Understanding Artificial Intelligence & Machine Learning: The Network
As I mentioned in my previous blog, I’ll be covering the basics of AI and ML in three separate posts. Artificial intelligence refers to the concept of machines being able to execute tasks that we would consider “smart” or “intelligent” such as visual perception or speech recognition. Machine learning refers to an application of artificial intelligence that revolves around the idea of developing of computer programs that can access data and use it learn for themselves. Today’s part will cover the neural network, also referred to as universal function approximation.
Artificial intelligence aims to utilise a network of artificial neurons to simulate what biological neurons do. The human brain is made of up hundreds of millions of neurons that communicate with each other to process and transmit information. In the AI world, this is what the neural network looks like:
So how does the above graphic resemble a biological neuron? For your reference, here’s what a biological neuron looks like:
For those who aren’t familiar with biology, below are some key elements to the neuron.
Axon: Also known as nerve-fibres, they are the primary transmission lines of the nervous system. They transmit information from one neuron to another.
Dendrite: The receivers of signals from other neurons and transmits the signals toward the cell body.
Soma: The cell body and effectively “the brain” of the neuron.
Synapse: The synapse is the structure that allows nerve impulses to be transmitted and received.
The fundamentals behind AI is that you are trying to artificially reconstruct a neuronal network by emulating a biological one. At the input layer, you’re introducing new information to the network. Like the brain, all inputs will have different weightings. Let’s use the example of attempting to cross a road in face of an incoming vehicle. Why is that sometimes you’ll run across the road despite the light being green or even where there is an absence of traffic lights, and other times, you will patiently wait?
If I see a car coming towards me as I attempt to cross the road, I will likely consider a range of factors, including the speed at which the vehicle is going at, how far it is from me, characteristics of the driver behind the wheel, model and type of the vehicle, whether I’m in a rush or not, and so forth. These all represent different inputs and consequently, are weighted differently, i.e if the car is going at 100km/hr, I most certainly won’t cross the road no matter how urgent my desire to cross the road is, for fear of getting hit, versus say 40km/hr where I think the risk is at a more appropriate level. All these factors are processed at the hidden layer, where the calculations take place — if the stimulus intensity is above the activation threshold, then the neuron will spike and send a signal. A neuron doesn’t just turn on and off. They will spike when they receive packets of new information.
If I’ve crossed the road about 100 times when the on-coming car has been travelling at 60km/hr, then next time I cross, I might brave it and cross when I perceive the car to be travelling even faster, say at 70km/hr. Taking this example further, if I make a habit of crossing roads with oncoming cars at increasing speeds, but never get hit, it will get to a point where I will cross the road no matter how fast a car is going. You often hear of surgeons getting desensitised to blood and the reason behind that is they have seen it so frequently, to the point that it no longer phases them. The sensory adaptation is as a result of numerous reason, most commonly where the synapse (the impulse receivers) become less responsive to constant stimulus. This will lead to a reduction in spikes associated with the stimulus, eventually falling to zero.
An important thing to note is that the activation threshold is not a logic gate that gives signals or not, rather it is a “continuous gate” of sorts. Once a neuron spikes, it will, through a set of weighted decision trees, provide an answer at the output layer.
But the reason I can make such decisions, is that in my lifetime, I’ve seen quite a few cars driving around and crossed many roads. This is fundamentally why parents will hold their child’s hand while crossing the road. They are aware the child does not yet understand the implications of crossing a road in the face of an oncoming vehicle. The child witnesses this on countless occasions and eventually learns to be able to make their own decision on whether it is safe to cross the road or not by adjusting the weighting of each factor that goes into computing the decision. The parents experiences are effectively used as training data for the child, who will use this to make their own assessments in the future. This represents the “learning” element, which I will cover in the next post.
As always, special thanks to my AI mentor Nick White for reviewing and providing feedback on this piece!
