Machine Learning (ML) is actually the correct term for what most of the public companies term as Artificial Intelligence (AI). And as
@Cloudyskies22 pointed out - it is not an Apple term. Machine Learning is actually a subset of Artificial Intelligence, but tends to be limited to just recognising patterns from data. When you build ML models, there are both assisted and unassisted learning to generate a model that is then used to basically categorize or recognize what the data provides (example of assisted which is more common - you have a human sit there and they see a picture for example and they tell the computer it is a cat, from all that human decisions - the computer builds an ML Model). Driverless car systems are basically the same, humans are telling the cars what things are and it builds models from it. Artificial Intelligence goes beyond that to analysing and contextualizing observed data and make decisions based on that.
I don't know any company that does not like (need) to trademark stuff - in fact, they would be stupid if they did not. Apple has thousands, so does Microsoft and Google.
Good points on the product.
However, the discussion here (not just in your post) is not quite exact in terms of the various concepts and relations in the milieu of AI/ML. I don't really want to ruffle feathers here, but as a long time researcher in this field, I just want to help paint a more accurate picture of what we do, and set the record straight.
AI (artificial intelligence) refers to the natural scientific and engineering disciplines that correspond to discovery and replication of human and other biological systems' abilities. AI is often discussed in terms of its various application sub-domains such as computer vision, NLP, robotics, etc. Those who work primarily in the various application doamins are probably behind those technologies from AI/ML that most people normally interact with. AI in its modern form (in about the last 15 years) primarily relies on various techniques in ML, but there are still certain applications where a more traditional types of non-ML techniques, such as formal logic (e.g. syntactic parsing in language) still play a major role.
ML (machine learning) is the theoretical (and for the most part mathematical) discipline that lies behind many of modern AI's capabilities. While the primary application of ML is in AI, there are a variety of other science and tech fields where ML also play some role, such as meteorology, agent based economics, simulation based physics, geostatistics, etc. In addition to mathematics, ML to a lesser degree also involves a certain amount of cognitive and neural science.
ML for the most part relies on, and is an outgrowth of three more fundamental types of mathematics:
- frequentist and Bayesian statistics (kind of goes without saying that statistics and probability theory are important here),
- numerical optimization (especially constrained convex optimization), which is critical any time when "training" of a model is involved
- differential geometry (including linear algebra, which is a special case of diff geom) which is intimately involved in the basic representations within parametric space of ML models, any time you hear the reference to terms like "tensor" or "embedding" (meaning embedded manifold in ambient representation space) in AI/ML, it likely has to do with diff geom.
ML also involves to various degree other areas in mathematics; such as dynamical systems (such as in recurrent neural nets and certain types of reinforcement learning); stochastic processes (such as using stochastic differential equations in diffusion models, e.g. DALL-E that you might have heard of); computational semantics (such as embeddings used in data space representations); and a bunch more.
AI/ML is actually a very diverse set of things that can be sliced and diced in many different way. The most important ways to distinguish different forms of ML probably are:
- the type of relation between the data and the model in the training procedure, sometimes referred to as the "setting", in other words how information from the data gets into the model. These include supervised, unsupervised, self-supervised, distributional filtering, normalizing flow, adversarial learning, reinforcement learning, guided diffusion, imitation learning, GA/ES (genetic algorithms evolutionary strategies), and a number of others.
- the type of flow of information through the structure of the model during inference. The two best known types are discriminative and generative models, although these are just the two extremes, there are various others in between such as parametric ensemble.
- the information density of the model within parametric space. Here again there are two very commonly seen types, which are point-estimate/frequentist, vs Bayesian/probablistically-coherent. But these are really just two within a larger continuum of options.
- the form of representation within model/parametric-space. It would be either parametric, meaning that the number of numerical parameters of the model is pre-determined at the beginning of the training process; or would be non-parametric, meaning that the number of parameters can vary, and usually grow throughout the training process (so the model can vary structurally as it gains knowledge from data).
What you have referred to in "you have a human sit there and they see a picture for example and they tell the computer it is a cat, from all that human decisions - the computer builds an ML Model", is a very specific type of ML system that is
- supervised (in how information from data is integrated into the model, in this case you mentioned this always treating human response as the gold standard)
- purely discriminative (in inference mode the information flows straight from input through a sequence of transformations to output)
- parametric, since the type of model and its structure is fully determined prior to training base on your description
- and likely frequentist, base on your description of a single model and just general knowledge these type of vision tasks do not involve Baysian systems
While your description is very good and correct, it is only a very narrow slice of what AI/ML is. This specific type of system is the simplest, and possibly the most common type in the industry. So it is unsurprising that it gets brought up all the time. But a lot of people often do not realize that there are many other variants, some are also quite often used in real world scenarios.
The other example application, driverless cars (or planes, helicoptors, other verhicles for that matter), are usually a very different type of ML system. Some of the simpler types of self-driving cars, such as those that run on mostly fixed route or only drive on highways, generally are of the simpler types such as using a regulator setting (e.g. linear quadratic regulator / LQR) as the control system, and some convolutional/residual neural nets + a filtering setting (e.g. Kalman filter) for the perception and sensor integration. Those typically are very far from the full capabilities of a human driver/pilot. Those systems that are much closer to human ability in these applications typically has a reinforcement learning based control system; but that is not very close to being realistically deployable in real life by a long shot.
Again, I don't want to bore you guys, or get into a whole a lot of minutiae. But there are so many misconceptions and gross simplifications out there right now about AI/ML, that I just want to bring a bit more clarity, and to help inform folks in these discussions, here in this forum. I actually primarily do research in reinforcement learning, among various other forms of AI/ML, so would be happy to answer things, time permitting from my busy schedule.
