Deep Neural Network

Many neural networks exhibit an overconfidence calibration bias, where output confidence scores remain close to 1 even when the error rate is high.

Modern AI models (especially deep neural networks and ensemble models) often have a black-box character, and their internal decision paths are difficult to trace, which creates a fundamental tension with the principles of auditability, traceability and accountability required by financial regulation.

The technical path of the LaserBATMAN project is divided into three steps: building an electrochemical cell model, training a reduced-order model with a deep neural network (DNN), and building a customized simulation app based on the reduced-order model.

In deep neural networks, the number of parameters can reach billions, and each parameter participates in the final prediction computation.

据IEEE Spectrum报道，随着深度神经网络规模持续膨胀，参数数量和计算量急剧上升，行业正越来越多地转向面向特定任务的AI芯片设计。

The study derives the marginal likelihood of a DNN using a probabilistic framework, and this quantity can be evaluated using only the weights of the network's last layer.

If the weight matrices after training exhibit a heavy-tailed spectral distribution, this is often a strong indicator of better model quality.

Convexity implies the existence of a unique optimal heavy-tail degree, and this optimal value can empirically predict the occurrence of overfitting on large-scale DNNs.

This study is one of the first attempts to systematically apply random matrix theory to provide a rigorous statistical foundation for empirical metrics of DNN weight spectra.

Although using only the last-layer weights is computationally efficient, it may miss information contained in intermediate layers.

Research teams at institutions such as MIT and Stanford have introduced deep neural networks to accelerate holographic computation, compressing single-frame hologram generation time from several minutes down to the millisecond level.

Whether the conclusions of this paper can be generalized to more complex modern models, such as deep neural networks, remains an open question.

Modern pricing models are typically based on gradient boosting trees (such as XGBoost) or deep neural networks, may take hundreds of input features, and produce results that are difficult to describe with human-readable rules.

Increasing network depth allows expressive capacity (number of segments) to grow exponentially, whereas increasing the width of a single layer yields only linear growth.

Deep neural networks are often more accurate than rule-based systems on target recognition tasks, but their decision-making processes are difficult for non-specialists to audit.

Neural networks encounter vanishing gradient problems as layers are stacked deeper, making training difficult to converge, and it was not until around 2010 that a series of technical breakthroughs solved the challenge of training deep networks, giving rise to the term deep neural network (DNN).

Deep neural networks contain billions or even trillions of parameters, and their black-box nature makes it difficult even for developers to accurately predict model behavior for specific inputs.

谷歌早在 2012 年就将深度神经网络引入其语音识别系统