Learning

     Learning is an optional feature in an intelligent system. An intelligent system that knows everything need not learn.

But the amount of knowledge in this universe is exponentially more than atoms. So no systems can have a marked answer to a marked question for every knowledge.

Learning is the ability of a system to change its configuration. It can change its algorithm.


    There are other metrics like effeciency of learning, its precision and accuracy. But that is for later on.


   Why is learning important in a system? It allows the system to pick up knowledge from its immediate surroundings, ensuring that it knows what is necessary.


Let's speak in biology terms:

    Evolution of learning:  There are multiple species in an ecosystem. One species feasts on the other, one species competes with the others for the same resources. Now, it is a constant arm's race between species thanks to evolution. Over the evolution period,  a structure evolved that allowed an organic species to be able to dynamically change its response to the surrounding's stimuli. No more, did a species have to wait for hundreds of generations to adapt its behaviour to a surrounding. Environment changes, ability to learn ensures an organism keeps up with the change.


    A neural network: A simple neural network is a line of neurons with an input neuron, and an output neuron that is connected to a muscle (a way to interact with the surroundings)

Fig 1.0: A basic neural network

    Learning involves: the change in the properties of connection between 2 or more neurons in the network. This is all basic knowledge, but what has me stumped is HOW do each of the neurons change connections to achive a desired result?

How does a dog learn tricks? How does a human learn to walk?
by practice. Each practice makes certain connections stronger and newer connections are formed. 
However, it seems that the intelligent beings also use feedbacks after a practice and somehow figure out what to adjust. And all this happens very fast.

Fig 1.1: a graph of accuracy vs time


This is the famous learning curve.

    So, do the neurons try changing each and every connections and keep the configuration that works out? Theoretically, that will be absolutely ineffecient. In higher intelligent organisms, not all neurons are tested. Only a select few neurons are tested using trial and error. 
    Basically, this points towards another learning algorithm layer on top of the basic neural network that governs the process of learning. Interestingly, the learning algorithm also learns ...to be continued







Comments

Post a Comment

Popular posts from this blog

What if you had infinite budget but only 1970s rocket tech to colonize space?

With an infinite budget and 1970s rocket technology, we could achieve some extremely ambitious space projects, even without modern advancements. Here’s what would be possible: 1. Mars Colonization (Apollo-Style) Using Saturn V or equivalent, we could launch multiple heavy-lift rockets to send habitats, supplies, and people to Mars. Without modern efficiency improvements, the mission would require tons of fuel and staging , but brute force can compensate for technological limitations. Nuclear thermal rockets (already theorized in the '70s) could reduce travel time. 2. Moon Base (Extensive, Permanent) A massive lunar base is very doable by continuously launching Saturn V-sized rockets with modules, rovers, and supplies. The Apollo program already demonstrated crewed lunar landings—scaling that up is just a matter of repetition and logistics . Nuclear power and regenerative life support (already conceptualized in the '70s) could sustain a base long-term. 3. Space Stations Bigg...
Read more

The paradox of the white screen

 I was thrown into a white room. Nothing was there, except a white screen. The white screen displayed a mysterious series of text. A sentence of an unknown language . I recorded them in my memory, what text followed what, and came up with a rulebook of the rules the initial sample followed. (even what correlation the x-th text had to a-th text.) (and by rules, I mean all POSSIBLE correlations were recorded) Suddenly, the screen showed another series of text. I wrote a separate set of rules, and tried to match the rules of the former and the latter. In the process, a huge amount of rules were discarded .  The discarded rules I recorded in another memory bank. The screen kept changing, each time the series of text changed. I got the grasp of what text might come next. The accurate rules began to emerge. Many rules happened 99% of the time.   I became so confident of the rules that I began to spot mistakes in some samples . But even if I became so good at predicting th...
Read more