Questions, ChatGPT’s answers, my elaborations — I asked some questions. ChatGPT gave reasonable high-level answers, which I copied verbatim below. I have highlighted, in bold, some words and phrases in ChatGPT’s responses. Ones that resonated with my experience. Generally speaking, ChatGPT’s answers set the stage for me to expand or clarify further. Which I did, in…

In (un)supervised learning, language modeling, and search — In this post, I cover various metrics commonly used in data science. Having them all here makes it easy to compare and contrast and get a birds-eye view of how they — and the problems they track — relate to each other. These metrics are primarily used to evaluate models…

A case study in sentiment classification — Say we have a large corpus of English sentences and we want to train a basic sentiment classifier that inputs a sentence and predicts the sentiment in it, which is one of positive, negative, neutral, or conflicting. Say we want to minimize the effort we spend on getting a labeled…

How attention and word embeddings help with learnable local alignments — Consider a task in which one is shown a few sentences, maybe an entire paragraph or two, and then asked to answer a specific question. This task, which tests reading comprehension, is widely used in the GRE test. In this post, we will start with insightful examples of this task…

The simple and powerful DDSketch — Say there is a web service and we’d like to track its response times suitably summarized over all the requests served in any one day. The average or median response time is not very useful. For instance, say the P50 — i.e., the median response time — is 100 ms…

A statistical NLP approach based on supervised learning — In a previous post on Medium, we discussed segmenting text into sentences [3]. Now we look at a related problem: segmenting text into paragraphs. At first glance, it may seem that the two problems are essentially the same, only at different levels of chunking. …

Streamability, mergeability, and how they relate to usual statistical models — Data Sketches: Definition And Distinguishing Features A data sketch is a compact summary of certain aspects of the data optimized for answering certain queries approximately. (We’ll elaborate on the italicized words soon.) Data sketches tend to be nonparametric — i.e. they don’t assume the data is drawn from a parametrized…

With concepts and examples in question-answering tasks — Well, it’s hard to argue against paying attention to the right things. Unfortunately, language modeling algorithms were not doing so for a long time. This all changed about half a decade back. The rest is history. As witnessed most recently with ChatGPT. In this post, I aim to provide a…

Examples, ML formulation, and brainstorming features — For quite a few years, I have been impressed with the suggestions offered by the smart replies feature in Gmail. How did they do it? In this post, we will brainstorm a first attempt at sketching out how such a feature might work under the hood. We will start with…