3 ms·
With regard to Chalmers emphasis on simulation: I think simulation is really a central point in AI and future human development. Not only that we crave the web
by visarga 9y ago
With regard to Chalmers emphasis on simulation: I think simulation is really a central point in AI and future human development. Not only that we crave the web and the VR, but AIs also need VR for virtual training. It is sometimes costly and other times impossible to encounter certain specific situations in real life, but agents need to be prepared. Human pilots train in simulators as well - for the same reason.
So simulation for AI agents is like school, or a lab where they can experiment. Simulation, in my opinion, is what can turn simple DL systems into agents capable of deep reasoning and creativity. AI will learn mostly from analysing and discriminating in the results of multiple simulation runs.
Gaming, VR and and DL run on the same kind of hardware - GPUs. In fact DL has benefited from the availability of gaming GPUs in the last five years. The same hardware, the same emphasis on simulation for both fields. Other simulation examples: AlphaGo did MCTS on top of neural nets (a simulation of Go playing), and self driving cars reconstruct a virtual scene of the world, where they do path planning and risk avoidance.
In my opinion, reasoning is also a simulation task - once we map into a graph all the relevant entities and the relations they hold between them, it is a matter of doing signal processing on graph (kind of like electronic circuits) to determine the answer for a number of reasoning tasks. So graph based, object relation representations could be used as cognitive simulators. DeepMind has a paper recently on graph based NNs.
Simulation might look different from games to weather, proteins and cognitive reasoning, but in all these cases we input a complex problem state and run the simulation to observe the outcome. We can do research and learning by this process.