the performance of neural network: A research on discrete memristor-based BP neural network

Artificial neural networks are receiving increasing attention from researchers. However, with the advent of big data era, artificial neural networks are limited by the Von Neumann architecture, making it difficult to make new breakthroughs in hardware implementation. Discrete-time memristor, emerging as a research focus in recent years, are anticipated to address this challenge effectively. To enrich the theoretical research of memristors in artificial neural networks, this paper studies BP neural networks based on various discrete memristors. Firstly, the concept of discrete memristor and several classical discrete memristor models are introduced. Based on these models, the discrete memristor-based BP neural networks are designed. Finally, these networks are utilized for achieving handwritten digit classification and speech feature classification, respectively. The results show that linear discrete memristors perform better than nonlinear discrete memristors, and a simple linear discrete memristor-based BP neural network has the best performance, reaching 97.40% (handwritten digit classification) and 93.78% (speech feature classification), respectively. In addition, some fundamental issues are also discussed, such as the effects of linear, nonlinear memristors, and initial charges on the performance of neural networks.

Models optimized for real-world tasks reveal the task-dependent necessity of precise temporal coding in hearing

Neurons encode information in the timing of their spikes in addition to their firing rates. Spike timing is particularly precise in the auditory nerve, where action potentials phase lock to sound with sub-millisecond precision, but its behavioral relevance remains uncertain. We optimized machine learning models to perform real-world hearing tasks with simulated cochlear input, assessing the precision of auditory nerve spike timing needed to reproduce human behavior. Models with high-fidelity phase locking exhibited more human-like sound localization and speech perception than models without, consistent with an essential role in human hearing. However, the temporal precision needed to reproduce human-like behavior varied across tasks, as did the precision that benefited real-world task performance. These effects suggest that perceptual domains incorporate phase locking to different extents depending on the demands of real-world hearing. The results illustrate how optimizing models for realistic tasks can clarify the role of candidate neural codes in perception.

Models optimized for real-world tasks reveal the task-dependent necessity of precise temporal coding in hearing

Neurons encode information in the timing of their spikes in addition to their firing rates. Spike timing is particularly precise in the auditory nerve, where action potentials phase lock to sound with sub-millisecond precision, but its behavioral relevance remains uncertain. We optimized machine learning models to perform real-world hearing tasks with simulated cochlear input, assessing the precision of auditory nerve spike timing needed to reproduce human behavior. Models with high-fidelity phase locking exhibited more human-like sound localization and speech perception than models without, consistent with an essential role in human hearing. However, the temporal precision needed to reproduce human-like behavior varied across tasks, as did the precision that benefited real-world task performance. These effects suggest that perceptual domains incorporate phase locking to different extents depending on the demands of real-world hearing. The results illustrate how optimizing models for realistic tasks can clarify the role of candidate neural codes in perception.

A regularized orthogonal activated inverse-learning neural network for regression and classification with outliers

A novel regularized orthogonal activated inverse-learning (ROAIL) neural network is proposed and investigated for reducing the impact of outliers in regression and classification fields. The proposed ROAIL network does not require extensive iterative computations. Instead, it can achieve the desired results with a single step of computation, allowing for the efficient acquisition of network weights. By extending the Gegenbauer polynomials to a multi-variate version, and integrating the ℓ2 regularization and Welsch loss function into the orthogonal activated inverse-learning framework, two forms of ROAIL are obtained, i.e., ℓ2 norm ROAIL (ℓ2-ROAIL) and Welsch-ROAIL (W-ROAIL). ℓ2-ROAIL neural network is proposed to minimize the empirical and structural risk simultaneously since taking the structural risk as a part of loss function can effectively reduce the complexity of the model and thus improve the generalization ability. W-ROAIL neural network further improves the robustness of the ℓ2-ROAIL neural network by replacing the original two-norm in loss function with Welsch function. The Welsch function can determine the weights of each sample according to its output error, and influence of outliers could be weakened since the weights of outliers would be reduced. Both regression and classification experiments show that W-ROAIL neural network has strong ability to suppress the influence of outliers.

An ensemble-acute lymphoblastic leukemia model for acute lymphoblastic leukemia image classification

The timely diagnosis of acute lymphoblastic leukemia (ALL) is of paramount importance for enhancing the treatment efficacy and the survival rates of patients. In this study, we seek to introduce an ensemble-ALL model for the image classification of ALL, with the goal of enhancing early diagnostic capabilities and streamlining the diagnostic and treatment processes for medical practitioners. In this study, a publicly available dataset is partitioned into training, validation, and test sets. A diverse set of convolutional neural networks, including InceptionV3, EfficientNetB4, ResNet50, CONV_POOL-CNN, ALL-CNN, Network in Network, and AlexNet, are employed for training. The top-performing four individual models are meticulously chosen and integrated with the squeeze-and-excitation (SE) module. Furthermore, the two most effective SE-embedded models are harmoniously combined to create the proposed ensemble-ALL model. This model leverages the Bayesian optimization algorithm to enhance its performance. The proposed ensemble-ALL model attains remarkable accuracy, precision, recall, F1-score, and kappa scores, registering at 96.26, 96.26, 96.26, 96.25, and 91.36%, respectively. These results surpass the benchmarks set by state-of-the-art studies in the realm of ALL image classification. This model represents a valuable contribution to the field of medical image recognition, particularly in the diagnosis of acute lymphoblastic leukemia, and it offers the potential to enhance the efficiency and accuracy of medical professionals in the diagnostic and treatment processes.

Many but not all deep neural network audio models capture brain responses and exhibit correspondence between model stages and brain regions

Models that predict brain responses to stimuli provide one measure of understanding of a sensory system and have many potential applications in science and engineering. Deep artificial neural networks have emerged as the leading such predictive models of the visual system but are less explored in audition. Prior work provided examples of audio-trained neural networks that produced good predictions of auditory cortical fMRI responses and exhibited correspondence between model stages and brain regions, but left it unclear whether these results generalize to other neural network models and, thus, how to further improve models in this domain. We evaluated model-brain correspondence for publicly available audio neural network models along with in-house models trained on 4 different tasks. Most tested models outpredicted standard spectromporal filter-bank models of auditory cortex and exhibited systematic model-brain correspondence: Middle stages best predicted primary auditory cortex, while deep stages best predicted non-primary cortex. However, some state-of-the-art models produced substantially worse brain predictions. Models trained to recognize speech in background noise produced better brain predictions than models trained to recognize speech in quiet, potentially because hearing in noise imposes constraints on biological auditory representations. The training task influenced the prediction quality for specific cortical tuning properties, with best overall predictions resulting from models trained on multiple tasks. The results generally support the promise of deep neural networks as models of audition, though they also indicate that current models do not explain auditory cortical responses in their entirety.

Deep problems with neural network models of human vision

Deep neural networks (DNNs) have had extraordinary successes in classifying photographic images of objects and are often described as the best models of biological vision. This conclusion is largely based on three sets of findings: (1) DNNs are more accurate than any other model in classifying images taken from various datasets, (2) DNNs do the best job in predicting the pattern of human errors in classifying objects taken from various behavioral datasets, and (3) DNNs do the best job in predicting brain signals in response to images taken from various brain datasets (e.g., single cell responses or fMRI data). However, these behavioral and brain datasets do not test hypotheses regarding what features are contributing to good predictions and we show that the predictions may be mediated by DNNs that share little overlap with biological vision. More problematically, we show that DNNs account for almost no results from psychological research. This contradicts the common claim that DNNs are good, let alone the best, models of human object recognition. We argue that theorists interested in developing biologically plausible models of human vision need to direct their attention to explaining psychological findings. More generally, theorists need to build models that explain the results of experiments that manipulate independent variables designed to test hypotheses rather than compete on making the best predictions. We conclude by briefly summarizing various promising modeling approaches that focus on psychological data.