Deep Learning 1.

Today I was watching a video from MIT: https://www.youtube.com/watch?v=ErnWZxJovaM&pp=ygUNZGVlcCBsZWFybmluZw%3D%3D about Deep Learning.

I learned about basics like what is deep learning, perceptron, forward propagation and some common parts of training like loss functions and activation functions.

Deep learning is just a possibility to extract patterns from data using neural networks. We can use large amounts of data to teach the neural network how to use some patterns and then use it to generate new stuff. That’s how GPT or Stable Diffusion is working. Such neural network is made out of perceptrons, which are building blocks. You stack a lot of perceptrons together into layers, and layers into whole network. You start with input layer where you give data. Next you have some hidden layers connected with each other by weighted connections. Hidden layers have some activation function. The most common functions are sigmoid, hyperbolic tangent or rectified linear unit (ReLU). Last layer is output layer which provides us the final result, e.g. prediction of next element like text, pixel or anything we want.

To teach such network, we start by using random weights on all connections. Then we show data - our inputs and expected outcomes. For each iteration we calculate the error between actual weights and expected value and by small step we are adjusting them up or down. Over time, we might end up with overfitting, so we can’t teach network with same data all the time. Spending not enough time is also not optimal as we will have under fitting, so the outcomes will be to “blurry”.

Making perfectly flat surfaces

As an engineer I spend a lot of time learning about different concepts in engineering. Not only software or hardware related engineering, but in all aspects. I found this video, which teaches you the concept of making a super flat surfaces: https://youtu.be/zckkvxwjIuM?si=oNoxMEidcVg69qB5

The whole idea is to have 3 different plates and rub them in particular order. Let’s call them A, B and C. So you start with A and B. Next you are doing A and C. After that B and C. You keep going with such order rubbing it in different directions. Over time, you will get perfectly flat surface. I feel like this is really creative way to manufacture something. I’m curious is it possible to apply it to 3D printed parts. To get smooth part, that can be used in laboratory.