Pick a card. And Remember it. We will need it in a bit. Got the card ? (if you are impatient, you can scroll down as see how I have removed the card you picked from the deck) Okay, lets take a dive into neuroscience and see how we, humans, are taking decisions and see how it fits into Artificial Intelligence. We make decisions every day, every second and are in fact very good at it (broadly speaking). We really, really like to believe that we are the masters of our own decisions. But are we? Lets see what internal and external factors secretly guide our decisions and perceptions.

We believe we are taking decisions rationally; but are we? (Forget about the times we are emotional) When you are at the supermarket, say you see some thing and you decide to buy it. What if I were to tell you that your brain took the decision approximately 8 seconds before you even started to take the decision consciously? You think you took the decision just now, but your brain took it 8 seconds ago and is just letting you know now. Its true. And its extremely offsetting and down right scary. (Find the Brun Knutson’s study on this, if you are interested)

Our decisions are almost all the time biased based on internal forces such as memory and experience and external forces such as subliminal cues rather than rationality. When we think about vision, the aspects like brightness, contrast and movement are triggers which quickly grabs our attention and affects how we perceive things. Why do you think we suddenly start to like things when they are re-branded? We are so easily distracted and our brains quickly gravitate towards what grabs our attention the most. There are many things that our brains are good at but one aspect we are poor in is the attention. ‘Change Blindness’ is one good example of this limitation in our brains. (The change blindness is “the surprising difficulty observers have in noticing large changes to visual scenes”)

The point is that; to make informed decisions we need to have good information and our brains are very easy to be tricked. At the beginning I asked you to pick a card and now I am going to remove that card. See if I got it correctly.

I did remove the correct card, didn’t I ? Is it magic? Of cause not. Almost everyone feels like that at first and then during the second time we realize how silly it is. Now the point is; this happens to us all the time and it can and is being used by everyone who knows the limitations (or shall we say the ‘loop holes’) in our decision making process.

With our mushroom colored 1.5 kg brain, there is a limited amount of information that one can process at any given time. So what the brain has evolved to do is to filter out everything and focus on what is important. Imagine what would happen if our brain had to process everything that our vision or auditory system is taking in. We would go crazy. So we have developed the mechanism which we call, attention. We pay attention to what we want and blur out the rest. This is exactly what happened with the card. You focused only on the card you wanted to remember.

Now this limitation (more of a safe guard) gets manipulated big time by magicians and also is used for marketing. Input of too much information can distract our brains from what truly matters and input of too little information can deceive our brains into taking bad and/or rash decisions. Its a very thin line we have to walk on.

Now coming to Artificial Intelligence; more specifically Reinforcement learning; an agent need to observe the environment and in many cases it cannot know everything about its surrounding. This is what we call the ‘partially observable environment’. Is this a good thing or a bad thing? Well, same as humans, a large influx of information could potentially distract the agent and an insufficient amount would cause it to never learn. For example if we want to build an agent which can navigate inside a room; giving it information about the soil moisture outside the building would only make it bad; or at the very least would make it take far more time to learn (as it has to learn that the moisture value is not important at all). At the same time if we were to remove all the sensor inputs expect for a light sensor, that would be really bad as well; as that is not in anyway sufficient enough information for it to learn how to navigate.

In Reinforcement learning it is very important how we represent the internal state of an agent. Just like in our brain when we filter out the noise and focus on what we believe to be important; design of an agent should also filter out the noise and only allow potentially useful information. Of cause we cannot do this all the time.

It is not as easy as it sounds to figure out what information to consider and what not to. Take a look at the below scenario (Image Credit : David Silver’s RL lecture). What do you think would happen in the 3rd sequence? Will you get the cheese or get shocked ?

If you were to consider the last 3 events as the information for your decision; then you would come to the conclusion that “well, last three items are similar to that of the 1st one, so would be the shock”

But one could also consider all 4 and say “hey, 2nd and 3rd scenarios both have one bell, one bulb and two levers, so its going to be cheese”. This shows that the amount of information that we are willing to process directly impact the decisions we or an AI agent would make ultimately.

For us the invisible controllers are the things such as memory, experience and external subliminal cues. We can similarly map it to an AI agent where its invisible controllers would be its internal representation of the outside world, past experiences and also the how observable is the environment. The decisions are not always rational, most of the time they seem so, but there are invisible controllers that are forcing the decision maker become biased. These biases could work as a positive or a negative for an AI just the same way they impact us.