On layer-wise representations in deep neural networks
نویسنده
چکیده
On Layer-Wise Representations in Deep Neural Networks It is well-known that deep neural networks are forming an efficient internal representation of the learning problem. However, it is unclear how this efficient representation is distributed layer-wise, and how it arises from learning. In this thesis, we develop a kernel-based analysis for deep networks that quantifies the representation at each layer in terms of noise and dimensionality. The analysis is applied to backpropagation networks and deep Boltzmann machines, and is able to capture the layer-wise reduction of noise and dimensionality. The analysis also reveals the disrupting effect of learning noise, and how it prevents the emergence of highly sophisticated deep models. Zusammenfassung Schichtweise Repräsentationen in Tiefen Neuronalen Netzen Es ist bekannt, dass tiefe neuronale Netze eine effiziente interne Repräsentation des Lernproblems bilden. Es ist jedoch unklar, wie sich diese effiziente Repräsentation über die Schichten verteilt und wie sie beim Lernen entsteht. In dieser Arbeit entwickeln wir eine Kernel-basierte Analyse für tiefe Netze. Diese Analyse quantifiziert die Repräsentation in jeder Schicht in Bezug auf Rauschen und Dimensionalität. Wir wenden die Analyse auf Backpropagation-Netze und tiefe Boltzmann-Maschinen an und messen die schichtweise Reduzierung von Rauschen und Dimensionalität. Die Analyse zeigt auch den störenden Einfluss des Lernrauschens: Dieses verhindert die Entstehung komplexer Strukturen in tiefen Modellen.
منابع مشابه
Cost-Sensitive Deep Learning with Layer-Wise Cost Estimation
While deep neural networks have succeeded in several visual applications, such as object recognition, detection, and localization, by reaching very high classification accuracies, it is important to note that many real-world applications demand varying costs for different types of misclassification errors, thus requiring cost-sensitive classification algorithms. Current models of deep neural ne...
متن کاملLearning to Prune Deep Neural Networks via Layer-wise Optimal Brain Surgeon
How to develop slim and accurate deep neural networks has become crucial for realworld applications, especially for those employed in embedded systems. Though previous work along this research line has shown some promising results, most existing methods either fail to significantly compress a well-trained deep network or require a heavy retraining process for the pruned deep network to re-boost...
متن کاملDeep Boltzmann Machines as Feed-Forward Hierarchies
The deep Boltzmann machine is a powerful model that extracts the hierarchical structure of observed data. While inference is typically slow due to its undirected nature, we argue that the emerging feature hierarchy is still explicit enough to be traversed in a feedforward fashion. The claim is corroborated by training a set of deep neural networks on real data and measuring the evolution of the...
متن کاملمعرفی شبکه های عصبی پیمانه ای عمیق با ساختار فضایی-زمانی دوگانه جهت بهبود بازشناسی گفتار پیوسته فارسی
In this article, growable deep modular neural networks for continuous speech recognition are introduced. These networks can be grown to implement the spatio-temporal information of the frame sequences at their input layer as well as their labels at the output layer at the same time. The trained neural network with such double spatio-temporal association structure can learn the phonetic sequence...
متن کاملExploring Strategies for Training Deep Neural Networks
Deep multi-layer neural networks have many levels of non-linearities allowing them to compactly represent highly non-linear and highly-varying functions. However, until recently it was not clear how to train such deep networks, since gradient-based optimization starting from random initialization often appears to get stuck in poor solutions. Hinton et al. recently proposed a greedy layer-wise u...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
عنوان ژورنال:
دوره شماره
صفحات -
تاریخ انتشار 2013