AI In Plain English
Go deeper with Artificial Intelligence, Machine Learning, Data Science, and Big Data
Image for post
Image for post
Thinking about logistic regression as a simple neural network gives an easier way to determine derivatives.

Deriving the softmax function, and cross-entropy loss, to get the general update rule for multiclass logistic regression.

This is continuing off of my Logistic Regression on CIFAR-10 article, so skim that for some details —especially the intuition on cross-entropy and softmax.

Here are a few details about the CIFAR-10 image dataset this article is being written in the context of.

I’d recommend a good knowledge of derivatives and simple neural network functions, like backpropagation for this part. Even if you do know those things, you won’t understand just from reading this. …

Image for post
Image for post
The first sample of the CIFAR-10 dataset, a 32x32 image of a frog. These are the R, G and B channels of that image.

Chapter I: Multiclass Logistic Regression

On PyTorch, data wrangling and analysis, softmax, categorical cross-entropy, mini-batch gradient descent, derivatives of the cost, and other concepts behind advanced logistic regression — why it doesn’t work.

Picking up a worn mechanical pencil from the floor, a half-used Staedtler eraser, a metal 6-inch ruler, a notebook from Japan, and my computer, at 1am on a school night, October 3rd, 2020, I began to teach myself calculus.

Okay, not teaching myself. There’s no such thing as teaching yourself, these days, with the plethora of content out there. …

Image for post
Image for post

This will be a very short article about the papers published on this topic.


A research paper by the authors, Wei-Han. Lee and Ruby B. Lee, from one of Princeton Universities Department introduces the idea of always keeping your smartphone safe by using a multi-sensors-based system that continuously authenticates the user using your smartphone’s sensors.

The system is able to perform this task by learning the owner’s behavior patterns and environment characteristics, and then authenticates the current user without interrupting the user’s interaction. This helps combat smartphone thefts from impersonation attacks.

The system would use a sensors that best reflects the user’s…

Image for post
Image for post

The main motive of any supervised machine learning algorithm is to reduce the model error. What is a model error? It is the difference between the true and the predicted values.

It consists of 3 parts — Bias Error, Variance Error, and Irreducible Error.

error in a machine learning model — bias, variance and tradeoff error
error in a machine learning model — bias, variance and tradeoff error

Irreducible Errors are errors which will occur irrespective of whichever model you are using. This error could have arisen from a missing input variable, measurement errors, etc.

In this blog post, I will be briefly discussing the first 2 parts — Bias and Variance Error.

Originally Published on My Website — Let’s Discuss

What is Bias?

Bias is the inability of the model to capture the true relation. A model with a high bias means that the model has low complexity. For example, Linear Regression is an example of a model with High Bias. …

Image for post
Image for post
Photo by Brian Erickson on Unsplash

Sand vs Shale Case Study based on Well Log Data

“In today’s analytics world, building machine learning models has become relatively easy (thanks to more robust and flexible tools and algorithms), but still the fundamental concepts are very confusing. One of such concepts is Hypothesis Testing.”

Essentially, hypothesis testing is a statistical method that performs the test of an assumption, so that the results can be declared accepted or rejected. Hypothesis testing is part of inferential statistics.

Hypothesis can be divided into two parts:


In this article, we use facies data which is derived from Well Log Data. …

In-depth walk through the complete mT5 research paper

Image for post
Image for post
Let's cook! Photo by Dan Gold on Unsplash

Hey Guys, Google recently released a new Transformer model named mT5 which is the multilingual version of T5 (both based on seq-2-seq architecture). If you recall T5, it is capable of performing almost any given text to text task in English. It was more of a One-for-all model in English(i.e. is one model for most NLP task). mT5 is a multilingual variant of T5 to with the capability of doing NLP tasks in more than 100 languages!😇

Image for post
Image for post
Photo by Jack Dong on Unsplash

If you haven’t read the T5 paper, check out my blog on T5. I have clearly explained all the details you need to know about the paper, and you will also have a much easier time understanding this blogpost. …

Eliminating bias with artificial intelligence

Image for post
Image for post

Engineering is no longer a man’s world. When we look at leadership positions, there are more female leaders today than there were 20 years ago. And yet, we still see too many female engineers choose a different career path after facing too many difficulties along the way.

Root causes inequality

Fortunately, not everyone is willing to accept this reality. Recently I had an hour-long conversation with an ambitious female engineer about what’s causing these women to leave the profession they’ve worked hard for and struggled to build. While I enjoyed speaking to her and value all the work she’s doing, it also left me feeling uncomfortable. …

Find out what is TFRecord and how to create TFRecord files to train a deep learning model.

In this post, you will learn the basics of TFRecord, benefits of using TFRecord. How to create a TFRecord file for an image dataset to train a deep learning model.

Image for post
Image for post
Photo by Anthony Martino on Unsplash

What is TFRecord?

TFRecord format stores structured data in a simple protocol buffer message format as a sequence of binary records for effiecient serialization

TFRecord uses tf.train.Example to create the protocol buffer(protobuf) message format that is represented by {“string”: value} where value is generated using tf.train.Feature.

Representation then is {“string”: tf.train.Feature}

tf.train.Feature accepts three different message type

Image for post
Image for post
Images, text, audio are classified as unstructured data

Natural language processing is the process of building machine learning models that can understand text or speech and perform desired tasks.

Text or speech data is unstructured i.e. by seeing the data or numbers behind text or image you can not make sense of it. You need context, meaning and a lot of other things to make sense of a sentence or paragraph or even a word.

Otherwise how would you differ between bat and a bat (bird) or weak and week (in speech).

This makes natural language processing for machine learning more difficult than doing for numbers or dates or similar structured data. …

Planning actions by simulating the Acting Engine

Image for post
Image for post


In the previous post, we discussed the RAE (Refinement Acting Engine) that plans the actions by refining tasks into executable actions and does that with the observed world state instead of the predicted world state. For the full article, please read here:

One thing that you may have noticed in that approach is the lack of planning in advance.

In some scenarios, planning in advance may give us a more optimal solution to our problem.

By planning, we can explore different courses of action and choose a good solution.

Sequential Refinement Planning Engine

SeRPE is based on RAE and it only supports a single task at a time. …

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store