5.1 Introduction to CNNs and RNNs

Two of the most common ANNs are convolutional neural networks (CNNs) and recurrent neural networks (RNNs). Building on your existing knowledge, we'll explore how these architectures process data, make predictions, and incorporate advanced techniques such as attention mechanisms and transformers in complex tasks such as vision and language processing.

You will:

understand how backpropagation, feed forward, and prediction fit into neural networks;
understand the use of attention mechanisms and self-attention;
be able to describe the use of transformers such as BERT and GPT.
We're going to explore two of the most common types of neural networks in modern
AI: convolutional neural networks and recurrent neural networks;
we'll also look at transformers.
These architectures are key to understanding how AI processes complex
data, such as images and sequences, building on the foundational concepts of
feed forward and backpropagation you've already covered.
CNNS are specifically designed to process visual data,
making them the backbone of image recognition tasks.
AlexNet, which we saw in Module 1, is a good example of a CNN.
Unlike traditional feed-forward networks, which consume information about pixels,
CNNs ingest low-level features by using initial "convolutional" layers—known as
filters or kernels—to analyze images and detect specific features, such as edges,
textures, and shapes.
These are then combined to predict the content of the image.
CNNs are powerful because they capture spatial hierarchies in data,
understanding both simple patterns and complex structures.
As in an image, they detect basic elements, then more
complex constructs, and then more abstract or complete
objects.
After the forward pass, the network adjusts its filters during
backpropagation to improve its accuracy over time.
Turning to RNNs, these are designed to handle sequential
data, such as text or time series.
The key feature of the RNN architecture is a global feedback-loop connection that
encompasses the hidden layers, allowing the model to create memories and
learn sequential information present in time-based inputs.
Whereas the ANN architecture that we looked at previously had data flowing in
one direction only, the loopback connections in RNNs allow
information to stick around.
This acts as a form of memory, making RNNs perfect for tasks in which
context is essential, such as language processing and
time-series-based classification.
RNNs process data step-by-step, with each step's output dependent on both
the current input and the information carried over from previous steps.
This involves backpropagation through time, or BPTT.
Just as in CNNs, backpropagation is used to adjust
network parameters, but in RNNs it acts across time steps,
helping the model learn from an entire sequence.
This enables the neural network to recognize long-term dependencies in the
data.
What if you want your model to focus on specific parts of the sequence?
This is where attention mechanisms and "self-attention" come in.
Attention mechanisms allow a network to weigh the importance of different parts
of the input data more heavily than others,
which is useful in tasks such as language translation or text generation.
Self-attention, a core component of transformers such as
BERT and GPT, take this idea further.
Instead of processing data sequentially, one input at a time like—RNNs—
transformers process all parts of the input tokens in an input sequence at the
same time.
Self-attention enables the model to consider relationships between these
different parts of the input data, regardless of their position.
This allows transformers to capture nuanced relationships,
and their parallel processing lets them process huge datasets,
leading to significant advances in natural-language processing and models
that can be further trained for specific domains.
RNNs are still useful for tasks that involve shorter sequences,
but transformers are now the dominant model for learning long-term dependencies
with multiple contexts in natural-language processing, computer vision,
and other time-series-related tasks.
As you explore more CNNs, RNNs and transformers,
you'll see how these architectures are tailored to different types of data and
tasks. By understanding the underlying mechanisms—
feed forward, backpropagation, and attention—
you'll gain insights into how modern AI achieves its remarkable capabilities in
areas such as image recognition and language understanding.
As you move deeper into the world of AI, where task complexity demands
increasingly sophisticated models, these concepts will be foundational.

Видео 5.1 Introduction to CNNs and RNNs канала CodeAI Academy