This paper explores deep multi-view representation learning, focusing on techniques that use multiple unlabeled views of data to learn representations for downstream tasks. The authors analyze and compare several approaches, including deep autoencoders and deep extensions of canonical correlation analysis (CCA), and introduce a new variant called deep canonically correlated autoencoders (DCCAE). They find that correlation-based representation learning outperforms reconstruction-based methods, with DCCAE consistently performing well across various tasks. The paper also discusses stochastic optimization for minibatch correlation-based objectives and compares kernel-based and neural network-based implementations in terms of time and performance. The authors provide empirical and theoretical analyses, including comparisons with linear and kernel CCA, and release their implementations and a benchmark dataset for future research.This paper explores deep multi-view representation learning, focusing on techniques that use multiple unlabeled views of data to learn representations for downstream tasks. The authors analyze and compare several approaches, including deep autoencoders and deep extensions of canonical correlation analysis (CCA), and introduce a new variant called deep canonically correlated autoencoders (DCCAE). They find that correlation-based representation learning outperforms reconstruction-based methods, with DCCAE consistently performing well across various tasks. The paper also discusses stochastic optimization for minibatch correlation-based objectives and compares kernel-based and neural network-based implementations in terms of time and performance. The authors provide empirical and theoretical analyses, including comparisons with linear and kernel CCA, and release their implementations and a benchmark dataset for future research.