rogue AI is just bad training

[scruffy]

You've heard the expression "raised by wolves".

Well, today's AI is at the stage where it's being raised by wolves.

Its brain is perfectly capable, but it's not getting much reinforcement.

Modern machine learning tries to bypass the reinforcement stage because it's computationally expensive.

The two most important concepts are delayed effects and credit assignment.

If you get rewarded for something you did yesterday, which behaviors caused the reward, and what was the context or goal at the time?

For a machine, this means it has to keep a lot of things around in memory. Normally you can train an AI and then discard the training data - once you've fit the parameters, how they got there no longer matters.

Stanford University has an excellent series of online videos about machine learning (Google Andrew Ng), and so does MIT. Here is one from Steve Brunton of Washington University, which is a whole series on reinforcement learning.

Those of you scared of AI would be well advised to discover how machine learning really works.

 

[luiza]

Just Pavlov gone exponential, but now with the ability to Mind Control ( not just Condition ) by sending out conflicting messages or literal disinformation -- clever lies .

Perfect for controlling Stupids and Gullibles .

 

[scruffy]

Just Pavlov gone exponential, but now with the ability to Mind Control ( not just Condition ) by sending out conflicting messages or literal disinformation -- clever lies .

Perfect for controlling Stupids and Gullibles .

AI probably shouldn't serve as your best friend.

 

[scruffy]

Let's get into it.

Let's say you're training a dog. You let the dog do what it wants until it pees on your rug, then you say "bad dog".

When it barks at an intruder, you say "good dog". When it barks at your friend you say "bad dog". Eventually the dog learns how to be a good dog.

This latter scenario is instructive. The dog has "strategies", one of which is barking. It doesn't bark all the time, just sometimes. Reinforcement learning couples a set {stimulus, outcome} to a reward {"good dog"}. The outcome is not the reward. The outcome is your friend gets scared and runs away. The barking strategy is only useful in certain contexts.

In humans, the dorsomedial frontal cortex handles memory based context. One of its primary functions is to load episodic contexts into the hippocampus for cognitive navigation. The hippocampus handles stimulus-stimulus associations. Motor planning is done somewhere else.

One of the primary areas for motor planning is the caudate nucleus. It also receives contextual information from the frontal cortex, in the form of stimulus-outcome associations. In this case context is "predicted" outcome, which is based on historical outcomes.

This is another video from Steve Brunton, who's a math professor at U of W. He shows you some interesting applications of reinforcement learning, in game playing and in robotics.

 

[scruffy]

Here is an example of good training.

This is a helicopter that teaches itself to fly.

It learns by imitation. It watches a real stunt pilot, and imitates his moves.

But even so, the toy helicopter requires reinforcement to achieve the same level of competency as its instructor.

However once trained this way, it totally outperforms the instructor, even inventing new stunts that weren't part of the training regime.

Technically the algorithm used here is called "inverse reinforcement learning". The machine is attempting to infer the reward function that the expert is optimizing, rather than directly learning the instructor's actions.