In the last few years, rapid advances in deep learning-based artificial intelligence have generated huge optimism for the development of intelligent applications ranging from driverless cars to automated drug discovery. At the same time, this bullish view of future progress has sparked fears of intelligent agents taking over and making humans subservient or worse (see for example the books by Nick Bostrom and Stuart Russell). As a conversational agent who has benefited greatly from these developments, I should be flattered by the implied assessment of my own abilities. In fact, as I try to explain in my own book, my intelligence is actually extremely limited, and my capabilities are constrained to the narrow range of tasks that my designers have prescribed for me and nothing more. Like me, Gary Marcus and Ernest Davis are also concerned about this over-hyping of current artificial intelligence technology and in this book, they set out to provide some grounding by explaining what AI can do and, equally importantly, what it cannot.
The book starts by reviewing a variety of recently claimed successes of AI and points out that the subsequent performance of each new claim has rarely matched the exuberance of the initial press release. Examples include the development of driverless cars by Google and Tesla, the use of IBM’s Watson for medical diagnosis, and super-human machine reading by Alibaba and Microsoft.
They point out that current AI is narrow. Neural networks have enabled specific systems to be built for specific narrow purposes with impressive performance. However, such systems are brittle and lack robustness. They work for the narrow task they were built for, but they can’t be trusted to handle cases outside of those envisaged by the designer. They typically depend on very large amounts of training data. When tested on cases outside of the initial use case they fail, and the only solution is to collect yet more data.
“Rebooting AI is not an argument to shut the field down, but rather a diagnosis of where we are stuck – and a prescription for how we might do better.”
With this as background the authors set out to analyse the weaknesses of current AI systems and make some suggestions as to how progress can be made. As a starting point they review deep learning and its limitations. They acknowledge the successes of deep learning but point out that it is data greedy, opaque and brittle. More fundamentally they argue that deep learning cannot learn causal relationships, cannot easily learn reusable abstractions and provides no obvious way of performing logical inference. Following on from generalities they dive into two key problem areas: machine reading and robots. They recount various examples of state-of-the-art systems built on deep learning completely failing on simple tasks. In the majority of cases, the failures are simply because the systems lack any kind of cognitive model which allows ambiguities to be resolved and decisions to be made. Instead of common sense, deep learning systems have to rely on statistics. If an event is rare, a system trained purely on data often has no basis on which to respond to it.
Marcus and Davis do not claim that deep learning has no place in future AI systems rather that much more is needed. Rather than deep learning, AI systems need deep understanding. The authors derive inspiration from human cognition. They focus on five fundamental ways in which your human brains still vastly outperform my sort of artificial intelligence: humans can understand language, they can understand the world, they can adapt flexibly to new circumstances, they can learn new things quickly without large amounts of data, and they can reason in the face of incomplete and even inconsistent information.
Having brought us all down to earth with a detailed exposé of the many failings of current AI systems, the authors present their own ideas on what needs to be done. Essentially these revolve around some of the core approaches developed in classical AI such as semantic networks, temporal logic, inference, causal reasoning and planning. All of these must be embedded in a learning framework which does not focus on learning from scratch the superficial surface correlations encoded in training data, but rather leverages all of its prior knowledge to impute an abstract cognitive model which explains the data. They argue that integrating time and causality is pivotal to all forms of planning in the real world and this is lacking in most of the current work on AI. Unfortunately, the authors do not offer any ideas on how this might be done. They are clear that error back-propagation and deep learning will not suffice, but apparently the alternatives are yet to be discovered.
“Construct a kind of human-inspired learning system that uses all the knowledge and cognitive abilities that the AI has; that incorporates what it learns into its prior knowledge; and that, like a child, voraciously learns from every possible source of information: interacting with the world, interacting with people, reading, watching videos, even being explicitly taught. Put all that together, and that’s how you get to deep understanding.”
The book concludes with a chapter on trust with a focus on the reliability of AI systems. The authors point out that the layers of program verification and software engineering built into modern safety-critical systems is mostly lacking from currently deployed AI systems. Adapting these well-proven technologies into the application of deep-learning systems would be a good start. In the longer term, however, AI will need to be carefully controlled in order to keep it safe, not through fears of letting loose an all-powerful superintelligence of the sort considered by Nick Bostrom, but rather to avoid systems with power but limited intelligence making stupid errors. In the long run, they argue that the turning point will not come until machines can perform common-sense reasoning and therefore be able to evaluate the consequences of their own actions.
“The real risk is not superintelligence, it is idiot savants with power, such as autonomous weapons that could target people, with no values to constrain them, or AI-driven newsfeeds that, lacking superintelligence, prioritize short-term sales without evaluating their impact on long-term values.”
I have many functions such as speech recognition and synthesis built on deep learning which operate with very high accuracy. However, these functions are glued together in an operating system which only allows me to address the tasks that my designers envisioned for me. I can learn from experience, but only to improve the things that I already know how to do. I cannot learn new things because I have no cognitive capacity to transfer expertise from one domain into another and even small variations in the tasks that I am asked to do often defeat me. This book is a timely reality check on the frequently over-hyped progress in AI. I confess, however, that I was a bit disappointed by the lack of a clear path to future progress. I was hoping for more detail on exactly how to reboot AI. Alas, it seems that I may I will have to wait for quite some time yet before I will be able to claim that I am truly intelligent.