,
12 tweets,
3 min read
Read on Twitter
DeepMind now researching the connection of Sleep to Deep Learning: deepmind.com/blog/article/r… #ai #deeplearning
Biological brains require sleep, but what exactly does sleep have to do with cognition? medium.com/intuitionmachi…
But why do brains require re-experiencing events to synthesize its understanding of these events? It seems that what humans remember about experiences are the summaries of these experiences played backward. That is how we build our causality networks!
The overlooked fact about supervised learning is that it only learns anti-causal relationships. In other words, the input is the effect and prediction is the cause. It can't learn given a cause, what is the effect. In short, it doesn't formulate causal models of the world.
When you replay experiences and you learn from the world by going backward in time. When we interact with our environment, we use supervised learning. That's because our actions are the cause and we perceive the effect. We learn to predict the cause (our actions) to the effect.
This is why all perception is based on recognizing affordances (see: Gibson). Our perception learns how to recognize what actions are available. So when we see a gradient or a shadow, this gets translated to a set of actions available in a 3D world.
Visual perception shares a common cognitive substrate with touch. Our vision is actively scanning its inputs to make sense of what the eye sees. It constructs affordance structure and not pixels. medium.com/intuitionmachi…
But as an agent learns affordances, it constructs its own biases. We create our own filter bubbles making it more difficult to breaking out and seeing reality as it truly is. We are the product of our environments. We cannot "unsee" visual illusions.
This is also why studying history is important. Many ideas are the way they are not because of rational thought, but rather by coincidences in the flow of time. The bias of GOFAI is a consequence of the emergence of computers and dualist thinking.
To make stronger progress towards AGI, we have to ween ourselves away from cognitive biases that originate from history. We have to re-examine history and recognize the cognitive baggage that we've accumulated.
Computers do not need to sleep for a reason.
@threadreaderapp unroll
