A deep learning framework is the software that does the heavy lifting of building and training neural networks so that researchers and engineers do not have to write it from scratch. The two dominant frameworks are TensorFlow (open-sourced by Google in 2015) and PyTorch (released by Facebook AI Research in 2017).
Two capabilities are the heart of any such framework. The first is automatic differentiation (autodiff): training a neural network means repeatedly computing how a tiny change in each of millions of parameters would change the error, and the framework computes all of those gradients automatically from the model code. The second is GPU abstraction: the framework maps the same model onto fast graphics-processor hardware without the user writing low-level GPU code, building on layers like NVIDIA’s CUDA. The PyTorch paper frames its version of this as an imperative, define-by-run style where “every aspect of PyTorch is a regular Python program under the full control of its user.”
Why open-sourcing them mattered: before free frameworks, these capabilities lived inside corporate labs or scattered academic code, and reproducing a published result was slow and error-prone. By releasing industrial-grade frameworks under permissive licenses (Google shipped TensorFlow under Apache 2.0), the major labs turned deep learning from specialist plumbing into something a competent developer could pick up. That standardization - shared tools, shared model formats - is a large part of why progress accelerated through the 2010s and into the large language model era.