4 ms·
"Analysing the agent’s internal representations, we can say that by taking this approach to reinforcement learning in a vast task space, our agents are aware of
by xcodevn 5y ago
"Analysing the agent’s internal representations, we can say that by taking this approach to reinforcement learning in a vast task space, our agents are aware of the basics of their bodies and the passage of time and that they understand the high-level structure of the games they encounter."
Wow, really amazing if true.
P.S.: After looking into their paper, it's not that impressive. They use agent's internal states (LSTM cells, attention outputs, etc.) to predict whether it is
early in the episode, or whether the agent is holding an object.
- Tarq0n 5y ago"Aware" is probably overly anthropomorphized language there. What they mean to say is that all these things have become parameterized within the model.
- jcims 5y agoIt would be interesting to see what would happen if they added social dynamics between the agents...like some space for theory of mind (what is that agent thinking), mimicry, communication, etc.
- leesec 5y agoFrom the article: "Because the environment is multiplayer, we can examine the progression of agent behaviours while training on held-out social dilemmas, such as in a game of “chicken”. As training progresses, our agents appear to exhibit more cooperative behaviour when playing with a copy of themselves. Given the nature of the environment, it is difficult to pinpoint intentionality — the behaviours we see often appear to be accidental, but still we see them occur consistently."
- phreeza 5y agoThere is also some other work from deepmind in this direction: https://deepmind.com/research/publications/machine-theory-mind https://deepmind.com/research/publications/machine-theory-mi...
- jcelerier 5y agothe main question of course being, aren't we anthropomorphizing ourselves too much ?
- Layke1123 5y agoAsking the real questions that will upset alot of people. ;)
- K0balt 5y agoI think this is a key insight. Human exceptionalism is, in my opinion, an extremely flawed assertion based on a sample size of one, yet it is widely accepted. Actual evidence does not support the idea that awareness of self and other “hallmarks of intelligence “ require anything more advanced than an insect, or perhaps even fungi.
- dougmwne 5y agoWhen people say this kind of stuff, I wonder whether there might not be philosophical zombies among us.
- the8472 5y ago[ ] To prove that you are human please describe how you are observably different from embodied, general, adaptive agents in 200 words.
- modeless 5y ago> it's not that impressive. They use agent's internal states (LSTM cells, attention outputs, etc.) to predict whether it is early in the episode, or whether the agent is holding an object. That seems like a decent definition of awareness to me. The agent has learned to encode information about time and its body in its internal state, which then influences its decisions. How else would you define awareness? Qualia or something?
- woeirua 5y agoBy that definition wouldn't a regular RNN or LSTM also possess awareness?
- modeless 5y agoI think it would be perfectly reasonable to describe any RNN as being "aware" of information that it learned and then used to make a decision. "Possess awareness" seems like loaded language though, evoking consciousness. In that direction I'd just quote Dijkstra: "The question of whether a computer can think is no more interesting than the question of whether a submarine can swim."
- pcl 5y agoOoh, that’s a great quote. I’d say that it’s no less interesting, either.
- imvetri 5y agoHaha. Thanks