Basics

Lets start with some of the basics. What is the simplest form of machine learning ? If you consider machine learning as utilizing previous data to predict future outcomes then taking average is one of the simplest ways. For example, if you predict the todays temprature from the historical temprature data of same day you will get a reasonable estimate. Don't underestimate this the theories of law of large numbers and central limit is very powerful especially if you don't have additional variables to explain the data. However, often we have explonotary variables. In machine learning nomeclature we call them features. In this setting, each feature lies in euclidean vector space and a function maps these features to target variable. Now the task is finding that function. Of course, you can just construct lookup table for input output pairs from previous data. If you don't have much variation from cases in runtime vs previous data this will do the job. But the real problem of machine learing is generalization. Consider, the continous values in features will have infinitysmall probabilities to match. For categorical variables these likelihood decreases exponentially. After those trivial solutions we have simplest models as linear models. What does those linear models do ? They simply take weighted average of input features to match output value. There could be various functions at the end of weighted average to satisfy some desired properties. For example logistic regression uses sigmoid at the end to warp the output between 0 and 1 which is good for describing binary class probabilities. Lets go more formal with this model in the next posting.

Comments

Post a Comment

Popular posts from this blog

When to do machine learning

I realized that before starting on technical details of methods I should write some about when to use machine learning, or when it is intelligent to use artificial intelligence ? In fact, in most of the cases avoid machine learning as much as possible. In first place, you use machine learning when your measurements, sensor values, data is not sufficiently descriptive to get final result that you are looking for. For example, covid tests works with very high precision currently. If your input is the result of the test you don't need machine learning. However, if you are detecting covid from chest x-ray you have many small tiny pixel transparancy measurements thus you need some kind of learning. In this case, you should ask while the alternative way is providing good quality result do we really need this learning ? In summary you should try to get descriptive data as possible in the first place. Later, if you have only small amount of samples, the best you can do is use them as te...
Read more

Lets start coding for linear regression

As from previous post, linear regression has analytical solution. For a given dataset the optimization problem reduces to linear algebra psudeo-inverse problem.  ``` W = (XX')^-1XY``` This would be the training step for the algorithm. Python code would be very easy.  np.linalg.inv(X.dot(X.T)).dot(X).dot(Y) Later once the parameters W is defined runtime would be a single matrix multiplication as our linear model. Y' = WX  Could be written as  Y_est = W.dot(X)
Read more

Linear algebra

Machine learning became available after being able to store and process large amounts of data. To be able to process we need to convert that data into numerical form. Often times we have tabular structures which is converted into matrices. For higher dimensional problems we could have tensors or 4d tensors. Then, in order to process that information we need to rely on operations on matrices. Thus we will utilize linear algebra. The linear regression formulation we have mentioned uses weighted sum of input features. Let us remind that these features are the rows of the data table converted into numerical form. This weighted sum is inner product of two vectors could be written as : w = [w1, w2, ...] x = [x1, x2, ...] s = w1*x1 + w2*x2 ... The dot product of two vectors is a scalar. If vectors are normalized the result would be cosine similarity. Now, this is model of single neuron in neural networks. It gives single value output for single input vector. If we have multiple dimens...
Read more