There's no deep learning course more visual and intuitive than this.

LLMs like ChatGPT are like magic.
How does the magic of generating poetry or writing code with a simple prompt happen? This book breaks down the concepts into easy-to-understand sections, presenting them with intuitive examples and clear illustrations.
Starting from the basics of neural networks and AI, it covers the latest technologies such as image classification and natural language processing step by step. PyTorch tutorials are provided for each major concept, so you can learn how to code neural networks from scratch.
The appendix provides illustrated explanations of mathematical concepts you may have learned in school but have now forgotten.

The most basic element when creating a neural network in PyTorch is the tensor.
Tensors are similar to Python lists, but are used to store data, weights, and biases.
The biggest difference between tensors and Python lists is that in tensors, all values ​​must have the same data type.
For example, if you create an integer tensor, all values ​​in it must be integers.
Another big difference is that tensors are optimized for use in neural networks.

--- p.18

Backpropagation consists of two steps.
The first step is to compute the derivative for each parameter to be optimized.
This set of differential values ​​is called the 'gradient'.
The name 'backpropagation' is given because this gradient is calculated from the output layer of the neural network towards the input layer.
/ Then, in the second step, we adjust the parameters using gradient descent or other optimization methods.
This process proceeds iteratively to find the optimal value.
That is, it is a method of gradually improving parameters through several small steps.

--- p.20

The first neural network, created in the 1950s, was a perceptron designed to recognize handwritten digits and letters.
However, these early CNN designs didn't work well in practice until the 1990s because the datasets were too small and the computers were too slow.
Fortunately, however, researchers working on GPUs (graphics processing units) discovered that they could run CNNs much faster. GPUs sped up CNNs, and CNNs became better at image classification, kickstarting the AI ​​revolution we're currently experiencing.

--- p.128

The first encoder-decoder model used LSTM (Sutskever et al.
2018), could be expanded to allow flexibility in the length of input and output sequences… and non-numeric inputs and outputs.
We also used a word embedding network, which allows for easy training.
I will explain how to use the encoder-decoder model to solve the sequence-to-sequence problem.
So, let's create a model that translates English phrases into Spanish phrases.

--- p.184

Now that we know the details of the encoder-decoder model, let's code two layers, each with two LSTMs.
Generating two word embeddings for each of the four tokens can be done with nn.Embedding().
We can create stacks and layers of LSTM units with nn.LSTM().
If you set input_size to 2, you can use all word embedding values ​​as input… and set hidden_size to 2, which receives the same input values, and create two stacks of LSTM units.
By setting num_layers to 2, we can create two stacks of LSTM units... and finally, we can create a fully connected layer with nn.Linear().

--- p.198

To understand the importance of keeping track of word order, imagine if Normalsaurus said... 'Squatch eats pizza.' In this case, Squatch would probably say... "Yum!!!" Conversely, imagine if Normalsaurus said... 'Pizza eats Squatch.' In this case, Squatch would probably say... "Eww!!!" These two sentences use exactly the same words, but have very different meanings.
Therefore, it is very important to keep track of word order.
Let's learn how transformers learn word order using positional encoding!

--- p.216