Rethinking Learning Dynamics in AI Models: An Early Theory from Experimentation

Source: Dev.to

Core Assumption in Deep Learning

• Typical view: Minimize loss → improve performance.
• Observation: Loss minimization alone does not always correlate with meaningful representation learning, especially in the early phases of training.

Proposed Idea

Representation Instability Phase – gradients first optimize surface‑level patterns before stable internal abstractions emerge.
I’m not sure whether this phenomenon already has a name, or if I’m mistaking training noise for structure.

Empirical Observations

These dynamics were observed while training a small transformer‑like model on synthetic data, by logging intermediate layer activations.

Training Loop Example

for epoch in range(num_epochs):
    optimizer.zero_grad()

    outputs = model(inputs)
    loss = criterion(outputs, targets)

    loss.backward()
    optimizer.step()

    # Track embedding norm without gradients
    with torch.no_grad():
        embedding_norm = model.encoder.weight.norm().item()

    print(f"Epoch {epoch} | Loss: {loss.item():.4f} | Embedding Norm: {embedding_norm:.2f}")

Note: Embedding norms and cosine similarities changed more drastically than loss values, especially early on.

Cosine Similarity Tracker

prev_embedding = None
for epoch in range(num_epochs):
    current_embedding = model.encoder.weight.clone().detach()

    if prev_embedding is not None:
        similarity = cosine_similarity(
            prev_embedding.view(-1),
            current_embedding.view(-1)
        )
        print(f"Epoch {epoch} | Embedding Stability: {similarity.item():.4f}")

    prev_embedding = current_embedding

The similarity score jumps erratically at first, then converges toward ~0.98–0.99 even when loss improvement becomes marginal.

Working Hypothesis

• Gradient descent initially prioritizes optimization shortcuts rather than semantic structure.
• Later it converges toward representations that are robust and generalizable.

Potential Implications

• Early stopping might prevent meaningful abstraction.
• Some overfitting phases could be necessary for representation formation.
• Regularization may delay, rather than prevent, representation collapse.

Open Questions

1. Dataset size: Is this an artifact of small datasets?
2. Existing theory: Does this align with concepts like loss‑landscape flatness, mode connectivity, or the information bottleneck?
3. Metrics: Are there better metrics than loss for tracking learning quality?
4. Noise vs. structure: Am I confusing emergent structure with random alignment?
5. Formal recognition: Is the “instability → abstraction → stabilization” pattern formally recognized in the literature?

I’m treating this more as a question than a claim. Any insights into where the reasoning might break—or how to frame it more rigorously—are greatly appreciated.