Monday, August 26, 2019

Dennett - why we need polymaths in the AI ethics debate

Daniel Dennett, is a renowned philosopher, but also a polymath, with a deep understanding of psychology, science and maths. His book Bacteria to Bach and Back is essential reading for those who want to see a synthesis of human evolution and AI, and its consequences for the ethics of AI. Dennett has the backgrounded depth in philosophy to understand the exaggerations and limits of AI and its consequences for ethics. He also grounds his views in biology an humans as the benchmark and progenitor of AI.

Competence without comprehension
He takes a holistic view of AI. Just as the Darwinian evolution of life over nearly 4 billion years has been ‘competence without comprehension’ the result of blind design, what Dawkins called the ‘blind watchmaker’, so cultural evolution and AI is often competence without comprehension. We have all sorts of memes in our brains but it is not clear that we know why they are there. Similarly with AI, Watson may have won Jeopardy! But it didn’t know it had won. This basic concept of competence without comprehension is something that has to be understood and assumes as there is far too much exaggeration and anthropomorphism around in the subject, especially in the ethics debate. His view, which I agree with, is that AI is not as good as you think it is and not as bad as you fear.

Bayesian brain
His vision, which has gained some traction in cognitive science, is that the brain uses Bayesian hierarchical coding (Hinton 2007; Clark 2013; Hohwy 2013), a prediction machine, constantly modelling forward. Interestingly, he sees this as the cause of dreams and hallucinations – random and arbitrary attempts at Bayesian prediction. This is an interesting species of the computational model of the brain and explains why the brain has been a productive, intuitive source of inspiration for AI, especially neural networks. 

Cultural evolution
He then examines cultural evolution as the invasion or infection of the brain by memes, primarily words, and that these memes operate in a sort of Bayesian marketplace, without a single soul or executive function. These informational memes, like Darwinian evolution, also show competence without comprehension and fitness in the sense of being advantageous to themselves. That brings us back to the ethical considerations around AI. He surfaces the contribution of Baldwin as an evolutionary theorist who saw 'learning' as an evolutionary accelerator.

AI
As he rightly says, we make children without actually understanding the entirety of the process, so it is with generative technology. Almost all AI is parasitic on human achievements, corpuses of text, images, music, maths and so on. He is rightly sceptical about Strong AI, master algorithms and super-intelligent agents.

We already trust systems that are way more competent than we humans and so we should. His call is for us to keep an eye on the boundaries between mind and machine, as we have a tendency to over-estimate the 'comprehension' of the machines, way beyond their actual competence, and investing in or anthropomorphising comprehension. We see this with even the most casual encounters with chatbots and devices such as Alexa or Google Home. We all too readily read intentionality, comprehension, even consciousness into technology when it is completely absent.

AI ethics
By adopting regulatory rules around false claims of anthropomorphism, especially in advertising and marketing, we can steer ourselves through the ethical concerns around AI. Over reach and concealing anthropomorphism and false claims should be illegal, just as exaggeration and side effects are regulated in the pharmaceutical industry. Tests, such as variations of Turing’s test, can be used to test their upper limit and actual competence. He is no fan of the demand for full transparency, which, he thinks, and I agree, are utopian. Many use Google Scholar, Google and other tools without knowing how they work. Competence without comprehension is not unusual.

Learning
His hope is that machines will open up “notorious pedagogical bottlenecks” even “imagination prostheses” working with and for us to solve big problems. We must recognise that the future is only partly, yet largely, in our control. Let our artificial intelligences depend on us even ”as we become more warily dependent on them”.

Tuesday, August 20, 2019

Neurotech – mind-boggling final frontier between mind & machine – and possible impact on learning

A slew of companies are working on neural interfaces in what could (emphasis here) change the future of our species through a deeper understanding of how we learn and the ability to accelerate learning.
The brakes on learning are well known, what Dennett (2017) calls the 'notorious pedagogic bottlenecks'; working memory, forgetting, inattention, distraction, interference, crude interfaces, low bandwidth interface of our meat fingers, inability to upload, download and network. Learning is often like trying to squeeze and elephant through a porthole. Our minds can only deal with a tiny fraction of the available sensory and other information that is available. So consciousness, and therefore learning, is severely limited by the evolved apparatus of our current, organic minds. Neural interfaces may free us from some of these holdups and blockages.
There is no shortage of brainpower, companies and investment behind the push. The stakes are high and the people working in this field are well-funded, multidisciplinary (neurologists, engineers, computer scientists, AI experts, mathematicians). They want to radically improve the interface, some non-invasive, some invasive, to improve cognition and performance.
This is a sort of cognitive moonshot, where frictionless movement between mind and machine could be possible or at least sophisticated hacks that make our minds more potent and efficient. Isn’t it curious, even wonderful, that the organ itself is now pushing for its enhancement?
On hearing about this stuff, many are immediately dismissive, without realising that much progress has already been made in animal studies and humans with cognitive enhancement and implants. In addition to animal studies, which may worry many, neurons grown from human stem cells are being used to develop the technology and new implant technology is awaiting approval.
Non-invasive technology
Mark Zuckerberg’s group, is perhaps the best known in this area, He has hired a high-powered team to create a non-invasive device targeted at ‘speech to text’. They use imaging to identify words as they are formed in the brain. We have this ability to rehearse and hear silently. Read that sentence again, internally, in a Scottish accent or in a squeaky voice. Your brain has this ability to rehearse silently and internally. Tapping into this phonological loop may allow us to by-pass our fingers, or actual speech, and type at speeds up to 100 words a minute. This is an admirable goal, of frictionless communication, but others are much more ambitious.
Serial entrepreneur Bryan Johnson has pumped $100 million into Kernel, which hopes to read and write to the brain. Kernel has its eye on serious medical conditions, such as dementia, Alzheimer’s and epilepsy. Building on decades of work on rats by Theodore Berger, the neuro-prosthesis technology has already been tried on real patients. They are developing algorithms which help brains learn faster or develop memories quicker – in other words to learn. 
His aim is to have a commercial product at an attainable price and the stakes are high, not just in the treatment of disease but in education and training. Education and training is an expensive and long-winded business. Young people spend nearly two decades in classrooms and lecture theatres to be even remotely ready for work, and that’s only the start, as once in work the learning continues. Medical treatment may be the stepping stone to enhanced learning.
Invasive technology
DARPA and others have been involved in some very strange research involving implants in insects, rats and sharks.  Humans have been able to control these living beings through electrical impulses controlled by humans. But in human Neurostimulators have long been used to relieve symptoms in neurological disorders, such as Parkinson’s ands epilepsy. 
A pioneer of micro-electrodes implanted in the brain, BrainGate can ‘decode’ the intentional signals that make cursors and limbs move. They are also building wireless devices that allow physicians to monitor brain activity to help diagnose and treat neurological diseases. Another company that is working towards building an implanted chip, that is a modem between mind and machine, is Paradromics. Their focus is also on healthcare.
Looking further ahead, as Musk tends to do, is Neurolink. As well as escaping from our planet and boring into it, Musk wants to bore into our brains. His is a technology play, with arrays of flexible threads, a neural lace, that can be inserted into the brain without tissue damage. They have also developed a robot for the automated insertion of these tiny threads. This builds on the success of neuroprosthetic control in cursor, limb and speech control. But he sees the problem of non-invasive techniques as one of fidelity. With non-invasive techniques, the skull distorts the data so that one is recording the noise of averages. This BMI (Brain Machine Interface) is designed to provide high-bandwidth communications, as the problem Musk is trying to overcome is low bandwidth human interfaces. It can be placed on different parts of the brain and should provide cleaner and more relevant data.
AI in learning
Much of the attention in this filed goes into the hardware – helmets, neural laces and implants but the real challenge is actually in software. These organisations are developing algorithms to interpret the data they receive from brains then constructing data that can be fed back into the brain. Recent advances in AI, and machine learning, especially neural networks, literally extend the brain by using silicon-based neuron-like structures. Daniel Dennett’s rather clumsily titled book ‘From Bacteria to Bach and Back’ makes the case for brains to be evolving Bayesian engines, both physically and culturally. It is this technology that seems to be interfacing and merging mind and machine.
Problems
We have made great progress in technology that relieves symptoms, makes the deaf hear with Cochlear implants, the blind see and the physically disabled walk and type. But the problems behind accelerated learning are immense. The brain has 86 billion neurons and the complex interactions between different types of neurons is still largely unknown. This complex Bayesian inference engine works as a huge parallel processor, so complex that we may be doing no more that reading the hum one hears from a large computer. The formation of memories in the brain may work in ways that are not possible to read from or write to. The technology may be literally like using a fork to do brain surgery.
Then there are the moral issues, such as privacy, identity and personal safety. It is vital that the patients do this voluntarily. That last bastion of privacy, your own thoughts, may now be open to examination, analysis and manipulation. As Daniel Dennett says, the brain is open to infection by memes, that is his definition of cultural evolution. But this means being open to both good and bad memes. We must make sure that these systems cannot be hacked and that they are under strict control.
Conclusion
The two major prizes, are in two sectors – healthcare and education. It seems likely that non-invasive techniques will produce results but limited results, as the channels and bandwidth are diffuse and noisy. Invasive techniques promise cleaner, high-bandwidth data, for two-way communication and the interpretation of data, along with memory formation. In terms of learning, this holds the greater promise, but is much harder to achieve. 
As science fiction becomes reality, we may reflect on what the future holds here. Is it cures for disabilities, so that they allow one to operate as an able-minded and bodied person would? Or is it a world where networked mind-control becomes possible? Will memories and skills be implantable? At the very least the process of learning may be enhanced. We have seen examples of this already. It’s all, literally, mind boggling.