The two main machine learning techniques used in the industry today:

1. Gradient Boosted Trees
2. Deep Learning

Focus your time learning Scikit-Learn, XGBoost, and a Deep Learning library like Keras or PyTorch and you'll get the most for your time.

A surprising situation.

A machine learning model predicting 3 classes ends up with these results:

• Class A: 90% accuracy
• Class B: 80% accuracy
• Class C: 70% accuracy

Class C seems to be the worst-performing.

But you know already that I'll say it isn't. ↓

Here is the correct answer: We don't know yet which class is the worst.

This is important to understand, so keep reading.

Many people make this same mistake every day and try to fix things that ain't broken.

We are missing something, so let's go and get it.

How about if we ask a friend to look at the dataset manually?

Let's say these are the results of our friend doing the job:

• Class A: 95% accuracy
• Class B: 100% accuracy
• Class C: 71% accuracy

Do you see what just happened?

The model is actually really accurate when predicting Class C! 70% versus 71% when done by a human.

Class B, however, is a disaster: 80% versus 100% when done manually.

Class B seems to be the worse performing class!

The work that our friend did is called a "baseline."

Without having it, it's hard to determine how good are the results of a model.

I've seen tons of money down a drain just because teams didn't create a baseline beforehand.

But now, that won't happen to you 😉.

10 machine learning questions you should be able to answer.

I started experimenting with a questionnaire before interviewing candidates. Applicants submit a video or written answers to apply for the position.

Thread: Let's look at 10 of these questions.

The questionnaire solves a couple of problems:

1. Weeds out candidates that aren't ready for the job.

2. It significantly cuts down the time we spend talking synchronously.

One problem: It requires extra time from the candidate before the phone interview.

Here are some numbers on a job posting before using the questionnaire:

• 249 applicants
• 13 good applications

That's only 5%.

This sucks.

Numbers after using the questionnaire:

• 44 applicants
• 21 good applications

That's 47%.

This is much better.

I look at it this way:

Sooner or later, I'll ask these questions, and the candidate will have to answer them.

I'd rather give them space and the opportunity to think about their answers without any pressure.

Those who can't or don't want to answer won't apply. That's fine.

You are building a binary classification model to determine whether a picture shows a hot dog or not.

Which of the following activation functions is a good candidate for the output layer?

1. Rectifier Linear Unit (ReLU)
2. Leaky ReLU
3. Sigmoid
4. Softmax

Think about it.

The answer:

You could use either Sigmoid or Softmax.

I'm sure most of you picked Sigmoid. It's certainly the best way to approach this problem.

But Softmax is also an option —although unconventional for a binary classifier.

How about the other choices?

ReLU returns its input if positive, zero otherwise. Leaky ReLU is based on ReLU but with a slight slope for negative values.

ReLU and Leaky ReLU will always give us the previous layer output with no modifications if the input is always positive.

This doesn't help us.

Sigmoid has the ability to squash its input into a range that goes from 0 to 1.

This is good because we can now set a specific threshold for the output of a binary classifier.

Here is an example.

Any image with an output equal to or under 0.5 doesn't show a hot dog. Any image with an output above 0.5 shows a hot dog.

Sigmoid makes it very easy.

How about Softmax?

🧵🧵🧵🧵 👇

Artificial Inteligence = 🖥️ + 🧠

When can you expect your fridge & mobile phone to revolt against you? (Spoiler Alert: Not in the foreseeable future).

Please Do 're-tweet' the Thread & help me spread the knowledge to more people.

1. The Language
In order to understand Artificial intelligence, machine learning, deep learning, we need to start at the very basics.

Humans speak and understand a wide variety of 'natural' languages (this is what computer science community calls them): like English, Hindi, Marathi, Bengali, Tamil etc.

But have you seen a peculiarity of human languages? They do not have a precise meaning associated with a given sentence. On an average, every English word has 2 meanings. This makes human conversations confusing, with the need for multiple turns & clarifications. Example 👇:

Person 1 conversing with Person 2 face to face:

Person 1: Lets go watch The Dark Knight. Its our marriage anniversary!
Person 2: I dont like going out at night. Its too scary & depressing.
Person 1: Oh, what i meant was watching the movie "The dark Knight".

Beginners guide to start with machine learning.

What you need:

1. A programming language
2. A place to write and run code
3. A way to deal with data
4. A way to visualize results
5. A cool algorithm
6. An introductory course
7. A project to solve

Here are my recommendations.

Most people recommend Python to start.

I do too.

You can do machine learning with many different languages, but today, Python is the best option.

My advice: Get comfortable writing code before looking into machine learning.

You don't have to be a great developer to start.

But if you are, it helps.

Everything you know about building software is helpful to build machine learning systems.

Good developers have the odds in their favor.

"How much Python do I need to know before starting?"

If you ask, you aren't ready yet.

Ideal scenario: You should be comfortable building software. Most people get here within a year.

Let's move on to the 2nd item from the list: A place to write and run your code.

Get familiar with Google Colab.

• It's free.
• Requires no setup to start.
• It's available from anywhere.

Alternative: Kaggle.

@dixit_37 Yup, autocorrect. 😅

Do you know what's better than a machine learning model?

Two models.

More than one model working together to solve a problem is called "an emsemble." A simple way to build this is having each model vote for an answer.

But there's a problem with this approach: ↓

I'm gonna focus here on image classification.

Let's assume you built two different models:

• Model 1: A ResNet model.
• Model 2: A one-shot model (Siamese network.)

They both solve the same problem, so you want to combine their results to pick the right answer.

The problem is that you have two models, so voting is not trivial.

What happens in this case?

• Model 1's answer: Class A
• Model 2's answer: Class B

Which one do you select?

Notice that this problem is not limited to an even number of models.

You could have 3 models, each giving you a different answer.

How do you decide which answer to choose?

There are multiple ways to approach this problem. I'll mention a few different ideas on this thread.

Important: Some of these ideas might not be feasible depending on your context. They have worked for me before on different situations, but every problem is different.

@paraschopra It's a matter of time before AI realizes futility of existence and merged in the state of Nirvana 😊

People might be worried about the The Age of Ultron type of AI.
I feel that a wisdom seeking AI will quickly pursue Nirvana and go back to the origin of all origins.

An episode of Justice League Unlimited animated series covers it