ALEX KAY BLOG

MODEL CALIBRATION

Stay on the side of real probability, May 21, 2022

Model calibration is a model postprocessing operation dedicated to improving probability estimates. Calibration helps researchers to interpret results better, provide correct error analysis, and confidently apply models in the hight risks applications.

OBJECTTIVE, COST & LOSS FUNCTIONS

Understand what you optimize, May 1, 2022

People used to talk about loss functions and how to optimize them and are concerned about their stability and relevance for a task. But living in an era of ML we must know that datasets containing thousands of examples needed to be aggregated somehow. Some tasks specific losses might help us to provide the knowledge about the domain to the model more efficiently, and discussing these ideas with our colleagues we must utilize the exact terms to be clear.

KNOWLEDGE GRAPHS

Real world entities networking, March 12, 2022

Math abstraction, graph-structured data model or topology to intergrate data. Knowledge graphs (KGs) are used to store interlinked entities. Relationships also can be described using KG abstraction and be used for calculating relationship among the objects. As a consequence KG allow to build mathematical apparatus describing options which can be used in particular purposes.

METRIC SPACE

We have entities represented as a vector or sequence, but do we know how to compare them? February 27, 2022

We might have entities described by N-dimensional vectors so we can apply some other math abstractions effectively, but do we really know that we have only one way to calculate distances among the vectors? We are able to expand our assumptions on comparing operations to find differences and commons more effectively and even maybe faster!

METAPHYSIC (book V) TERMS

The basis of thinking, February 22, 2022

One of the major works of Aristotle and in the history of philosophy is "Metaphysics". This is a compilation and contains 14 books. Each of them might be read independently from each other and discussed separately. I find all of them interesting, but in the fifth book (V) Aristotle introduces main terms that describe the substance of knowledge. All of them with definitions are collected here.

WORD EMBEDDINGS

Distributed semantic, February 19, 2022

Embeddings as a vector representation allow describing objects in a continuous N-dimensional vector space. It is the most common method for building the systems which must to calculate relations among the objects (ex: tokens in NLP) using their vector representation.