Can Artificial Intelligence Replicate the Human Brain?
Working as a machine learning engineer in biotech, I am often asked how the algorithms I build compare to the human brain. Can artificial intelligence really replicate the functions of the mind? There isn’t an easy answer to this question as there are many different types and applications of AI, and the field is continually evolving. It’s a great question though, so here are my two cents on the subject.
Currently, artificial neural networks are the most popular and fastest growing branch of artificial intelligence. These networks are made up of many simulated neurons that process and propagate information through connections to subsequent neurons, eventually converging to an output. Using an algorithm called back-propagation, these networks can be trained from data to classify images, process text and speech, play games, navigate complex environments, and accomplish a host of other tasks.
One function of neural networks that appears to closely mimic how the mind works is the analysis and classification of images. An example, well-illustrated by Ben Dickson, is to consider the picture below. If you recognize this animal, a bunch of neuron activations in your brain have linked what you saw to a name and any other information you know relating to it. However, if you’re like me, you’re rifling through your repertoire of animal species, comparing tails, ears, and fur coats in search of an appropriate classification bin. Your biological neural network is reprocessing your past experiences to inform a novel situation.
Well, the picture is of a large Indian civet, a rare mammal native to Southern Asia. Did you get it right? It doesn’t belong to the rodent, feline, or canine families as one might guess, in fact it falls into a category of its very own called the viverrid family. And now you have a new and appropriation bin for the next time you come across a civet.
Many artificial neural networks work a lot like this, pattern matching based on past experiences to interpret new inputs. Artificial neural networks are really quite good at this type of classification and clustering of data, in fact they can crush human opponents in strategy games such as chess and Go; however, they fall short when it comes deduction and reasoning. Many of the simplest behaviours for even primitive animals turn out to be deceptively challenging for artificial neural networks to replicate. For example, we cannot train an artificial neural network to stalk prey in the wilderness or build a nest from available twigs and mud. Furthermore, children who are just beginning to explore the world don’t rely on structured supervised learning, instead they learn a vast majority by investigating and discovering their surroundings.
One explanation for this might be the vast difference in size. Our brain consists of about 86 billion neurons and over 100 trillion synapses. The number of neurons in an artificial neural network is much less than this, usually in the ballpark of 100 to 10000 tops. This inevitably puts the artificial neural networks at a disadvantage computationally, but what’s interesting is that increasing the number of neurons in an artificial neural network does not always yield better results. This highlights two other key differences between artificial and biological neural networks: architecture and power consumption. In artificial neural networks, only neighbouring layers are connected which, in most cases, are activated sequentially. Biological neurons on the other hand often fire asynchronously in parallel, coordinated in part by their myelin coating and diversity of neuronal cells. Furthermore, the human brain runs on less than 20 watts of power, while a computer will usually consume at least 200 watts and generates a lot of heat in the process.
So, while artificial neural networks are far from being able to replicate the full function of the brain, they do have value in their own right. They can help people solve complex problems in real life such as optimizing logistics for transportation networks and processing raw photos and videos in medical imaging, robotics, or facial recognition.
Excitingly (and somewhat surprisingly), artificial neural networks may also prove useful for us to better understand the brain. Studying how a trained model generates its outputs and changes in response to new stimuli may well provide useful insights into their biological counterparts.
Thank you to the Cansbridge fellowship, SmartARM, and Queen’s University for their support. Thanks very much to you for reading and stay tuned for my next blog post!
Additional Resources
https://bdtechtalks.com/2020/06/22/direct-fit-artificial-neural-networks/
