RIP Theano

Before TensorFlow, PyTorch and Caffe; Theano was the major library for deep learning development. However, the library’s development and support will end after the upcoming Theano 1.0 release.

The news came in an email from Theano’s main developer Pascal Lamblin and Yoshua Bengio, notable expert on artificial neural networks and deep learning. “We will continue minimal maintenance to keep it working for one year, but we will stop actively implementing new features. Supporting Theano is no longer the best way we can enable the emergence and application of novel research ideas,” wrote Prof. Bengio.

Lack of continuous support and maintenance is considered the biggest reason for Theano’s demise. Unlike Google and Microsoft which have maintained support for TensorFlow and Caffe, the Montreal Institute for Leaning Algorithms (MILA), creator and maintainer of Theano, cannot afford to keep the library running on a regular basis.

Founded in 2010, Theano is an open-source library that helps users to provide mathematical expressions in Python, and combines the convenience of NumPy’s syntax with the speed of optimized native machine language. Theano can use GPUs and perform efficient symbolic differentiation.

Theano inspired both academia and industry with various innovations, such as expressing models as mathematical expressions, rewriting computation graphs for better performance and memory usage, transparent execution on GPU, and higher-order automatic differentiation. Notable contributors to Theano included experts such as Ian Goodfellow, Staff Research Scientist at Google Brain and godfather of GANs(Generative adversarial networks).

Theano supporters responded to its discontinuation with both regret and gratitude. Reddit user “Deskates” wrote: “Its important to think about the impact which Theano actually created. Many people approached deep learning (thanks to Theano). I feel a debt of gratitude towards those who contributed to it over the years, making it such a great tool. I remember reading the tutorials at deeplearning.net which was probably the starting point for me.”

As an early inspirational library, Theano was deeply involved in the spawning of state-of-the-art deep learning algorithms. It may be inevitable to see a once mainstream library retire, but Theano’s legacy probably lives in many of today’s deep learning frameworks.

Journalist: Tony Peng | Editor: Michael Sarazen