Softmax is one of the most commonly used functions in machine learning.
It is used to transform high-level features into probabilities. Based on the formula, it is hard to imagine how it is done exactly.
Softmax might not be what you think it is. Let's find out why!
🧵 👇🏽
First, we start with the exponential function eˣ, which transforms a real number into a positive one.
It has a feature that shows the geometry of this transformation: it turns addition into multiplication.
In particular, eᵃ ⁺ ᵇ = eᵃ eᵇ holds.
The input x = (x₁, x₂, ..., xₙ) consists of the highest level features: the class scores.
For two vectors x and y, xᵢ - yᵢ expresses the difference between features.
After the exponential function, this is transformed into their ratio.
To turn these transformed features into probabilities, we have to normalize them so that the sum is one.
This is why we divide with the sum of all exponentials in the definition of Softmax.
Although Softmax is very popular, there is a big issue with it.
Since exponentiation turns addition into multiplication, normalizing causes the function to map very different feature vectors to the same probability distributions.
If the features represent the class labels "cat", "dog", "kangaroo", the vector (1, 2, 3) intuitively means that it is a bit more likely "kangaroo" than the others.
(-10, -9, -8) signals that the input is neither. Yet, their Softmax is the same.
Why?
Because Softmax is invariant for translating each feature with the same value.
If this is not clear, check out the calculation below.
This is the reason why prediction uncertainty is hard to estimate with Softmax.
Overall, you have to be careful when using Softmax. Instead of thinking about the output as the true class probability distribution, view it as an estimation for which class is the most likely.
If you enjoyed this thread, consider following me and hitting a like/retweet on the first tweet of the thread!
I regularly post simple explanations of seemingly complicated concepts in machine learning, make sure you don't miss out on the next one!
An important point by @tpk_in: the exponential function can lead to overflow, so the translational invariance mentioned above can be useful for implementations.
In the last 24 hours, more than 400 of you decided to follow me. Thank you, I am honored!
As you probably know, I love explaining complex machine learning concepts simply. I have collected some of my past threads for you to make sure you don't miss out on them.
Convolution is not the easiest operation to understand: it involves functions, sums, and two moving parts.
However, there is an illuminating explanation — with probability theory!
There is a whole new aspect of convolution that you (probably) haven't seen before.
🧵 👇🏽
In machine learning, convolutions are most often applied for images, but to make our job easier, we shall take a step back and go to one dimension.
There, convolution is defined as below.
Now, let's forget about these formulas for a while, and talk about a simple probability distribution: we toss two 6-sided dices and study the resulting values.
To formalize the problem, let 𝑋 and 𝑌 be two random variables, describing the outcome of the first and second toss.
One of my favorite convolutional network architectures is the U-Net.
It solves a hard problem in such an elegant way that it became one of the most performant and popular choices for semantic segmentation tasks.
How does it work?
🧵 👇🏽
Let's quickly recap what semantic segmentation is: a common computer vision task, where we want to classify which class each pixel belongs to.
Because we want to provide a prediction on a pixel level, this task is much harder than classification.
Since the absolutely classic paper Fully Convolutional Networks for Semantic Segmentation by Jonathan Long, Evan Shelhamer, and Trevor Darrell, fully end-to-end autoencoder architectures were most commonly used for this.
There is a common misconception that all probability distributions are like a Gaussian.
Often, the reasoning involves the Central Limit Theorem.
This is not exactly right: they resemble Gaussian only from a certain perspective.
🧵 👇🏽
Let's state the CLT first. If we have 𝑋₁, 𝑋₂, ..., 𝑋ₙ independent and identically distributed random variables, their scaled sum is a Gaussian distribution in the limit.
The surprising thing here is the limit is independent of the variables' distribution.
Note that the random variables undergo a significant transformation: averaging and scaling with the mean, the variance, and √𝑛.
(The scaling transformation is the "certain perspective" I mentioned in the first tweet.)