AI Model Compression Part V: The LSTM: A New Kind of Memory

Ateeb TaseerPosted on January 1, 2025

3-4 Min Read Time

After overcoming the initial struggles with maintaining context over time, the next leap in AI’s memory capabilities came with the development of the Long Short-Term Memory (LSTM) architecture, which added a new level of sophistication to how machines retain and process information.

Birth of the Long Short-Term Memory

In 1997, Hochreiter and Schmidhuber introduced the Long Short-Term Memory (LSTM) architecture. But to understand its genius, let's first understand the human memory system it mirrors:
 

Human Memory System          LSTM Gates
------------------          ----------
Attention Filter        →   Input Gate
                           (what to remember)

Working Memory          →   Memory Cell
                           (current state)

Memory Consolidation   →   Forget Gate
                           (what to forget)

Memory Retrieval       →   Output Gate
                           (what to use)

Explore how memory systems in AI evolved in Part IV of this series.

The mathematics of the LSTM tells this story of selective memory:
 

Input Gate:
i_t = σ(W_i[h_(t-1), x_t] + b_i)

Forget Gate:
f_t = σ(W_f[h_(t-1), x_t] + b_f)

Memory Cell:
c_t = f_t ⊙ c_(t-1) + i_t ⊙ tanh(W_c[h_(t-1), x_t] + b_c)

Output Gate:
o_t = σ(W_o[h_(t-1), x_t] + b_o)

Each equation represents a crucial aspect of conscious memory, forming the backbone of deep learning solutions that drive today's intelligent systems.

• The Input Gate decides what new information is worth remembering
• The Forget Gate determines what old memories can fade
• The Memory Cell maintains the current state of understanding
• The Output Gate chooses what memories are relevant now

The Dance of Memory: How LSTMs Learn

Think of an LSTM as a master storyteller, constantly deciding:
 

• Which details to emphasize
• Which to let fade
• How to connect distant events
• When to recall earlier information

This mirrors how human consciousness works with memory:

Example: Reading a Mystery Novel
 

Human Process          LSTM Process
-------------         ------------
Note key clues    →   Input Gate activates
                      for important information

Hold suspects     →   Memory Cell maintains
in mind               key details

Discard red       →   Forget Gate removes
herrings              irrelevant information

Connect final     →   Output Gate retrieves
clues                 stored information