This paper introduces a pseudo-Siamese Network for Asset Pricing (SNAP) model that integrates deep learning with conditional asset pricing theory. The model uses three LSTM sub-networks to estimate time-varying alpha, beta, and factor risk premia from high-dimensional firm characteristics and macroeconomic variables. The pseudo-Siamese architecture allows these components to be learned jointly in an adversarial manner, improving out-of-sample prediction. Applied to monthly U.S. stock returns from 1970 to 2019, the model achieves an out-of-sample Sharpe ratio above 2 and outperforms traditional factor models and machine learning benchmarks like LASSO and feedforward networks. The authors also propose a novel test for mispricing by comparing residuals from the full model and a version without the alpha sub-network, using the Mann-Whitney U test, which rejects the null of zero alphas. Finally, they construct arbitrage portfolios based on deep alphas and apply K-Means clustering to analyze their Sharpe ratios over time, finding that while popular and unpopular portfolios persist in the short run, their performance differences diminish in the long run, empirically validating Benjamin Graham's quote about the market as a short-term voting machine and long-term weighing machine.