During this lockdown, I was engrossed in two amazing history books — Sapiens by Yuval Noah Harari and The Alphabet vs the Goddess by Leonard Shlain. Collectively, the content of both these books gave me a fair picture of human, societal, economical and most importantly neural evolution. However, this blog is not about summary of these books but rather the comparison between cognitive revolution of humans and AI.
Before cognitive revolution, Homo Sapiens were the most disadvantage predator species owing to their low muscular power and bipedalism( resulting in slow speed). Thus human cognitive revolution began as a result of physical disadvantage of human body against contemporary predators. During cognitive revolution, the brain size of humans increased contributing to 25% of total body energy consumption( from 8% earlier). But the cognitive revolution wasn’t completely identical in both the sexes. Bigger brains meant longer childhoods and slower growth resulting in greater nurturing. Thus human females being mothers of marsupial mammals (infants are weak after birth) had to refrain from hunting during nurturing period and depended on males for meat. Males on the other hand mastered hunting. But because hunting wasn’t always successful, gathering edibles was a vital source of nutrition for the females. Thus males became hunters and females became gatherers.
But what is it about being a hunter or a gatherer? Hunting demands cold-bloodedness, being cruel, focused and time-conscious while gathering/nurturing requires being holistic, aware and empathetic. The different roles played by men and women lead to redesigning of our nervous system by mother nature. It is because of this that unlike other vertebrates, we humans have two functionally different lobes of brain ( called as hemispheric lateralization).
Above image encapsulates the different functions of left and right human brain. To make the distinction simple, let us put this according to Charles Sanders Peirce’s three categories of relation — symbols, indices and icons. According to Peirce’s definition, icons are direct resemblance of the object. E.g. an image of elephant to refer a real elephant. Indices leverage some temporal or physical connection to things they refer to. For example, mercury in thermometer refers to the temperature of the specimen — in this case mercury is the index. Symbols, on the other hand, are the agreed conventions which are used to refer to some object. For example, the word “elephant” has no relationship with the real elephant, but we humans have mutually agreed that these seven alphabets together would refer to a real elephant.
In his book mentioned above, Leonard Shlain theorise that whenever(or wherever) symbols emerged as a prominent medium of record keeping or communication, males dominated the society/kingdom and vice versa. Thus in ancient Egypt where the prominent script was based on cuneiforms( icons), egalitarianism emerged while the opposite is true about rest of societies in later ages where alphabets gave way to patriarchy. In the modern, with the emergence of iconic media like cameras, televisions and Instagram, we see the egalitarian shift emerging again. The reason for this, author argues, is that the symbols are nothing but a mere abstraction of real objects making them easier to deal with by left brain functionalities like abstract thinking, analytical and logical thought process and ordered sequencing which are most important to a good hunter(male). On the other hand icons requires key faculties of right brain like creative thinking, intuitive thought process and spatial intelligence which are most required by a good gatherer/nurturer. The dichotomy between left and right hemispheres mirrors the differences between hunter/killer and gatherer/nurturer strategies.
Like brains, even our eyes have evolved to give two complementary functionalities. The eyes divide every scene in two major elements — figure and ground. Figure is visualised sharply and in detail while ground provides the context within which the figure resides. The cones cells located densely at central part of retina best see the figure while rods cells located evenly throughout the periphery of the retina see the ground best.Thus rods allow the all-at-once computation of brain (like right brain) while cones allow to scrutinise (concentrate) at one object and process one-at-a-time (like left brain). Interestingly women have more rods in their retina while men have more cones. Thus the evolutionary changes in both these organs (and also in some other body parts) allowed humans to become most notorious predators of the planet and build the sophisticated civilisation.
Now let us move to the cognitive revolution in machines i.e. emergence of AI. AI practice is broadly divided into two parts — Connectionist AI and Symbolic AI. While symbolic AI consists of rule based, classic and good old fashioned AI (GOFAI), connectionism consists of algorithms that try to emulate human brain — the Neural Networks. In the earlier years of AI research, symbolic AI algorithms dominated. Symbolic AI primarily emerged through logic based programs of if-else followed by present day machine learning algorithms of Decision tree, SVMs, Naives Bayes, etc. As for connectionism, Alan Turing’s anticipation of connectionism in 1948 went unnoticed, the experiment by Dr Hubel and Dr Wiesel gained some traction and it led to the founding stones of concept behind CNN. This experiment was carried out by inserting micro-electrodes inside the visual cortex of anaesthetised cat. Images of lines were shown at different angles and behaviour of neurons triggered was noted. It was observed that some neurons fired very rapidly by watching the lines at specific angles, while other neurons responded best to the lines at different angles. The fate of connectionism finally changed with advent of high end computer graphics and invention of LSTM (that solved vanishing gradient problem) in 1997. But what is all this has to do about cognitive revolution?
If you compare symbolism versus connectionism with hemispheric lateralization of human brain, it all boils to the same one-at-a-time versus all-at-once computation. The difference is just that, in humans the left brain faculties came into existence much later than that of right brain. It was when the humans became time conscious, invented sophisticated languages and scripts to converse with each other (which required logical and analytical aptitude — all left brain faculties). In machines however, left brain faculties of logical reasoning have been in place since the symbolic wave of AI. However, abilities of right brain — to compose awesome music tracks, to make intuitive decisions or to produce a highly intuitive and simple design are still missing in AI (GANs have many limitations). But what has connectionist AI produced then? A set cones and rods connected to a weak brain I would say.
Rod cells (plus the part of right brain to which they are connected) are analogous to CNN architecture in AI which provides all-at-once computation and provide the ground for image recognition. On the other hand, the combination of cone cells and time conscious left brain is analogous to the RNN architecture which has memory states to store the events in ordered timestamp and thus handles symbolic activities like languages(NLP) just like left brain.
I called the neural networks as weak brains because they don’t have as powerful relational inductive bias as human brain has which solves much of the computational power of human brain. It is due to these biases that we always keep some assumption minds while understanding some concepts which in turn help us to reduce time by just validating the biases. (Biases are also the ones that cause quarrels among us!). Neither does the neural networks have something called as intuition to save their computational time. It is well known that best of the human beings rely on intuitions to make their choices in life( e.g. Steve Jobs). After all, even DARPA admits that third AI wave is all about intuitive machines.
Moreover, most of the human activities require combined efforts of left and right brain. This, in AI, has only been limited to image captioning. and handful of other applications. We humans (or most of the AI practitioners) have a bias that CNN is only for images and RNN is only for time series and text. Rather I most certainly feel, that CNN+RNN architectures would undoubtedly be the part the AGI — Artificial General Intelligence!
However, the results of GPT-3 are promising and hence it has been regarded as the most sophisticated symbolic AI algorithm of date. Also, the connectionist approach may only be the future of AI. The paper by Deepmind emphasis on instilling Symbolic behaviour in AI.
In next part I will be writing about how conversations have helped human to grow, how it can help machines and what role reinforcement learning can play to build the AGI.
Your views on this are welcome in comments, on Twitter or on LinkedIn!
