AI and Machine Learning – The Basics
steveblank.com
steveblank.com
1) NLP performance characterized to be better than vision systems. I don't think that is true.
2) Some minor facts are not right. E.g. OpenAI did GPT-3, not Google
3) I expected the set of exemplars for technology to be better researched. Siri and Alexa are NOT recommendation engines. Similarly, lacked the best of class examples on many fronts. This was the biggest issue in the paper.
For recommendation systems, the top three examples that come to mind are TikTok, Layer 6 (a Canadian company that TD Bank acquired a few years back) and Netflix.
You may want to add Nerfs to the paper. They are the hot new algorithm out there. I am a scientist at a Canadian research lab and my very smart colleagues tell me it is the next best thing.
Automated vision is far head of NLP IMHO. NLP had it's Imagenet moment only at the advent of BERT (which was cira 2018? or so). Also, Transformers, which BERT and its progeny rely on, are massively compute and data hungry. They are also slow to run on today's chips. In my opinion, reason is that language benchmarks aren't as clearcut as vision. For instance, NLP researchers use BLEU scores, which are a pretty blunt instrument. I'd say NLP is even further behind than speech processing (which is now mostly based on DL). A key person behind Siri is Adam Cheyer btw .. he did Siri and then Bixby. The way these NLP systems work is pretty simple conceptually .. they break the problem to two steps. Intent Identification and then Slot filling. You can use DL for both steps but don't have to. Key issue with NLP systems is they are extremely brittle (a ton of work to customize). Dialog is pretty weak today, and that is partly due to the challenge in training signal.
You say 5000 images per images of a class. I know that was a ballpark but this seemed misleading. There are at least 2 problems I see. First, you need "different" examples .. seeing the same examples (e.g. from different viewpoints) does not help. Second, it really matters what the set of classes your model is trying to discriminate against. E.g. to differentiate apples and bananas, I likely need far fewer than 5000 examples since they are so visually distinct. Imagenet was a seminal moment not just because of the number of examples per class but because of the humongous number of classes (10K+).
For RL and robotics, there have been some neat advances. I was skeptical about RL's utility in practice (due to reasons you point out .. simulations vs. real-world, and especially the issue of faster than real-time) but am seeing it more and more in practice. E.g. 5/6G applications exist.
You may want to add coverage of some important emerging topics: multi-modal (matching vision to text and vice-versa), sensor fusion, student-teacher.
Paper didn't talk about any work from MIT's Han lab or their startup OMNIML? They just launched at the TinyML summit this year and they are hot! Also, TVM tech (startup behind it is called OctoML) is pretty important for on-device AI.
Those were some initial thoughts. If this is useful, can add to it later.
One thing missing in the article is the exponential growth rate of progress. The rate of progress is something I try to explain to non-tech friends. I love seeing difficult problems solved and then simply become new engineering tools to build with.
Transformer models like GPT-3 and Co-Pilot have so quickly transformed my work flow and especially GPT-3 has transformed things that I can do.
> An AI can see and understand what it sees. It can identify and detect an object or a feature in an image or video. It can even identify faces.
Why do we switch to this kind of language when we wouldn't for a standard algorithm.
You might say casually "the algorithm has a bug so it doesn't recognize barcodes with a 0 in them" or this door opens when it recognizes a human but it's still very clear that it's not thinking whereas AI people seem to intentionally blur this line to get headlines and attention.
Another example:
> It’s taken decades but as of today, on its simplest implementations, machine learning applications can do some tasks better and/or faster than humans
You wouldn't say this about a pocket calculator, even though it's true. Why say it here?
Would. ("Better" is legitimate for "more precisely".)
In fact, to an inquiry, not only «An AI can see and understand what it sees. It can identify and detect» can be read through usual duly interpretation of text as a typical imprecision for «An AI can see and "understand" what it sees: it can identify and detect» (the full stop stands for a colon), where rhetoric 'understand' is explained immediately after, but also the term "understand" is not that necessarily bound to Intelligence: it means "to have entered into a relation, to create a relation" (that 'under-' is a case of 'inter-'), hence "to approach" - "understand" is generally legitimate for progressive (for only progressive) definition of a concept. The use of "understand" for Intelligence must be some elision of "properly, duly understand", proportionally to what is achievable to a human. Which also implies that a more limited entity can "understand" to the best of its nature.
This noted: little problem as per the «barcode», since «it's still very clear».
Surely, if _some_ «people seem to intentionally blur this line», we can censor them.
But surely again, the formulation «whereas AI people seem to intentionally blur this line» is just plain offensive, and the addition «to get headlines and attention» is well over offensive. You have to add the quantifier, "whereas _some_ AI people" - otherwise typicality (beyond statistics: qualification) is implied.
The sales/pr guy wasn't even sure why or not willing to disclose the why. I wonder if it had to do with unions. However that is total speculations based on an experience with gov't and unions approximately 20 years ago where the unionw was very concerned with computer and software I had written taking over union jobs.
I think it will just be another tool available and we'll slowly see it creep into places without making earth shattering improvements. An example that comes to mind is IDE auto-complete that GitHub and others are doing - better than what we had before, but not exactly revolutionary
Apart from within academia and niche places like deepmind, I get the feeling that ML roles and positions will become the new DBA roles of the past - ultimately super-dull jobs where you are just shuffling around storage space, migrating data, or provisioning new tables/models etc without actually adding any business value or getting involved in any of the user-facing work etc.
It is already starting to get heavily commoditized with downloadable pretrained models etc... I see a lot of interns etc who want to come solve cancer with machine learning. I hope that there are not too many youngsters better their career on making it big in machine learning (unless as a researcher) because I genuinely feel like machine learning will just be some library/black box that 99.9% of the time will just be a downloadable pre-trained model that you add like you would if you needed to add OCR support to your product.
I am sure there will be some exciting new things that we'll see of course, but I think they'll be step-wise improvements, rather than huge leaps that open up entirely new worlds of opportunities. E.g. CNNs made image recognition much better (and anyone can now download a really powerful pre-trained model and beat state-of-the-art from just a few years prior etc), but it only made image recognition better - we could do it before, just not as well.
- Uber/Google Maps/Waze: ETA prediction
- Gmail: Smart Compose & spam filtering
- Instagram/Snapchat/Any camera app: Computer vision
- Siri/Google Assistant: Speech-to-text
- FaceID: Facial recognition
- Facebook/Netflix/All content aggregators: Recommendation engines
- Any banking app: Fraud detection
ML's use is extremely widespread at this point. The above list is just a tiny snapshot. "AI" is term thrown around by marketers and hypemen all the time, no arguments there, but ML's usage is anything but niche these days.
This article seems to be an exception to the rule
Probably not considered AI these days.
- Expert system / Rule-based system (CLIPS https://clipsrules.net/)
- Knowledge-based system
- Rational agent planning that doesn't involve ML (search, heuristics, MDP, POMDP)
- NLP that doesn't involve ML (Markov model, etc)
If you Google AI taxonomy you can get a good understanding of it.
Most so called algorithms taught in universities are after all some form of basic AI, there's very little conceptual difference between an A* planner and a minimax tree, yet one of these is essentially Stockfish, which is "smarter" than all humans when it comes to chess.
- Automated Theorem Proving
- Satisfiability (SAT) and constraint solving
I believe things come and go in cycles, so we might see the rise of non-ML solutions if the current circumstances change. Just like the last time: https://en.wikipedia.org/wiki/AI_winter.
Stuff like this: https://en.m.wikipedia.org/wiki/A*_search_algorithm
Such algorithms are still a viable approach in many situations (and I'm sure A* is used in production somewhere today).
Random forest can predict tabular data extremely well without overfitting, A* is the mathematical best approach to finding the shortest path through a graph, alfa-beta pruning with heuristics consistently beats any known neural net in chess, quicksort will be better than a neural sort because you don't have to first offload data onto a damn gpu, etc.
Even Tesla doesn't give the car's accelerator and steering reigns to the detection neural net outputs (that would be actually insane), instead it lets them extract data from the world into its vector space and then uses deterministic motion planning to determine where the car should drive based on A* GPS routing. That way the system won't do anything undefined and stupid, like neural nets are usually prone to.
DeepMind have downplayed the use of MCTS in their Alpha-x family, to the point of obfuscating the fact that is part of their system at all and have sowed much confusion about this, but their systems ain't going nowhere without good, old-fashion game-tree search.
Stockfish only recently adopted neural nets, btw.
Eg Wikipedia: "Reinforcement learning (RL) is an area of machine learning " https://en.m.wikipedia.org/wiki/Reinforcement_learning
In the spirit of the cutting edge, any chance you could give me a chain-of-reasoning on that inference?
It all depends on whether you consider the new use as a particular application of a more general thing or as a thing on its own. (But I agree that if you call it with than name it’s not that general.)
It seems we agree with Wikipedia that ML contains RL?
It's true RL is also studied in other fields.
I struggle to see how that means that RL is a good answer to "What are some AI algorithms that are still in use, that is not Machine Learning".
It may be the case if you define ML broadly enough. One may also define RL broadly to refer to things that existed well before ML was a thing (not that I would do it, but one may). I guess that may still be within the AI umbrella, but I’m not sure.
For a bit of history on machine learning I recommend Rodney Brooks' seminal series of articles on machine learning, beginnign here:
https://rodneybrooks.com/forai-machine-learning-explained/
The first article in the series, linked above has one section titled "Machine Learning Started with Games". In that section he goes over Arthur Samuel's checkers-playing program that beat a human champion in 1961.
The section also contains Brooks' description of Donald Michie's MENACE, which is widely considered to be one of the first reinforcement learning algorithms. For lack of a computer, it was implemented on a set of match boxes:
In 1960 Surgical Science did not have much pull in getting access to a digital computer. So Donald Michie himself built a machine that could learn to play the game of tic-tac-toe (Noughts and Crosses in British English) from 304 matchboxes, small rectangular boxes which were the containers for matches, and which had an outer cover and a sliding inner box to hold the matches. He put a label on one end of each of these sliding boxes, and carefully filled them with precise numbers of colored beads. With the help of a human operator, mindlessly following some simple rules, he had a machine that could not only play tic-tac-toe but could learn to get better at it.
The answer to the latter question is obviously “no”. Did anyone argue otherwise? mdp2021 suggested that redytedy may have meant that but what he or she actually wrote is “You can USE ML for RL, but the field itself is considered separate from ML and under AI in general.”
mdp2021: “The poster probably meant: Reinforcement Learning is not restricted to Neural Networks.”
feral: “Hmm, I don't see that.” [that RLinrtNN is probably what redytedy meant]
“In the spirit of the cutting edge, any chance you could give me a chain-of-reasoning on that inference?” [the inference that redytedy probably meant RLinrtNN]
I could be wrong.
However, I didn't think of reinforcement learning when compiling that list, because in my experience non-ML RL solutions are rarely better than ML solutions. Happy to be corrected on that front.
We subscribe to the gratitude.
https://drive.google.com/file/d/1GReNZ9fbI4203esUwGzqqLVg1El...