The Master Algorithm is about 5 different machine learning algorithms, and how these may be harmonized into a single “master algorithm.” A knowledge of machine learning will help you to ride the AI wave, and make smarter decisions.
#1 Induction
Induction is the application of logic backwards, where we use a conclusion and a specific premise to form a general premise. This means that if you’re a machine, and the data tells you that Jane is both a doctor and university educated, you then generalize that all doctors are university educated. This approach of generalizing may sound risky, but we can normally get around this problem by providing our machine learner with big data, so exceptional cases can surface.
#2 Neural networks
We can think of a regression as a single neuron, where input data is activated by a slope to produce output predictions. Adding multiple inputs and hidden layers will give us a deep neural network. The algorithm learns through gradient descent and backprop.
Gradient descent: iteratively update all weights based on cost function’s gradient and algorithm’s learning rate hyper parameter. Note when I say ‘slope’ in this video, the more technically correct term is ‘weight’ and is NOT ‘gradient’ here. The cost function’s is the difference between the real and predicted output across all available data, so its gradient is its partial differential. Potentially runs into the problem of stopping at local minimum instead of global minimum, however practically this is not a problem if learning is done in a high dimensional space.
Backprop = Backward propagation: there are 4 hidden layers, gradient descent is done for layer 4, then 3, 2, 1. Must be done backward as gradient descent for layer 3 requires gradients for layer 4. Once all gradients are learnt from back to front, the algorithm forward propagates all data inputs through the neural network to find the predicted output and new cost function value, which it uses to backprop the next iteration of gradient descent till cost function is minimized.
#3 Evolutionary fitness
Here, gene pool composition is knowledge, which is learnt through selection for evolutionary fitness. You can think of this as 'caviar learning.' Fish lay thousands of eggs, and only a small fraction of the fittest survives to adulthood. Similarly, you can compete a multitude of different neural network structures against one another, and then select those that perform the best. To find out how the beautiful Gouldian finch came to be, check out: • Why Beauty beats Beef (Book: Evolution of ...
#4 Bayes Theorem
An example of an algo that uses Bayes theorem is Naïve Bayes. This estimates the subjective degree of belief of the cause, given the observed effects, making the naive assumption that the effects are independent given the cause. I find Bayes theorem interesting because it gives us an elegant way to update our beliefs based on new information. Notably, it’s a simple formula that solves a blind spot in our intuition.
#5 Analogies
This involves learning through associations. A good example is word embeddings by comparing word vectors in a hyperdimensional space of meanings. More broadly, relational learning is what this channel is really about, helping you to take knowledge from one field and apply it to another field when you spot a relationship. In my view, this is one of the most powerful ways to innovate.
Photo credits
By User:BBODO on english wikipedia - English wikipedia site: Image:060622 Woodpecker Finch
composite.jpg, CC BY-SA 2.5, https://commons.wikimedia.org/w/index...
By Peter Wilton - Large ground finchUploaded by Magnus Manske, CC BY 2.0, https://commons.wikimedia.org/w/index...
Информация по комментариям в разработке