Emergent Sparsity in Frozen-Random-CNN Deep RL by Scott Norton
Details
Playing Pong Through 3 Neurons: Emergent Sparsity in Frozen-Random-CNN Deep RL
Something unexpected happens when you freeze a randomly initialized CNN and train a deep RL agent on top of it. The agent spontaneously compresses all task-relevant information through as few as 1–3 of the 64 neurons in its first fully-connected layer, even with no sparsity objective anywhere in the training setup.
Scott Norton presents his manuscript (arXiv pending) documenting this phenomenon in Atari Pong, under both deterministic and sticky-action training. Come see how it is possible to do AI research even if you're not at a frontier lab.
No Reinforcement Learning background required. The setting is informal and open to everyone!
Suggestions for future papers and topics welcome.
Find us in the Eagles conference room attached to the main Coworking area.