Can human brain be the key to true Artificial Intelligence?

AI (Artificial Intelligence) today is far from human level intelligence. Some of the AI applications like Alpha Go, speech recognition, handwriting recognition, biometrics etc. makes a common man believe AI is very close to a conscious humanoid General AI - but this could not be far from reality. Having recently developed a keen interest in neuroscience, I realized something interesting - subconscious mind that is responsible for about 95% of decision making is very much like a narrow AI (AI trained to do only 1 task). Obviously, this is just a personal opinion and I would love to hear if I am missing something critical in my chain of thoughts. The reason I think this is very interesting is that this can be the "key" to the general AI, we read in books. A few callouts before you read - lots of facts in this article are taken from various credible sources (no credits to me), the brain representation is an over-simplification and I am not a neurologist - just a data scientist with a belief that I would see the Artificial General Intelligence (AI that can do everything like humans) in my lifetime.

Why I think narrow AI is like the subconscious side of the brain?

Answering this question would need a bit of understanding of the brain. So if you are a neuroscientist, please skip the next 2 paragraphs. Neuroscience has made strong progress to understand human brain over last 2 decades. On a high level, we know there are two mechanisms through which we think - very elegantly put by Daniel Kahneman as System 1 and System 2.

System 1 is the sub-conscious mind that dominates most of the decision making and is almost out of our immediate control. It is built with a set of beliefs and a number of causal relationships that we have learnt in our lifetime (even if most relationship might just be un-directional associative level). Our body inherently prefers to give most of decision making to this sub-conscious mind to save energy. Why save energy you ask? The brain has a very very high energy consumption density - it consumes 20% of energy while it only weighs 1.5 kgs. So it needs to be in energy saving mode all the time until unless it is absolutely necessary otherwise. Most of sub-conscious activities happen through the network in deeper brain called the limbic brain (the brain we share with animals) and the basal ganglia (the brain we share with reptiles) while conscious activity happens through a network where the Neo - cortex (this part if what make humans unique) is very actively involved. Below picture is a simplistic and intuitive representation of the brain (source - danielbrian.com; based on Triune brain model).

It might be quite counter intuitive to feel that 95% of decisions are made through the sub - conscious mind but here is how to think about this. While I am writing this article, I feel in full control of what I am writing, but all that Neo-cortex doing at this point is instructing my sub-conscious to write what would logically make sense. When brain instructs my fingers to hit keys on the keyboard, sit properly on the chair, keep my eyes blinking, tell me to drink water, get up every now and then - it's all done through the subconscious. A typical example given to demonstrate the power of the subconscious mind is car driving on a familiar road - notice how conscious mind takes control of a situation while driving whenever you see a dog on the road or two trucks overtake you from both sides. The brain does this team work between the two systems quite effortlessly and without us even knowing. Enough about brain - lets now come back to the topic how subconscious mind is similar to the narrow AI. Below are few things we know about the subconscious mind -

1. It is fast & effective for high volume procedural tasks. Driving, using keyboard etc. are all procedural as the rules of the game do not change much.
2. It lacks flexibility. This is an extension of the last point, but worth repeating. Sub-conscious mind fails if you give it a new task that we have never seen before. For instance, car driving is quite sub-conscious for me most of the times, but it is very much conscious for one of my friend, who is still learning to drive a car. By the way she can drive a two wheeler quite well, but sub-conscious is unable to adapt most of those learning and screams conscious side for help.
3. It takes little effort. Going back to driving - it is refreshing for me to take a long drive that can be straight 3-4 hours, but it drains my friend to go on driving lesson for just an hour. Why - is the car different? Is the road different? No and No. It's just a different part of the brain we are using.
4. It is hard to interpret. Even though, subconscious mind takes decisions for the majority of tasks, it is super hard to understand "Why?". If you ask why you are so anchored by the price tag in a paint gallery of abstract art OR if you ask why you chose the medium size of coffee OR why you prefer I-phone when other brands can give you better features for less price OR why a Coke lover prefers Coke over Pepsi - you possibly would not get any concrete answers. The best answer you get is - "It just feels better/right". And the answer is absolutely right, even though hard for our neo-cortex to understand as the answer lacks specifics. Studies have shown that many of our choices are governed by our sub-conscious, which happens to be in limbic brain which controls emotions. Simon Sinek in his book "Start with Why" explains this concept quite well. He says - the limbic part of the brain that controls emotion and drives a lot of decision making is not the same part that controls language. This makes it hard to express these preferences and feelings.
5. It needs lots of data to make changes in the model. I would speak for many when I say I want to adopt a fitter lifestyle - start with getting up early in the morning and go to the Gym. However, everyday I struggle. At the beginning of the year, I took a challenge to change this and force myself to follow a fitter routine. To my surprise, it became almost natural and effortless in a few weeks to get up early. I even started feeling incomplete when I missed the gym in the morning (I wish it lasted). I am not proud of the fact that I slipped back, but it still makes a point. To code into the subconscious, we cannot convince it with reasons "Why getting up early is important" but body needs, behavioural proof, in other words many data points to code this rule in the subconscious. I wish there was a short cut to explain the sub-conscious what I want and align it to what I like.

Do points 1 to 5 looks familiar - lacks flexibility (narrow learning), fast & effective (becomes pretty good for selected tasks), little effort (can work with less layers/neuron so that our computers can handle), hard to interpret (the biggest challenge with Machine Learning), and needs a lot of data to retrain (if AI needs to learn new pattern, it needs a lot of data points). If you still want to make it explicit - these are all you learn in AI 101.

How can this similarity help?

Hopefully, I was able to convince you to some extent that Deep learning and AI are indeed similar to the subconscious mind and not to the overall brain - the difference might look insignificant but it is an important one because this makes us understand what we lack in building the Artificial General Intelligence. I am no way saying I know the answer, but it excited my neocortex enough to write this long article and I am curious to listen to experts what they think of this chain of thought. Let's try to push this chain a bit more on few different scenarios to get a few tangible outcomes using this similarity -

Direction towards Artificial General Intelligence

One of the articles on Exploring your mind (From the Subconscious Mind to the Conscious Mind - Exploring your mind ) talks about how the human brain evolves with age. In short, human children start with a very low usage of neo-cortex and primarily use their subconscious mind. They then progressively develop the system 2 i.e. conscious mind. After the age of 12, the two systems start working independently while still evolving the network towards higher frequency brain waves. Fascinating, isn't it.

When as a child your parents explained you to make a paper plane, did you try to understand "why it flies" or you simply copied the mechanism and it worked. This is what we did for deep learning networks too - we knew brain has a net of neuron passing signals, so we created computer programs that could do exactly that with few mathematical assumptions - many many trials and errors later we have the world's strongest predictive algorithm.

Now, following the logic that today's AI can at best mimic subconscious mind, we can follow the evolution curve of humans. This is exactly what one school of thought on AGI believes. Josh Tenenbaum says this very well - "We're trying to take one of the oldest dreams of AI seriously: that you could build a machine that grows into intelligence the way a human does—that starts like a baby and learns like a child." I am not saying AI is at par with the subconscious mind (far from it) nor I am saying following this path would definitely lead to an Artificial General Intelligence - I am just saying this makes us realize that there are absolutely few missing pieces that we don't know how to build yet. Like, the programmed instinct, which humans are naturally born with (they are definitely not a set of hard coded rules), the tiny bit of neocortex even a 2 year old starts using, and so many more unknown unknown, but training AI on childlike intelligence would be a good start for sure.

Photo Credit - WSJ

Answering some of the hard questions about future of AI

One of the questions, I am often asked is "What jobs are at risk in next 5-10 years with strong advancement of AI?". You can say something like "high volume, low complexity" tasks - but would people who have a very basic understanding of Data and AI understand such an abstract language. Can we instead say - tasks that you can do effortlessly after some knowledge & practice can be automated, which is essentially those tasks done mainly through our subconscious minds, which are in-turn an AI like structure. These would be things like driving, simple queries at a call centre, payroll managers, editing/proof reading, etc. In my mind, the later response covers a lot more cases than the generic "high volume" response. So, you better be using a lot of pre-frontal cortex in your work to be assured that AI is not really coming to take your job.

Source - Builtin.com

Unleash the power of augmented intelligence

At the beginning of this article, I spoke about how effortless it is for brain to bounce between System 1 and System 2. If we are unable to build system 2, why not borrow it from humans to support system 1 at scale i.e. leverage human enabled bots and bot enabled humans. We already use this in many places like when you chat with Amazon customer service, you start with the bot but you have an option to switch to human - but is this transition seamless. Not really, we absolutely know when we switch from humans to a machine. I believe there is a lot of opportunity to create systems where AI takes calls most of the times, but quickly hands over to humans if it encounters a question it is unable to answer without consumer realizing the switch happened.

Companies like Tesla are already able to make this happen, where the car is on auto pilot mode, but human can take over the control at any point in time. Additionally, every time human touches the steering the car notes this as an error in its prediction to learn. This is what I think the near future state of the world of AI should be - "harness the pre-frontal cortex of human" - sounds insanely mean I know!

Source - Analytics Insights

Is artificial Neocortex a possibility

Judea Pearl in his book "The book of Why" articulates very well on how humans rose in the food pyramid by asking a simple question - "Why". He illustrates the model of thinking through the ladder of causation - finding association (seeing) , Intervention (doing) and Counterfactuals (understanding). He argues, and I agree, AI alone can only find an association relationship - AI would never know if the rise in barometer reading causes a storm or other way round just based on data. Humans on the other hand can do counterfactuals almost effortlessly through their conscious mind. Counterfactual questions can be posed as "What if I had done" or "Why" - both have the same essence comparing the observed world to a counterfactual world.

For example, consider statement - "my flight is delayed because of a storm" - here I can only make this causal statement if I know in a counterfactual world, the flight would be on time if there was no storm. Notice that counterfactual data cannot exist as this is an imaginary world - so AI cannot learn these relationships from data exclusively. Human do this thought experiment all the time using their neo-cortex to learn true causal relationship and then imprint these learning to the sub-conscious brain.

So what, you ask? There are at least two learnings here - 1. if we want to impart such pre-frontal cortex skills to AI - we need to create an AI that could imagine a counterfactual world. 2. if we simply want to give AI (that can behave like a subconscious mind) all the causal relationships we know of, we need a new vocabulary/mode of input, i.e. it cannot be learned through data nor it can be imparted through rules (as they are too inflexible).

So what do you think?

Writing this article was a novel experience as I am not a neuroscientist and can only understand the subject superficially. The core intent to write it was to get your views on what you think? Do you agree with me that AI is indeed like subconscious mind? While reading the article, were you able to stretch the thought more than I did or in a different direction than I took? I would absolutely love to hear your thoughts on the subject.