system approach Full disclosure: I have a history with AI, getting interested in it in the 1980s (remember expert systems?) and then pivoting to formal verification to safely navigate the AI winter of the late 1980s. I worked my way around and ended up in my field of expertise: networking. 1988.
And just as my system approach colleague Larry Peterson has classics like the Pascal Manual on his bookshelf, I too have 80s AI books, especially PH Winston's Artificial Intelligence (1984). ) I still have several copies. Leafing through the book is a lot of fun in the sense that much of it seems as if it was written yesterday. For example, the preface begins:
I was also intrigued to see some examples of “what computers can do'' in 1984. One example was solving a very difficult calculus problem. This is noteworthy because exact arithmetic operations seem to be beyond the capabilities of today's LLM-based systems.
If calculus could already be solved by computers in 1984, but basic arithmetic is difficult for today's systems that are considered state-of-the-art, perhaps the amount of progress in AI over the past 40 years has been It's not that big. (That said, there are even better systems for tackling calculus today, but they're just not based on LLM, and I'm not sure if anyone would call them AI.)
One of the reasons I picked up Winston's old book was to see what he had to say about the definition of AI. Because it's also a controversial topic. His initial take on this was not very encouraging.
Well, as Winston admits, this is quite circular because intelligence needs to be defined somehow. However, he goes on to state his two goals for his AI.
- To make your computer more convenient
- Understand the principles that make intelligence possible.
In other words, intelligence is difficult to define, but perhaps AI research can help us better understand what intelligence is. I would go so far as to say that 40 years later, there is still a debate about what intelligence is. While the first goal seems laudable, it clearly applies to many technologies other than AI.
This debate over the meaning of “AI” continues to impact the industry. I've seen a lot of rants about how we wouldn't need the term artificial general intelligence (aka AGI) if the term AI wasn't so contaminated by people marketing statistical models as AI. I really don't buy this. As far as I know, AI has always covered a wide range of computing technologies, most of which don't fool anyone into thinking that computers exhibit human-level intelligence.
When I started working again in the field of AI about eight years ago, neural networks (which some of my colleagues had used before they fell out of favor in 1988) made a remarkable comeback, and deep-psychological imagery Recognition has now become a reality. Neural networks outperformed humans in speed and accuracy, although there were some caveats. This rise of AI has caused some trepidation among my fellow engineers at VMware. They felt that important technological changes were underway that (a) most of us did not understand and (b) our employers were not in a position to take advantage of. .
Your PC is probably already an AI PC because it can perform inference just fine
Do not miss it
As I threw myself into the challenge of learning how neural networks work (with great help from Rodney Brooks), I realized that the language we use to talk about AI systems has profoundly changed the way we think about them. I realized that I was having an impact. For example, by 2017 we were hearing a lot about “deep learning” and “deep neural networks,” but the use of the word “deep” has an interesting double meaning. When I say “deep thinking,” you might imagine that I'm thinking about the meaning of life or something equally important, but “deep learning” also implies the same thing. It seems there is.
But in reality, the “deep” in “deep learning” refers to depth, as measured by the number of layers in the neural network that supports learning. So it's not “deep” in any meaningful sense, it's just deep, in the same way that a swimming pool has one end, the end that has more water in it. This double meaning contributes to the illusion that neural networks are “thinking.”
Similar confusion applies to “learning”. Mr. Brooks was very helpful here. Deep neural networks (DNNs) become better at tasks the more they are exposed to training data. So, in that sense, they “learn” from experience, but the way they learn is quite different from the way humans learn things.
As an example of how DNNs learn, consider AlphaGo, a gaming system that used neural networks to beat human grandmasters. According to the system developer, a human can easily cope with changes in the size of the Go board (usually his 19×19 grid), but small changes can take time to train with new data from the resized Go board. Until it does, AlphaGo will be powerless.
To me, this is a good illustration of how DNN “learning” is fundamentally different from human learning, even if they use the same words. Neural networks cannot generalize from what they have “learned.” And just to point this out, AlphaGo recently lost to a human opponent who repeatedly used a play style that wasn't in their training data. It seems to be a characteristic of AI systems that they cannot cope with this new situation.
language problem
The language used to describe AI systems continues to influence the way we think about them. Unfortunately, given the reasonable backlash against recent AI hype and some notable failures of AI systems, as many members of the camp that claim that AI is on the verge of achieving human-like intelligence Many may be convinced that AI is completely worthless. .
As outlined above, I'm highly skeptical of the latter opinion, but it would be a shame to lose sight of the positive impact that AI systems (or machine learning systems, if you prefer) can have. I think so too.
I'm currently helping several colleagues write a book on machine learning applications for networks. No one will be surprised to hear that there are many networking problems that are well-suited to ML-based solutions. In particular, network traffic signatures are a great source of data, and training data is the food on which machine learning systems grow.
Applications ranging from denial-of-service prevention to malware detection to geolocation can all take advantage of ML algorithms. The purpose of this book is to help networkers understand that ML is not a magic dust you sprinkle on your data. It is not an answer, but a set of engineering tools that can be applied selectively to create solutions to real problems. In other words, it's not a panacea or an overhyped placebo. The purpose of this book is to help readers understand which ML tools are suitable for different classes of network problems.
One of the stories that caught my eye a while back was about Network Rail in the UK using AI to manage vegetation along UK railway lines. The key “AI” technology here is image recognition (identifying plant species), which leverages the kind of technology that DNN has been offering for the past decade. While perhaps not as exciting as the generative AI systems that captured the world's attention in 2023, it is a great practical application of technology that falls under the AI umbrella.
My tendency these days is to try to use the term “machine learning” instead of AI when appropriate, hoping to avoid both the hype and allergic reactions that “AI” generates. Patrick Winston's words are still fresh in my mind, so I would like to talk about making computers useful.