Why our approach to Artificial Intelligence needs a rebootBy Paul Laughlin
Do you see the limitations & overhyped expectations of today’s approach to Artificial Intelligence? Does it need a reboot, a redirection, to finally achieve the potential of AI that truly understands us & we can trust?
That is the premise of a great book on the subject, “Rebooting AI: Building Artificial Intelligence we can trust“, from Gary Marcus and Ernest Davis. Gary is both a Professor of Psychology at New York University and co-founder of Robust.AI. Ernest is a Professor of Maths at New York Uni and one of the world’s leading scientists on commonsense reasoning for AI. A quote on the cover nicely sums up why this book is needed:
“Finally, a book that tells us what AI is, what AI is not, and what AI could become if only we are ambitious and creative enough.”Garry Kasparov (chess grandmaster)
A conceptual understanding of AI past and present
I recommend this book for both those working in this field and non-technical leaders. One of the reasons for that endorsement is the care these authors take to avoid jargon and explain both technical terms and the conceptual thinking that underpins them.
Packed full of examples to support their points, the big picture of this book is that it gives a realistic picture of current progress with AI. Building on that, it highlights why there is an important gap between this progress and the vision many people have of a SciFi future where machines are at least as intelligent as people. The authors explore what positive progress has been made in the field of Deep Learning and using available Big Data. They also point out the limitations of this approach and what else is needed.
Through real world challenges like reading, navigating changing environments and learning from general principles – they highlight how limited many current models (parlour tricks) really are. They build on this by considering then human mind and how we think and learn. But this is no theoretical exercise. From a societal point of view they highlight the need for machines to have common sense, gain deeper understanding & be trustworthy.
So, despite the justified critique in this book, these are wounds from a friend. The authors obviously care about the future of AI and are here making a case for a fundamental reboot to ensure it does not (once again) fail to reach its potential.
Topics covered within this book (so you can dig deeper)
In just over 200 pages, so a very accessible paperback, the authors cover a lot of ground. Here are my takeaways of some of the highlights to explore in each chapter.
Chapter 1: Mind the gap
Identifying the limitations of current “Artificial Intelligence“. Building on the Turing Test with 6 useful questions to ask any apparently intelligent system. It paints a challenging but honest picture of progress to date.
“The bitter truth is that for now the vast majority of dollars invested in AI are going toward solutions that are brittle, cryptic and too unreliable to be used in high-stakes problems.”Rebooting AI, Marcus & Davis
This chapter also explains 3 key problems: the fundamental over-attribution error; the illusory progress gap; the robustness gap. They close by making the case that to move past the current ‘AI Chasm’ requires a clear sense of what is at stake, understanding of why current systems aren’t getting the job done, and a new strategy.
Chapter 2: What’s at stake
This rather chilling chapter brings to life what can go wrong if we don’t make AI smarter. If we let loose systems and robots that are too “brittle, cryptic & unreliable“. Through considering the real world jobs we would like to automate they really help bring this to life. So, we are no longer so impressed by AI machines winning games or gameshows.
Through examples in practice, they bring to life 9 key risks:
- The fundamental over-attribution error
- The lack of robustness
- Relying heavily on the precise details of training data sets
- The perpetuation of obsolete social biases (encoded in data)
- Echo-chamber effect (learning from data it generated)
- Gaming the system (relying on publicly generated data)
- Amplification of social bias (due to combination of above effects)
- Too easy to end up with the wrong goals
- Risk of being used deliberately to cause public harm
Given the limitations of existing AI systems, yet still the seriousness of a number of these risks, it’s a sobering list. The authors make a compelling case for why this needs to be addressed before greater automation (with insufficient intelligence) puts us all in danger.
Chapter 3: Deep Learning and beyond
Drilling down beneath general concepts, this chapter does a good job of introducing the reader to field of Deep Learning. As someone who was working supervised and unsupervised neural networks back in the 1990s, I liked this introduction. It covers the concepts that matter without burying the reader in technical jargon. There are also many useful references to leaders in the field or studies to read further.
Here they also introduce a framework and nomenclature for AI that usefully positions Machine Learning as a subset and Deep Learning as a subset within that. Once again, as someone with a history in AI usage prior to the AI Winter, I was glad to see the acknowledgement of the role of other approaches. Too many modern writers equate AI to variants of Deep Learning, ignoring Decision Trees, Probabilistic Learning, Rule Induction and other knowledge representations in Classic AI. More on that later.
This is a balanced review. It acknowledges the tremendous and exciting progress that has been made in this field. But, it is also honest about the limitations and risks of limiting AI to this approach. The authors cite 3 core problems with Deep Learning and related approaches:
- Deep Learning is greedy (huge amounts of data needed to cope with variability of real world)
- Deep Learning is opaque (we still lack explainable or accountable models)
- Deep learning is brittle (images & language only need tweaking from familiar to fool it)
Chapter 4: If computers are so smart, how come they can’t read?
I hadn’t given this question serious thought before reading this book. But the authors bring the topic to life as a way of highlighting the limitations of current AI approach. When you consider the subtlety and variability of how language can be used and the questions we can ask, you see the issue. Once a system relies on past examples & pattern matching, rather than understanding of the content, it is easily fooled.
The authors walk through numerous examples of questions that are too difficult for current AI systems (text or voice based). They also help explain how the current approach will never get there. Deep Learning systems will never be able to truly learn from the wealth of humanities written record. Not while they rely on matching to past examples rather than comprehending context and composition.
It’s this chapter that begins to show how an older approach to AI has a role to play. How work on knowledge representations and encoding existing human understanding, from ‘Classic AI’ is part of the solution. They rightly suggest that both the power of memory in Deep Learning & the cognition models from Classic AI are needed to read. But also highlight a missing piece, common sense.
Chapter 5: Where’s Rosie
Now the book turns the world of robots (including autonomous cars). If the inability of AI systems to read and understand was concerning, some of these examples are terrifying.
The authors review the progress that has been made in robotics. Examples of progress on localisation and motor control are both impressive. But they then go on to highlight very important gaps between current capability and what is needed. Explaining the importance of robots being able to:
- use situational awareness (in a changing world)
- figure out what is the best thing to do now (fast enough to react)
- possess a general-purpose understanding to apply to unfamiliar challenges
A few real world applications really bring to life the risks to humans of such limitations. The robo-carer that drops an elderly patient because they react in a way not seen before. The autonomous car that protects its driver at the cost of many other lives. We don’t have general-purpose domestic robots yet, because they are not yet flexible enough to cope with the real-world.
Chapter 6: Insights from the Human Mind
OK, if AI can’t yet compete with a young child with regards to reading and navigating the real world, can AI learn from how we think? This chapter explores the potential of our current understanding of the human mind and cognition. They offer 11 clues to help AI, drawn from the Cognitive Sciences:
- There are no silver bullets (multiple improvements are needed)
- Cognition makes extensive use of internal representations (mental models)
- Abstraction and Generalisation play an essential role in Cognition
- Cognitive Systems are highly structured
- Even apparently simple aspects of Cognition require multiple tools
- Human thought and language are compositional
- A robust understanding of the world requires bottom-up & top-down info
- Concepts are embedded in theories
- Causal relations are a fundamental part of understanding the world
- We keep track if individual people and things
- Complex cognitive creatures aren’t blank slates
Chapter 7: Common Sense and the path to Deep Understanding
Here the book turns to a more progressive mindset. Exploring options and the progress made in doing what has been advised earlier. That is, learn from human cognition and make progress in building a deeper understanding and common sense within AI models. In summary, it is difficult and progress has been slow.
The review progress with representations of knowledge and developing an ability to learn. A helpful way to think about this is their use of Kant’s three knowledge frameworks (ability to recognise & reason about time, space and causality). It is insightful how much humans rely on all three millions of times a day, yet each is challenging for AI.
Once again, we see the benefit of blending the best thinking that has happened within Classic AI and technological progress within Deep Learning. The authors recommend:
- Develop systems that can represent those 3 knowledge frameworks (time, space, causality)
- Embed those within an architecture that can be freely extended to new knowledge
- Develop powerful reasoning techniques (able to work top-down & bottom-up)
- Connect these to perception, manipulation & language
- Build the above into rich cognitive models of the world experienced
- Construct a human-inspired learning system (learn from prior knowledge & available sources)
No one said that building Deep Understanding was easy!
Chapter 8: Trust
In many ways we are back where we first began but seeing the place anew. Why does this matter? Plus, what else is needed for AI to truly help us? The answer is trustworthy AI is needed. So many examples used throughout this book show how critical it is that our increasingly reliance on AI and automation is matched by AI that we can trust.
The authors make a number of points here that don’t get enough airtime. They go beyond the regular chatter about Ethics policies, inclusion and the risks of social biases in data and Deep Learning models. For instance, they highlight the need for AI models to learn from good engineering practice. How good engineers: design solutions to be stronger than the limits in brief; design for failure; incorporate fail-safes. Such safety-first thinking is needed.
Trustworthiness can also be aided by best practice from wider software engineering. The use of modular design for improved understanding & future flexibility. The use of good metrics for testing. Documentation and the ability to debug code. In addition to all this, there is also the need to encode ethical values into the reasoning of AI systems. Not just filtered data, but something closer to Asimov’s 3 Laws of Robotics. These all need to be supported and challenged by a regulatory system that is up-to-date and fit for purpose (knowing what questions to ask).
Why this book should encourage Data Leaders
I’ve mentioned before that as someone with a background in both Deep Learning and Classic AI, I am encouraged by this book. It does indeed advocate a reboot, but not a withdrawl. Far from being an argument against the use of AI, it is a challenge to ‘build back better’. To use the best thinking from Classic AI & Cognitive Sciences, coupled with the innovations in Deep Learning and Robotics. As the authors themselves put it:
“The only way out of this mess is to get cracking on building machines equipped with common sense, cognitive models and powerful tools for reasoning.”Rebooting AI, Marcus & Davis
So, as well as being a wake-up call on the need to change, this book should be an encouragement to Data Science leaders. Set higher-goals for your AI capability. Think bigger and engage more with what the real world needs, not the chosen focus of current technology solutions. There are a number of resources at the end of this book that can help you take that forward. A curated list of ‘Suggested Readings’ which the authors navigate with you. There’s also a comprehensive set of footnotes and endnotes, together with a lengthy bibliography & topic index.