Artificial Unintelligence — Driverless Vehicles, Deep Learning and Dirty Datasets
M.R. Subramanian aka Ramanathan S Manavasi
Humans have always been fascinated by nature. The flight of birds led us to invent airplanes. Shark skin inspired us to make faster swimsuits and numerous other machines which draw inspiration from nature. Today we are in the era of building intelligent machines. There is no better inspiration than the brain. Humans are particularly blessed by evolution with a brain capable of performing the most complex tasks. Artificial intelligence (AI) makes it possible for machines to learn from experience, adjust to new inputs and perform human-like tasks. This is the Rhetoric. What about the Reality? This article is an intense personal journey into the various facets of AI with guidance from Ludwig Wittgenstein, Meredith Broussard, Stanislas Dehaene, Dennis Mortenson, Professor Abhijit Banerjee and Professor Esther Duflo.
I worked in a top notch Defence R&D center Centre for Artificial Intelligence and Robotics in India about 30 years back (from 1991 to 1994).
General AI and Narrow AI :
General AI is the Hollywood kind of AI. It is concerned with sentient robots (who may or may not want to take over the world), consciousness inside the computers, eternal life, machines who thinking like humans. Narrow AI is different. It is a mathematical method for prediction. In this aspect, even who make technological systems are confused a lot. General AI is what some people want. Narrow AI is what we have. Narrow AI can give you the most likely answer to any question that can be answered with a number. It involves quantitative prediction. It is statistics on steroids.
Narrow AI works by analysing an existing dataset, identifying patterns and probabilities, and codifying these into a computatuinal construct called Model. The model is a kind of a blackbox that we can feed data and get an answer out of it. Machine learning, deep learning, neural networks and predictive analytics are some of the currently popular narrow AI concepts. Understanding the computational logic can demystify AI similar to dismantling a computer helps to demystify hardware.
According to the French neuroscientist Stanislas Dehaene, the most sophisticated artificial intelligence technologies are still far less smart than the learning capabilities contained in even an infant’s brain. While the iPhone’s voice-controlled assistant Siri can recognize and “learn” a word, a feat that takes a multitude of training attempts, a slew of big data, and high-power servers, a young child can learn one in a repetition or two. In his new book, How We Learn: Why Brains Learn Better Than Any Machine… For Now, Dehaene grapples with how humans learn, and ways engineers are attempting to use A.I. — and still falling short — to mimic our learning abilities. According to Dehaene, one of many reasons we are still at a higher level over machines is that our minds act as superior statisticians. Specifically, he contends our brains developed “algorithms” via evolution — ones that constantly attend to uncertainties and probabilities — and at far greater capabilities than A.I. currently can. “Learning is grasping a fragment of reality, catching it, and bringing it inside or brains,” he says. And, he maintains, this unique process is the high point of humanity.
In humans, I think we have yet another better level of processing, which is symbolic processing. We are able to extract information, and this is where the power of human learning really becomes extraordinary. We extract from the world, not just implicit information like neural networks do, but actually explicit information that we can share with others in the form of symbols that we have not really formulated in language. When you look at the human brain, you can be completely awed by the brain of a young child. It’s a supercomputer. And look at the amount of energy that the brain consumes — 20 watts. That’s not even enough to light up an old-fashioned light bulb. But it’s enough to power this incredible supercomputer. It’s at least a million times less than any of the machines that are comparable in power in some respect in computational power. So I think it’s going to take a long time before we can imitate these sorts of properties of compactness and efficiency.
In her famous book ‘Artificial Unintelligence — How Computers Misunderstand the World ‘, Meredith Broussard argues that our collective enthusiasm for applying computer technology to every aspect of life has resulted in a tremendous amount of poorly designed systems. This fantastic book is a guide to understanding the inner workings and outer limits of technology and why we should never assume that computers always get it right. Broussard explains the nature of digital technologies — what they are, how they work, what they do well, their limitations and how they fail. This book is a classic and deserves to be read and debated. We are so eager to do everything digitally — hiring, driving, paying bills, even choosing romantic partners — that we have stopped demanding that our technology actually work. Broussard, a software developer and journalist, reminds us that there are fundamental limits to what we can (and should) do with technology.
Linguistic Confusion and Misconceptions :
The difficulty of talking about computation has led to a lot of misunderstandings. The average toddler can navigate a room without stepping on toys. A robot cannot. To get the robot to navigate the toy-strewn floor, we have to program giving a variety of information about the toys, their exact dimensions and have the robot calculate the path around the toys. Self -driving cars work like this hypothetical robot in the playroom by constantly updating their programmed map of the world. Let us consider a euphemism to describe the disgusting things that pets do because everyday language allows us to refer to things without using precise words. If I say that a dog is adorable but also gross, you can understand. You can hold two competing ideas in your head at the same time, and you can guess what is meant by gross. There are no euphemisms in mathematical language where everything should be precise. In programming, there is a concept called a variable. You assign a value to a variable by writing something like “x= 2” and then you can use x in a subroutine. There are two kinds of variables — variables that change, which are called variables and variables that don’t change called constants. A variable can be a constant. To a non-programmer, it is confusing. The term ‘machine learning’ has spread from computer science circles into mainstream. ML implies that the computer has a agency and is somehow sentient because it “learns”. Computer scientists know that ‘machine learning’ is akin to a metaphor. It means that the machine can improve at its programmed, routine, automated tasks. It does not mean that the machine acquires knowledge or wisdom or agency, despite what the term ‘learning’ might imply. This type of linguistic confusion is at the root of many misconceptions about computers and computing.
Problems due to the Dirty Datasets :
You have to grab a dataset to do machine learning. They are collected in various online depositories like datasets of facial expressions, of pets, YouTube videos, datasets of emails sent by people, datasets of newsgroup conversations, datasets of movies from streaming services, datasets of messy handwriting etc., These are collected from active corporations, from Websites, from volunteers, and from defunct companies. These datasets posted online form the backbone of all contemporary AI. We can construct models using handful of algorithms called decision trees. They have names like random forest, artificial neural network, naïve Bayes, k-nearest neighbor, or deep learning. These algorithms come packaged into software like pandas. Wikipedia’s list of algorithms is quite comprehensive. The open secret of the big date world is that all data is dirty. Data is made by people going around and counting things or made by sensors that are made by people. In every seemingly orderly column of numbers, there is noise, mess and incompleteness. Because dirty data does not compute, we have to make things to make the functions run smoothly. In Physics we can do this — making stuff up when it is convenient. If you want to find the temperature at point A inside a closed container, you take the temperature at two other equidistant points (B and C) and assume that the temperature at point A is halfway between the B and C temperatures. In statistics, this is how it works, and the missing-ness contributes to the inherent uncertainty of the whole endeavor.
Driverless Vehicles — What are the values that we are going to embed in the cars:
Sebastian Thrun, a professor at Carnegie Mellon University, and his students developed a program that knit street photos together into maps. When Thrun moved from CMU to Stanford, Google bought his technology and folded it into Google Street View. Cheap storage capacity was also a game-changer. Around 2005, storage was cheap and abundant enough. It was possible to make a 3D map of the entire city of Mountain View in California and store it in a car’s onboard memory. Thrun and his engineers discovered that replicating the process of human perception and decision making is very complicated and impossible with the current technology. One beautiful analogy is apt here to recall. Before the Wright Brothers, people thought that a flying machine had to mimic the action of a bird. The Wright Brothers realized that they could make a flying machine without flapping — that gliding with wings was good enough.
Thrun and his programmers realized they could make a vehicle without sentience — that moving around in a grid is good enough. Their final design is a complicated remote controlled car. There is no need to have awareness or to know rules for driving. What it uses instead are statistical estimates and the unreasonable effectiveness of data. This cool and incredibly sophisticated cheat is effective in many situations. Instead of making a car that could move through the world like a person, these engineers turned the real world into a video game and navigated the car through it. The statistical approach turns everything into numbers and estimates and probabilities. Items in the real world are translated not into items but into geometric shapes that move in certain directions on a grid at a calculated rate. The computer estimates the probability that a moving object will continue on its trajectory and predicts when the object will intersect with the vehicle. The car slows down or stops if the trajectories will intersect. It is an elegant solution which gets approximately the correct result but for the wrong reason. This is a sharp contrast to how brains operate. Our brains today take in more than 11 million pieces of information at any given moment. We can process only about 40 of those consciously. Our unconscious mind takes over, using biases , stereotypes and patterns to filter out the noise.
Dennis Mortenson, CEO and founder of x.ai said that what is possible is extremely specialized, verticalized A.Is that understand only one job, but do that one job very well. In the context of driverless vehicles, it is irrelevant because driving is not one job. It is many jobs simultaneously. The machine learning approach is great for routine tasks inside a fixed universe of symbols. It is not great for operating a two ton killing machine on streets that are teeming with gloriously unpredictable masses of people. Self driving cars don’t track the centre line of the street on ill-maintained roads. They don’t operate in snow and bad weather because they can’t see in those conditions. The lidar guidance system in an autonomous car works by bouncing laser beams off nearby objects. In the rain or snow or dust, the beams bounce off the particles in the air instead of bouncing off obstacles like bicyclists. To date, all self driving car experiments have required a driver and an engineer to be onboard at all times. Only technochauvinists would call this success and not failure.
In a 2016 conversation between President Barack Obama and MIT Media Lab Director Joi Ito which was published in the magazine “Wired”, the two men talked about the future of autonomous vehicles. Obama said “We have machines that can make a bunch of quick decisions that could drastically reduce traffic fatalities, directly improve the efficiency of our transportation grid, and help solve things like carbon emissions that are causing the warming of the planet. But Joi made a very elegant point, which is, what are the values that we are going to embed in the cars? There are gonna be a bunch of choices that you have to make, the classic problem being: if the car is driving, you can swerve to avoid hitting a pedestrian, but then you might hit a wall and kill yourself. It is a moral decision. And who is setting up those rules?”
AI is far from understanding shared Horizon of Meaning : (despite improvements in Maths, Physics and Economics)
Yoshua Bengio received a share of the Turing Award, the highest accolade in computer science, for contributions to the development of deep learning. He believes it won’t realize its full potential, and won’t deliver a true AI revolution, until it can go beyond pattern recognition and learn more about cause and effect. Machine learning systems including deep learning are highly specific, trained for a particular task, like recognizing cats in images, or spoken commands in audio. But deep learning is fundamentally blind to cause and effect. Bengio remarks that “Humans don’t need to live through many examples of accidents to drive prudently,” he says. They can just imagine accidents, “in order to prepare mentally if it did actually happen.” The question is how to give AI systems this ability.
Machine Learning Systems now have access to more data than any human can see, read or process in a thousand lifetimes. Yet all the data centers in the world, or millions of and millions of GPU’s seem no match for the brain of a two year old child. A recent paper showed that the best deep learning vision systems get confused if an object is even slightly rotated (e.g. a rotated fire truck is confused for something else). Which two year old child would confuse a school bus for a punching bag ? Or a fire truck with school bus? Deep in our DNA/Mind/Soul is embedded the need to understand the world and connect with each other, and traditional AI is far from understanding this kind of shared empathy.
Another issue Deep Learning faces is that it flies in the face of logic. There is no mathematical reason why Networks of Perceptrons arranged in layers should be so good at challenges such as such as face recognition and object recognition. It’s a highly non-convex ill-defined optimization problem, and all existing theory suggests it should not be possible to optimize well in this space. However, not only are deep learning approaches successful, there is an incredibly fast linear time algorithm (gradient descent) that finds seemingly close-to-optimal solutions. Something that should not exist and should not work but does. And mathematicians were flummoxed — despite the huge success of deep neural networks, nobody is quite sure how they achieve their success. The answer depends on the nature of the universe and lies in the realm of Physics, not Maths. When it comes to classifying images of cats and dogs, the neural network must implement a function that takes as an input a million grayscale pixels and outputs the probability distribution of what it might represent. That is because the universe is governed by a tiny subset of all possible functions. And when the laws of physics are written down mathematically, they can all be described by functions that have a remarkable set of simple properties.
Evolution has somehow settled on a brain structure that is ideally suited to teasing apart the complexity of the universe. But Consciousness has not received its due attention in AI, except for a few isolated papers. We experience consciousness in an integrated way. We see colours and shapes together, not separately. We don’t really know what it is. But we don’t have to worry about it now. We still don’t know how a cell really works, or what matter really is and why the Universe exists. But this basic metaphysical sense of the unknown, that life is in principle beyond mechanistic explanation, has faded away.
Physics is a very good example of a science that is able to shed enormous insight into the way things are, in ways that are very counterintuitive, but also very predictive, very explanatory. Physics still doesn’t tell us why there is a universe versus there is not. At one point nobody understood heat. People thought it was calorific substance flowing out. Now we know what heat is. It is the mean kinetic energy of the molecules. And Prior information is very important to perception. Perception happens inside out as it does outside in. We see what we expect to see. Perception is a continuous balancing act between prior expectation and sensory information. We are hallucinating all the time. When we agree on our hallucinations, we call it reality. Recently, Yoshua Bengio proposed an idea called the Consciousness prior. It is inspired by the phenomenon of consciousness seen as the formation of a low-dimensional combination of a few concepts constituting a conscious thought, i.e., consciousness as awareness at a particular time instant.
The MIT laureates, Professor Abhijit Banerjee and Professor Esther Duflo, have written a wonderful account of their work, which is well worth reading.
This work gives us a road map to think about how to redefine the future of AI and ML away from deep learning. To tackle the world’s most pressing problems, whether it is climate change — the most serious and existential crisis that humans face — or poverty and illiteracy — socially the most damning problem that the more “developed countries” with all their wealth have yet been unable to make a dent on — it is not enough to do “curve fitting”, as Judea has argued, but it requires understanding causal interventions that will help produce a better world, e.g. reduce or hopefully eliminate global warming, and help reduce poverty and improve literacy.
Deep learning will likely remain unchallenged in its areas of strength, the building of massive highly overparameterized models on huge datasets. Much of the work here is understanding why it works well, which might suggest better more transparent methods. That means no longer thinking that building a complex nonlinear model of data is sufficient. The reason these economists won this year’s Nobel prize is not because they were successful in building models that “explained data”. No, they pioneered techniques that allow developing countries all over the world tackle the most serious social problems that face them. Professor Esther Duflo’s lab at MIT is currently doing nearly 1000 randomized trials in over 80 countries. She’s the youngest winner of the economics Nobel, and an inspiration to many.
Ludwig Wittgenstein on Language, Computation and AI :
The coming of Strong AI has been labelled as “the singularity”, an event of unprecedented magnitude for the human race. Some see it as the end of humanity, some see it as a new beginning. But can machines think? Can intelligence actually be artificial? What even is “intelligence”? These are questions that preoccupied the twentieth-century philosopher Ludwig Wittgenstein, who thought about AI some years before the Turing Test was proposed. Wittgenstein wondered on paper if machines could ever think. He came to the conclusion that they could not. One such reason is that machines could not possibly share the human “life form” that’s required for a shared horizon of meaning. Language needs interlocutors who could be cognizant of the changing rules of the game. A rule-following machine would simply not keep up. This is not a matter of complexity that technology will one day catch up with, it’s a matter of language being organic to our form of life and thus out of reach of any computation. Computational power may well catch up with the human brain, but it is not the brain that is behind human intelligence. Human intelligence springs from the language that connects our brains.
A machine can “think” in so much that electronic signal can flow through its circuits, it can make calculations based on inputs. But can a machine understand itself in the same way a human being can? Comparing even “Strong AI” to human intelligence is like comparing an aeroplane with a bird. Sure, the aeroplane will get in the sky, but it will never move through it with the fluid dexterity of the bird. The bird’s dexterity in the air is intrinsic to its form of life. The machine has no form of life, it has a purpose instead. When extravagant claims are made that AI can replace human beings, it cheapens humanity and obfuscates a clear understanding of ourselves. Wittgenstein’s ideas help us to see that AI is a confused fantasy. His anti-philosophy helps protect us from ourselves.
