The Hello World! of machine learning: Classifying handwriting digits.
But everyone solves this problem the same way.
Here is a different, non-boring approach that you haven't seen before.
1 of 9
But everyone solves this problem the same way.
Here is a different, non-boring approach that you haven't seen before.
1 of 9
I used Contrastive Learning to solve this problem.
Nobody gets away with listing MNIST in their portfolio unless you use a different, exciting approach.
Contrastive Learning is just that.
2 of 9
Nobody gets away with listing MNIST in their portfolio unless you use a different, exciting approach.
Contrastive Learning is just that.
2 of 9
Here is the high-level idea:
1. Create a neural network that turns a picture of a digit into an embedding (a vector of numbers.)
2. Embeddings belonging to the same digit should be similar.
3. Embeddings belonging to different digits should be far apart.
3 of 9
1. Create a neural network that turns a picture of a digit into an embedding (a vector of numbers.)
2. Embeddings belonging to the same digit should be similar.
3. Embeddings belonging to different digits should be far apart.
3 of 9
Whenever we receive a new picture:
1. Use the network to create the embedding.
2. Compare it to every digit's template embedding.
3. Correct answer is the digit whose embedding is the most similar to the one we created.
4 of 9
1. Use the network to create the embedding.
2. Compare it to every digit's template embedding.
3. Correct answer is the digit whose embedding is the most similar to the one we created.
4 of 9
To make this happen:
1. We need a network with 2 heads.
2. Take two images as the input.
3. Compute the distance between the results.
The loss function will help us minimize the distance between images of the same digit.
5 of 9
1. We need a network with 2 heads.
2. Take two images as the input.
3. Compute the distance between the results.
The loss function will help us minimize the distance between images of the same digit.
5 of 9
Look at the attached picture.
• Each input expects an image
• The "model" layer computes the embeddings
• The "lambda" layer computes the distance.
We want the distance to be small if both images are the same. If they are different, we want the distance to be large.
6 of 9
• Each input expects an image
• The "model" layer computes the embeddings
• The "lambda" layer computes the distance.
We want the distance to be small if both images are the same. If they are different, we want the distance to be large.
6 of 9
There's a name for this: "Siamese network."
Take a look at the attached code. It shows the Keras implementation of this model. I'm using @ylecun's Contrastive Loss.
The entire code is here:
deepnote.com/@santiago-vald…
7 of 9
Take a look at the attached code. It shows the Keras implementation of this model. I'm using @ylecun's Contrastive Loss.
The entire code is here:
deepnote.com/@santiago-vald…
7 of 9
@ylecun I trained the model for 15 epochs. It reached 92% accuracy on the test set.
Not ground-breaking results, but it's an excellent way to showcase Contrastive Learning.
Attached you can see 10 of the results (including 2 mistakes.)
8 of 9
Not ground-breaking results, but it's an excellent way to showcase Contrastive Learning.
Attached you can see 10 of the results (including 2 mistakes.)
8 of 9
@ylecun Siamese Networks are handy in real-life applications. I've used them a ton.
For classification problems, they are ideal if you only have a few samples of each class.
9 of 9
For classification problems, they are ideal if you only have a few samples of each class.
9 of 9
• • •
Missing some Tweet in this thread? You can try to
force a refresh
