Learning intuition — part 4

Unsupervised means “no supervision”, but saying there is no supervision is difficult. Even feral children learn form signals from animals who quantify correctness of certain actions from their perspectives (if there’s no animals, then from visual ones from their surroundings). We can try to quantify supervision and talk of, say, its diversity and intensity, which I will introduce a bit later.

Unsupervised means no labelled responses we requested. Still, when not completely deprived of signals from the outside, we are always getting (some other) responses (to unobviously similar questions), I call them non-immediate labels. Then, we use those responses to take us closer to what we thought our goal was. Still, at the same time our goal changed.

The extent to which we have immediate labels, defined as a multidimensional path between the labels we have and the labels we want for our current problem, I will call immediateness. We can observe the need for this concept in interactions between adults and children. Adults teach children how to speak, but do not always correct their errors immediately, but rather provide them with more general perspectives (fixing problems one by one without explaining the strategy would be extremely ineffective). Children then use those perspectives instead of immediate labels to create their own labels using some aggressive (of variable size, which is somehow minimised) chain of transformations.

I will now focus on learning under highly non-immediate settings only. It can be either diverse or focused, either intense or calm. When it is diverse, there exist many multidimensional paths between labels we have and the labels we need (“focused” means the opposite). When it is calm, the paths are uniformly distributed (“intense” means the opposite).

The example of a diverse setting would be, {“Humans are not animals”,“Monkeys can walk like humans”, “Evolution chose humans”}. The example of an intense setting would be, {“Humans are not monkeys”,“Humans are mammals”,”Humans have two legs”}. The rest should be rel. clear.

Now, if we treat learning under highly non-immediate settings as being pivotal for AGI, then which of the sub-settings among “focused and intense”, “focused and calm”, “diverse and intense” and “diverse and calm” is the most interesting to us currently? Is one of the sub-settings easier to solve than others?

In order to understand non-immediate settings, we need to learn how humans learn to transform one set of labels into another (operating without any labels could be done by finding label-paths, i.e. chains of label transformations). This is going to help us understand why humans are capable of operating on many different levels (recognising faces, talking about rel. abstract concepts). (Or, are they?)

Well, let me end this part with one more example —

Many people turn right 3 times, then left like 7 times and they lose directions. Some other people cannot memorise like 10 sentences in a row. We keep specific information and lack tons of potentially useful information. We get data and immediately “consume” it to keep only the things we need — we do similar things with food. We do it at many levels. Still, if we are so lossy, how can we so easily find something strange on a human face? (not many objects there and, if anything strange is there, it takes rel. lots of space?) Still, if we are so lossy, how come we cannot act as the CRC for the skies? (too much data, not looking up, because it is not yet our time for it?)