CIF: Small: Learning Signal Representations for Multiple Inference Tasks

高性能计算的快速发展和海量数据集的广泛使用正在导致信号表示的理论和实践的范式转变，面向推理和学习。信号表示是一种压缩的摘要，它只保留信号的那些对于一类推理任务来说是显著的特征。该项目为信号表示提供了一个全面的理论和算法框架，该框架足够广泛，既涵盖了传统类型的信号表示，如矢量量化和稀疏码，也涵盖了受大数据机器学习和信号处理最新进展启发的更现代类型。在该框架下，通过对编码映射、解码映射和表示的模型空间施加结构约束，同时将这些对象定制为感兴趣的任务类别，以统一的方式解决信号表示的统计性能和计算复杂性。这种统一导致了对高度结构化的内部表示的新的理论和算法见解，这些内部表示是深度神经网络最近在视觉、音频和语音分析方面的挑战性任务取得惊人成功的关键因素。

Rapid advances in high-performance computing and widespread availability of massive datasets are bringing about a paradigm shift in the theory and practice of signal representations, geared towards inference and learning. A signal representation is a compressed summary that only retains those features of the signal that are salient for a class of inference tasks. This project provides a comprehensive theoretical and algorithmic framework for signal representations, which is sufficiently broad to cover both the traditional types of signal representations, such as vector quantization and sparse codes, and the more modern types, inspired by recent advances in machine learning and signal processing for Big Data. Under this framework, the statistical performance and the computational complexity of signal representations are addressed in a unified manner by imposing structural constraints on the encoding map, the decoding map, and the model space of the representation, while simultaneously tailoring these objects to the class of tasks of interest. This unification leads to new theoretical and algorithmic insights into highly structured internal representations that are a key factor in recent spectacular success of deep neural networks on challenging tasks in visual, audio, and speech analytics.