This paper introduces Deep Networks for Global Optimization (DNGO), a scalable Bayesian optimization method that uses deep neural networks to model distributions over functions. Unlike traditional Gaussian processes (GPs), which scale cubically with the number of observations, DNGO scales linearly, making it suitable for large-scale hyperparameter optimization. The authors demonstrate that DNGO performs competitively with state-of-the-art GP-based approaches while achieving significantly higher degrees of parallelism. This allows for efficient hyperparameter tuning in complex models, such as convolutional neural networks for object recognition and multi-modal neural language models for image caption generation. The effectiveness of DNGO is demonstrated through experiments on benchmark problems and real-world applications, showing that it can find competitive models with significantly fewer evaluations compared to existing methods.This paper introduces Deep Networks for Global Optimization (DNGO), a scalable Bayesian optimization method that uses deep neural networks to model distributions over functions. Unlike traditional Gaussian processes (GPs), which scale cubically with the number of observations, DNGO scales linearly, making it suitable for large-scale hyperparameter optimization. The authors demonstrate that DNGO performs competitively with state-of-the-art GP-based approaches while achieving significantly higher degrees of parallelism. This allows for efficient hyperparameter tuning in complex models, such as convolutional neural networks for object recognition and multi-modal neural language models for image caption generation. The effectiveness of DNGO is demonstrated through experiments on benchmark problems and real-world applications, showing that it can find competitive models with significantly fewer evaluations compared to existing methods.