Super Alphago

[Bran Bast]

LinkedInand 3rd parties use essential and non-essential cookies to provide, secure, analyze and improve our Services, and to show you relevant ads (including professional and job ads) on and off LinkedIn. Learn more in our Cookie Policy.

I just went back to my normal routine after an amazing week at AMLD 2020. The machine learning community is very vibrant right now especially here in Lausanne and there are several amazing opportunities out there both in terms of new exciting products as well as new unbelievable technologies.

The successful adoption of machine learning and ultimately the survival of this relatively new market depends not only on skilled scientists and engineers to support the demand with novel solutions but also on leaders and decision makers to drive its adoption, on politician to understand and support the inevitable changes in society and HR to invest on the right talents. For all of this to succeed it is paramount to be on the same page with the terminology. Coming from a background of NLU and having spent the past five years in teaching a machine how to understand text I cannot stress enough the importance of having a clear lexicon.

Data science. Data science is the act, profession or discipline of applying scientific methods to extract or generate information or knowledge from data. Data science is often used as an hypernym from as simply plotting a graph in a spreadsheet to training the most advanced machine learning model. In my opinion "working in data science" and in particular "being a data scientist" are expressions that should be used very carefully because they are very generic and they define the whole field.

Machine learning. Machine learning is the application of data science to design a computer system capable of learning how to solve a problem from data instead of being explicitly programmed to do so. I think we should make a point of not calling "machine learning engineer" all those engineers and data scientist which are machine learning users, e.g. those which deploy or use models without actually training them. We should also make a further distinction between those machine learning engineer which are feeding new data to well known models and those which are actually inventing new models. For the lack of a better terminology we should really start professionally differentiate between machine learning users, machine learning trainers (or ML ops) and machine learning inventors.

Artificial Intelligence. Artificial intelligence can be vaguely defined as the use of machine learning towards the development of computer systems able to perform tasks normally requiring human intelligence (or up to human level). This is also known as Narrow AI. Although in our conversational use of the language ML and AI are almost synonyms if we follow this and the above definitions I would argue that not all data science is machine learning, and that not all machine learning is artificial intelligence. In fact it would be better to talk about simulated intelligence and to strictly define as such the application of machine learning to replicate human intelligence by learning human decision from historical data. The latter definition also underlines the inability of simulated intelligence to cope with scenarios not covered by the training data. Please note that here I am intentionally drawing the line between simulated intelligence and what I consider being more advanced solutions.

Augmented Intelligence. Augmented intelligence is the application of machine learning to develop assistive technologies designed to enhance human capabilities. This is one of my favorite concepts and I love the idea of machine learning designed to improve the capability of the human race rather than to entirely replace them.

Adaptive Artificial Intelligence. I consider adaptive artificial intelligence the application of machine learning to design systems that can self learn and eventually adapt to new situations without the need of existing data. Computer systems trained with reinforcement learning are a good example of systems that will adapt to new scenarios.

Cognitive Artificial Intelligence. A subset of simulated intelligence solutions can be defined cognitive if they can learn or be taught to imitate or apply cognitive human behaviors. Cognitive behaviors, for the lack of a better definition, is the idea of intentionally either failing or sub optimally achieving a goal in order to satisfy or achieve alternative irrational or emotional objectives. These can be derived from all sorts of human cultural bias including the need of super-seeding predictions with policies or constraints that cannot be violated. Imagine an AI playing chess which would change its objective according to its opponent and if it is playing with a child and that would intentionally let him/her win. Or imagine a navigation system learning that the user prefers to drive using a certain route (just because) rather than taking the shortest or fastest.

Super Intelligence. Super Intelligence is the idea of using AI to train new better AIs ideally without the use of human labelled data. For instance AlphaGo Zero bested AlphaGo because by being trained without human data, AlphaGo Zero, was not biased toward human mistakes. AlphaGo Zero however is based on reward functions and models which are still designed by humans. Maybe in future there will be a AlphaGo Super which will learn it's own reward functions and therefore be able to best any human defined reward function.

Artificial General Intelligence. AGI is the idea of using machine learning to train an AI capable of being self aware. By being self conscious an AGI should then be able to define its own purpose and choose its own path. This obviously poses all the kinds of philosophical and theological questions. The concept of AGI and super intelligence overlap and go along very well. Imagine an AGI being able to self train itself every time ti wants a new skill. Interesting right ;)

We have been talking about building an Artificial General Intelligence agent, or even a Super Intelligence agent. How are we going to get there? How are we going get to ECW and SLP? What do researchers need to work on now?

I think we need to work on architectures of intelligent beings, whether they live in the real world or in cyber space. And I think that we need to work on structured modules that will give the base compositional capabilities, ground everything in perception and action in the world, have useful spatial representations and manipulations, provide enough ability to react to the world on short time scales, and to adequately handle ambiguity across all these domains.

Currently all AI systems operate within some sort of structure, but it is not the structure of something with ongoing existence. They operate as transactional programs that people run when they want something.

So the very first thing we need is programs, whether they are embodied or not, that can take care of their own needs, understand the world in which they live (be it the cloud or the physical world) and ensure their ongoing existence. A Roomba does a little of this, finding its recharger when it is low on power, indicating to humans that it needs its dust bin emptied, and asking for help when it gets stuck. That is hardly the level of self sufficiency we need for ECW, but it is an indication of the sort of thing I mean.

The seven examples I gave, in Part III, of things which are currently hard for Artificial Intelligence, are all good starting points. But they were just seven that I chose for illustrative purposes. There are a number of people who have been thinking about the issue, and they have come up with their own considered lists.

And for the hard core learning festishists here is a question to ask them. Would they prefer that their payroll department, their mortgage provider, or the Internal Revenue Service (the US income tax authority) use an Excel spreadsheet to calculate financial matters for them, or would they trust these parts of their lives to a trained Deep Learning network that had seen millions of examples of spreadsheets and encoded all that learning in weights in a network? You know what they are going to answer. When it comes to such a crunch even they will admit that learning from examples is not necessarily the best approach.

Others will have different explicit lists, but as long as people are working on innate modules that can be combined within a structure of some entity with an ongoing existence and its own ongoing projects, that can be combined within a system that perceives and acts on its world, and that can be combined within a system that is doing something real rather than a toy online demonstration, then progress will be being made.

And note, we have totally managed to avoid the question of consciousness. Whether either ECW or SLP need to conscious in any way at all, is, I think, an open question. And it will remain so as long as we have no understanding at all of consciousness. And we have none!

Meanwhile real researchers were competing in DARPA competitions such as the Grand Challenge, Urban Grand Challenge (which lead directly to all the current work on self driving cars), and the Robot Challenge.

I have never been a great fan of competitions for research domains as I have always felt that it leads to group think, and a lot of effort going into gaming the rules. And, I think that specific stated goals can lead to competitions being formed, even when none may have been intended, as in the case of the Turing Test.

If we are going to make real progress towards super, or just every day general, Artificial Intelligence then I think it is imperative that we concentrate on general competence in areas rather than flashy hype bait worthy performances.

2 year old Object Recognition competence. A two year old already has color constancy, and can describe things by at least a few color words. But much more than this they can handle object classes, mapping what they see visually to function.

3a8082e126