This is a pleasant surprise. The more people that work on TensorFlow the better, especially as the DeepMind team will be more aligned with extending TensorFlow's research potential.
I am curious how well TensorFlow fits for many of DeepMind's tasks though. Much of their recent work has been in reinforcement algorithms and hard stochastic decision tasks (think gradient approximation via Monte Carlo simulations rather than exactly computed gradients) which TensorFlow hasn't traditionally been used for.
Has anyone seen TensorFlow efficiently used for such tasks? I'm hoping that DeepMind will release models showing me what I've been doing wrong! =]
(note: I produce novel models in TensorFlow for research but they're mostly fully differentiable end-to-end backpropagation tasks - I might have just missed how to apply it efficiently to these other domains)
I am curious how well TensorFlow fits for many of DeepMind's tasks though. Much of their recent work has been in reinforcement algorithms and hard stochastic decision tasks (think gradient approximation via Monte Carlo simulations rather than exactly computed gradients) which TensorFlow hasn't traditionally been used for.
Has anyone seen TensorFlow efficiently used for such tasks? I'm hoping that DeepMind will release models showing me what I've been doing wrong! =]
(note: I produce novel models in TensorFlow for research but they're mostly fully differentiable end-to-end backpropagation tasks - I might have just missed how to apply it efficiently to these other domains)