Good Post.

Machine Learning Theory, I think, in its most renown version is now just applied statistics to inference hidden states from either complete or incomplete observations (Markov chains stuff + Expectation-Maximization Algorithms) or more complicated patterns, i.e. speech recognition, nested maps of related words and so on, by means of artificial neural networks.

I think one significant point here is that of engineering: whether it is worth solving an cognitive-like problem with AI or with an 'easier' network solution: for instance the google self-driving car: its solution came up from google maps, the car simply retrieve in advance the info from a online database aboút what lies before it.

I see what you mean; however the information retrieved from Google would be historic (street layout, gross map, etc) and therefore still require real-time sensory information and on-line decision making for the vehicle (to deal with new objects on the street, other vehicles, etc). Sebastian Thrun has shown with his efforts at Stanford in the DARPA Grand Challenge that AI and ML play an incredibly important roe in the development of robotic response models that operate in real-time to solve these problems.

Some of such an info is on-live like current traffic flow, weather features, road conditions and so on, already available in google-map premium. But I got your point, I think that addressing properly the issue, it is a matter of dealing with the old divide/conquer approach. Following with self-driving car example: some parts of the total "cognitive" system will be on the wireless "cloud" ( from servers, other cars, wireless sensor networks, retrieving Dijkstra solutions for optimal routing ...) so that really let machines catch on the fly most of the info required and translate it into specific actions. Those features not able to be retrieved from the cloud, ought to be implemented in terms of standalone AI.

But one remark here: machine learning as you know ( i am also thinking in neural networks) is very limited to the size and type of sampled items given during the training stage. That poses a serious limitation when the range of inference outcome lags far behind the requirements of the tasks, like those faced by a self-driving car: that is the why the cloud + AI together solution seems to me an easier and more feasible way to deal with such a computationally complex problem than AI alone.

Yes .. I agree. A combined engineering & AI solution is normally the most practical approach to an applied problem. A remark on your remark - the LML paradigm insists on continuous learning and not a train once and use forever approach.

Thanks Wai-keung .. Most insightful. The ability to learn how to learn has been widely considered a key step from the narrow AI of today and the Big AI of tomorrow. This certainly implies the retention (an I propose consolidation of learned knowledge (see the attached)

Oops forgot the attachment.

Good review on LML!

You came up in your excellent review with the key point, the relationship between brain computation and IA. In the human Amygdala for instance, it has been tested 3 different learning models for predicting a reinforcement outcome given some random associated cues.

Models based on previous expectations (agent has already built a sort of map consisting in probabilistic relationship among variables before being exposed to a real scenario) are significant more predictable than the classic delta-error model, based on pure associative learning where agent is surprised by the non-expected events . That means that in human-brain a LML might consist in updating hypotheses arranged into probabilistic maps every time agent is fetching new observations. So as far human amygdala is concerned, a way to figure out how amygdala can learn to learn is by building and updating new expectation maps based on previous ones for different scenarios. So in that way, the first map is built stochastically (i.e. following an E-M algorithm ) , but the second one is built faster taking as pattern the previous one and so on.

http://www.ploscompbiol.org/article/info%3Adoi%2F10.1371%2Fjournal.pcbi.1002918

I think the point by Wai Keung hits the hammer on the nail when it comes to machine learning. Our brains have the ability to sense a reality, store it and then convert it to a paradigm to apply it to a new situation. The complexity arises when one reality possesses more than one paradigm (as it should). The brain seamlessly selects the most suitable paradigm from the retrieved reality. Applied to machines, how do we get the machine to "sense" a different paradigm...only by explicitly giving it to the machine. Then the machine needs to determine that the new paradigm is more suitable to be used in the specific scenario and then apply it. The analogy is that of a picture scene with different objects, animate, inanimate, colour, event etc. How do we teach the machine to recognise based on preconditioned criteria? This is like an inspector at a crime scene and is looking for clues. Whilst I understand that the examples are fairly complex, I would be intrigued with algorithms that enable machines to be selective in their learning paradigms, and update those learning paradigms.

Dr AbdulBasit

An important example pressing the need for Lifelong machine learning is that of a robot for household. We cannot train it for everything before hand as called Narrow IA earlier. This is because we can't assume the vocabulary, needs, actions and communication style of the people in that house. So in such a case, LML based approach as Daniel called it Big AI will help each robot adjust to the needs and liking of its housemates. However, we have to accept that automatic learning is prune to errors more than traditional approach. So there should be some ground truth provided to such a robot so that it will continue to learn new things without hurting itself or anyone else. A ground truth would be like a frame to which every rule should fit. For example the NELL model is developed with a semi-supervised approach that needs manual corrections from time to time. With a ground truth provided ones, may be the multiple time manual effort could be saved.

Taimoor .. Thanks for your comment. A Household robot is a great example of how an Intelligent Agent will need to learn over a lifetime. There are many nuances from home to home that could not be dealt with in any other way. And I agree that such agents will need to come primed with learned knowledge from a practical perspective so that it arrives as 5 year old ready for nuanced learning versus a baby that must learn everything. .. Danny