Why Artificial Intelligence Needs Philosophy
How philosophical insights can guide the field of artificial intelligence
The Cognitive Revolution
The prospect of genuine artificial intelligence has circulated in the popular consciousness since at least the advent of mainframe computers in the 1960s,
but the feverish enthusiasm of the 1970s and 1980s soon petered out when the scientific community came face-to-face with a problem that has preoccupied philosophers for centuries: the problem of consciousness.
That consciousness should be in the radius of scientific inquiry at all had been taboo. This was the result of the dominance of methodological behaviourism in psychology, which rejected the systematic study of mental phenomena because they proved so resistant to experimentation.
This view reigned supreme until around the 1960s, when the limitations of modeling human behaviour on reinforcement and operant conditioning could no longer be ignored.
Philosophers themselves had experienced their own behaviourist phase and articulated positions about consciousness that tried to reconcile the vocabulary of mental phenomena, such as pains and feelings, with materialist science.
Many of these views had a reductionist agenda in the sense that they argued for the identity of mental states such as thoughts and desires with brain states described in neuroscientific language, say c-fibers firing.
Is the Mind a Computer?
Integrated Circuitry. Image by Umberto.
One of the positive outcomes of the cognitive revolution was that it recognized the limitations of behaviourist methodology, namely the sole study of observable human behaviour, and launched a multidisciplinary study of the mind that enlisted the insights of philosophers, psychologists, linguists and neuroscientists known as cognitive science.
What became clear from this multidisciplinary approach was that reductionist theories and models were sorely inadequate in the sense that they tried to deny the reality of inner phenomena instead of explaining them.
Parallel to the cognitive revolution, the field of artificial intelligence was born in Dartmouth College in the 1950s spurred by Alan Turing’s theory of computation and the rise of digital computers in the same decade.
Since digital computers model mathematical computation and symbol manipulation, the idea arose that symbol manipulation was the essence of the mind. This can be expressed by the metaphor that the mind is to the brain as the software is to the hardware in computers.
Enthusiasm for this identification of computation with human thought was so great that Herbert Simon, one of the founders of AI, proclaimed that machines will be able to do what humans can in a matter of decades. This, alas, did not come to pass. What went wrong?
The Consciousness Dilemma
Mind as a nebulous phenomenon. Image by Jr Korpa.
This is where philosophy comes in. One of the aspects that many philosophers recognized that computers could not emulate was the subjective, first-person experience that characterized human consciousness. Some philosophers and scientists find this problem so recalcitrant to explanation that they dubbed it the hard-problem of consciousness.
The philosopher who coined the term, David Chalmers, laid out his views in his book The Conscious Mind: In Search of Fundamental Theory (1996).
In it he argues that empirical science has made little to no strides in explaining how the brain gives rise to inner, subjective experience, in part because first-person awareness cannot be broken down into components like other phenomena, and may be fundamental.
Other philosophers, like Daniel Dennett, disagree.
These philosophers think that much of what we experience as a unified, internal, subjective theatre with the self at the helm is rather an elaborate illusion generated by massive information-processing systems in the brain that take sensory input and yield complex behaviour as output.
The parts that we experience — namely thoughts, desires, beliefs, pains and pleasures — are the tip of the iceberg of an ocean of unconscious processes.
Dennett therefore denies the reality of first person, subject experience, also known as qualia, and instead claims that the stream of awareness is the result of a vast bundle of parallel and almost independent processes that create the illusion of a unified field.
The secret to cracking the code of the mind is not in overcoming a hurdle that somehow qualitatively separates the mind from other phenomena, but rather lies in letting empirical science run its course.
Dennett explained his views in his seminal book Consciousness Explained (1993), which was both lauded for its efforts to naturalize consciousness and criticized for evading the problem of first-person subjective experience altogether.
Philosophers, therefore, fall into two camps: those who think that subjective experience is reducible, and therefore identical to, brain states, and those who think that subjective experience, while causally generated by the brain, cannot be reduced to it or explained away.
The varieties of positions are in reality are much more nuanced than I’m able to get into, but in essence the distinction can be summarized as follows: while nearly all philosophers argue for physical identity, namely there’s only one physical reality,
some deny property identity, namely that some properties, such as being in pain, are identical to physical properties, such as neuronal firings.
Despite philosophical disagreement as to whether experience constitutes something ineffable that we cannot assimilate into materialist science, today philosophers and scientists are almost unanimous in their rejection of the computational theory of the mind.
Minds are not computers; we are, in fact, on the whole bad at math and reasoning, and symbol manipulation captures only part of what the mind does, nor does it do it serially like a computer.
Daniel Kahneman’s acclaimed book Thinking Fast and Slow (2011) popularizes a wealth of psychological evidence indicating that mental processes divide into two parallel systems that sit uneasily alongside each other: one heuristic and domain-specific managed by the evolutionarily older parts of the brain, such as the amygdala and brain stem, and the other slow and domain-general managed by the evolutionarily younger parts of the brain, such as the cerebral cortex.
The sheer variety of cognitive biases covered in the book attests to the limitations of our mental capacities, and how separate systems in the brain evolved to cope with environmental problems by simplifying informational input.
Are Artificial Neural Networks the end of the road?
Connectome: MRI tractography of white matter tracts in the brain. Photo from Wikipedia.
Today all the rage with AI has shifted to the burgeoning field of artificial neural networks. Not serial processing, modelled on computation, but artificial neural networks, implemented in computers, better model the mind.
Unlike serial processing, neural networks model information processing on biological nervous systems consisting of neurons that transmit electro-chemical signals through a network.
The neuronal equivalents in the model are nodes endowed with activation weights. The signal input is a real number, and the output is computed as a non-linear function of the sum of inputs in a layer of nodes.
The signal is propagated if the output meets a threshold, whose value changes with each iteration of input or learning.
Perhaps artificial neural networks are the answer. After all they are the closest model to our understanding of how the brain works.
The only problem is that present understanding of biological neural networks is feeble at best. A biological signal is not a number, nor is it computed through a function.
We do not yet know how it is that connections through billions of neurons give rise to the cognitive systems we have identified such as memory, attention, and learning mechanisms, though we have very good guesses about parallel neural circuitry.
Most elusive of all, subjective experiences like pains, thoughts, and the self remain even less understood.
The question to be answered remains: how does a recurrent network architecture implement the mind?
Until this question is answered, AI has no hope of rivaling human general intelligence. Often called artificial general intelligence (AGI), or strong AI, this type of intelligence emulates all human capabilities, including creativity and reproduction.
Perhaps it will turn out that phenomenal consciousness is not the distinguishing feature of our intelligence, but some causally inert byproduct (epiphenomenon) of our biological hardware.
If that turns out to be the case, our fascination with consciousness will have been little more than an anthropomorphic obsession with no broader significance than quenching our native curiosities.