AI ?
AI is all the rage these days, Only shadowed by the crypto hype. In this post I would like to try and distinguish the hype from the change, A big change is happening around us and with lack of better words we often use the A word to express it. Artificial intelligence, The word intelligence should be used with more caution as we are far from understanding its meaning.
Since my childhood I saw people around me treating the processor as the computer brain (It might even be the reason I made a career out of deep diving processors) and as time went by I saw this comparison fade away, Its coming big time these days as a special magic dust enters our technology all around. You see, We really understand how processor work, Actually at Intel we had amazing teams spending years being familiar with every corner of its logic , neural network processors are about to change that and while the underlying machine mechanics are explainable, the results interpretations wont be as easy.
A small neural networks promo
Neural networks are in the ground zero of the revolution, What are neural network ? I will brief it shortly:
Neural network is a network of connected artificial neurons. A neuron is a point which is connected to other neurons and decide whether the neuron should transfer the signal (on/off signal, just like your light switch) . Each neuron learns when it should “fire” (turn on its output) given on/off combinations on its inputs. This simple principle is behind the entire AI buzz, We tend to call it AI as its kind of resembles on how our brain works (Biological neurons have more complexity).
These networks are with us for decades, so what has changed ?
Around 2010, Processors became strong enough to actually be able to run very large networks. In addition the big data era of 2006 generated enough datasets to feed these networks. So strong enough compute power to run these networks and enough data to feed them have opened a new window. Time to get some perspective on network intelligence capacity.
The human brain as processor.
Neural network strength usually measured by the number of its neurons and connections (synapses in the bio world), this is a gross factor to the amount of pattern recognition the network is capable of, lets get into perspective with some numbers, take these with grain of salt:
- Number of neurons in the brain ~ 100 Billion
- Number of Neurons on very large networks ~ 100 Millions
We get 1000X factor of the human brain over our largest networks, I saw the same math goes to 10000X as well (many assumptions and hand waving), but you get the point, Our best artificial networks have 0.1% of the human brain capacity.
Assuming Moore’s law to continue (2 years for 2X compute power, wild assumption but that’s for a different post), We can expect AI to reach a single human brain capability within ~ 20 years.
So why is all the excitement ? Apparently 0.1% is a lot, and given the fact its keeps growing then we can expect acceleration of AI capabilities all around us.
So what can you do with 0.1% brain ?
Apparently a lot, In dataloop we learn about new business transformations on a daily basis, from self driving cars and robotics to automatic medical analysis and personal assistance, the emerging opportunities are all over and likely more applications will become possible as neural network processors are getting stronger. The rule is simple, If a human expert can detect a pattern (A damaged railway or bad product on assembly line for example), Machine will likely do it better, Its only a matter of (short) time.
The challenges ahead ?
There are many ways to divide/map cognitive capabilities , I like these 3 layers as they keep thing simple:
- Recognition — When we recognize patterns like speech, visual perception and others.
- Knowledge representation — When we connect past recognition to present recognition and getting stronger understanding of the present, and store current data for future use.
- Reasoning — When we conclude from the present & past to the future.
Our today’s 0.1% systems are partially applying the recognition (its getting better) , Experimenting with knowledge representation and have very limited reasoning.
In the above picture AI detects (tested with Google vision API) the children and the birds as drawn (actually pretty hard but doable if you are Google) , You are able to understand pretty fast they are feeding the parrots, though you have never seen this picture, your knowledge representation allows you to conclude that, The feeding conclusion is where AI limits today on very basic images.
What about the rest of the scene ? Feel free to guess, your guess is probably better then all AI out there.
But the really big deal in my mind are emotions and the intelligence behind them and we are going for short emotional ride, one that is decades away from AI capabilities.
Child and a bird, take #2
Again a child and a bird, But a photo that sends shivers down one’s spine. You can probably feel your emotions rise, Yes you have seen the bird and the child like before but you probably have other millions words running in your head: The bird is a vulture feeding of dead animals and you reason what it’s waiting for (the future this image represent).
But much more happens in your brain, What a world are we living in ? How can we save her (notice you couldn't tell its her,so don’t expect AI to ever be able) ? What happen to her ? Obviously the photographer was there so it cant be that bad ? Some maybe thinking on God and others maybe on their own children.
The vulture and the little girl, also known as “Struggling Girl”, is a photograph by Kevin Carter.
Kevin won a Pulitzer prize for this photo and committed a suicide four months after winning the prize, at the age of 33. Did adding this knowledge changed the way you perceive this photo ?
He chased away the bird and reported the girl went back on her feet, No one knows what happen after.
What AI has to say about this photo ?
Today’s AI and Ill take a wild shot, 2025 AI as well wouldn't be able to describe the scene with 15 words(How many words do you have?).
I tested two visual image analyzers out there, Google’s Vision API and IBM Watson (both are considered state of the art), These are the results:
IBM watson :
Google Vision:
As you can see, both failed to recognize the most “simple” objects in the image (The most basic level of intelligence ) , This image besides being emotionally hard is also technically hard, even on the basic level.
How much time until a computer will be able to “feel” what we do about this picture ? Does it ever needs to or our emotions are nothing but a distraction or weakness?
Summary
Automation and pattern recognition are probably better words then AI but they don’t buzz as well so we are likely to continue using the A word.
I tried to walk you, the reader, in the same path as AI is walking today. The logic and analytics are “easy”, The reasoning is very hard and emotions and feelings are close to impossible.
So in the coming years when you hear about new state of the art company or algorithm feed it with the struggling girl and judge for yourself if it manages to produce meaningful 15 words.