You've learned about feedforward neural networks, also known as Multi-Layer Perceptrons (MLPs), and how they can learn complex, non-linear relationships in data. Their structure, typically involving densely connected layers where each neuron connects to every neuron in the previous layer, makes them powerful function approximators when the input is a fixed-size vector of features.

However, this very structure introduces limitations when dealing with specific kinds of data that possess inherent structure not easily captured by a simple vector representation. Let's examine two significant areas where standard feedforward networks often fall short: processing grid-structured data like images, and handling sequential data like text or time series.

Handling Spatial Structure in Images

Images are not just arbitrary collections of pixels; they have a strong spatial structure. Pixels close to each other are often related, forming textures, edges, and objects. When you feed an image into a standard MLP, the typical first step is to flatten the image matrix (e.g., height x width x channels) into a single, long vector.

A 2D image is typically flattened into a 1D vector before being fed into an MLP. Notice how pixels that were adjacent vertically (e.g., 1 and 4) are now separated in the vector.

This flattening process discards the explicit 2D spatial arrangement. Pixels that were close neighbors in the image (like one directly above another) might end up far apart in the input vector. An MLP treats each element of this vector somewhat independently in the first layer, making it difficult to learn features based on local spatial patterns efficiently.

Furthermore, consider the connections. In a densely connected layer, every neuron connects to every input pixel. For even moderately sized images (e.g., 256x256 pixels), this results in an enormous number of parameters (weights) in the very first hidden layer. This parameter explosion leads to several problems:

Computational Cost: Training becomes very slow and memory-intensive.
Overfitting: With so many parameters, the model can easily memorize the training data instead of learning generalizable features.
Ignoring Locality: The network doesn't inherently know that pixels nearby are more relevant to each other than pixels far apart. It has to learn this from scratch, which is inefficient.

Another challenge is translation invariance. An object (like a cat) should still be recognizable as a cat whether it appears in the top-left corner or the bottom-right corner of the image. MLPs lack built-in translation invariance. Because each input pixel is treated uniquely by the weights connecting to it, the network essentially needs to learn to detect the features of a cat separately for every possible position it might appear in. This is highly inefficient.

Handling Sequential Order and Variable Lengths

Data like natural language text, speech audio, or financial time series are inherently sequential. The order of elements matters significantly. For example, "dog bites man" has a very different meaning from "man bites dog".

Standard MLPs run into two main difficulties with sequences:

Fixed-Size Inputs: MLPs require inputs to be vectors of a predefined, fixed size. However, sequences naturally vary in length. Sentences can have few or many words, time series can cover different durations, and audio clips vary. How do you fit a 10-word sentence and a 100-word sentence into the same fixed-size input layer? Common workarounds like padding shorter sequences or truncating longer ones can work but often introduce noise or discard potentially relevant information.
Lack of Memory/State: An MLP processes its entire input vector "at once" during the forward pass. It doesn't have an intrinsic mechanism to remember information from earlier parts of a sequence while processing later parts. When processing the word "bites" in "dog bites man," a simple MLP doesn't inherently retain the context that "dog" came before it. While you could potentially feed the entire sequence (padded/truncated) as one large vector, the network struggles to learn long-range dependencies. The relationship between words far apart in a long sentence is hard to capture when each input element is processed through independent weights without a mechanism for maintaining state or sharing information across time steps.

Similar to the image case, MLPs also lack parameter sharing across time steps. If a specific pattern or feature is important regardless of where it appears in the sequence (e.g., detecting a specific phrase), an MLP would need different weights to detect that pattern at the beginning, middle, or end of the sequence. This is inefficient and requires more data to learn effectively.

These limitations highlight the need for network architectures specifically designed to handle spatial hierarchies (like in images) and temporal dependencies (like in sequences). Convolutional Neural Networks (CNNs) address the challenges of spatial data, while Recurrent Neural Networks (RNNs) are tailored for sequential data. We will explore the fundamentals of these architectures next.