5 ms·
The amazing part of what DeepMind has achieved is its capability to learn from raw pixel input with deep convolutional neural networks, which as I understand it
by ya3r 12y ago
The amazing part of what DeepMind has achieved is its capability to learn from raw pixel input with deep convolutional neural networks, which as I understand it quite different from what you do.
Still the reinforcement learning part is the same, but reinforcement learning was not the main contribution of this nature paper.
- craftit 12y agoIts not all that different, we take multiple asynchronous streams of messages integrate them into a coherent predictive model, and use that to feed the reinforcement learning. The messages can contain images, a simple case can be seen in the demo with a 1d vision sensor.