architecture-design-principles

What problem does it solve?

This Skill provides foundational principles for designing neural network architectures, helping you avoid common mistakes like ignoring inductive biases, over-engineering, or failing to use skip connections. It ensures your network is well-suited to your problem, data, and computational constraints, leading to more efficient and effective models.

Core Features & Use Cases

• Inductive Bias Matching: Learn to align your architecture (MLP, CNN, RNN, Transformer, GNN) with the inherent structure of your data.
• Complexity Management: Master the "start simple" approach, incrementally adding complexity only when justified, saving development time and compute.
• Use Case: You're building a deep image classification model. This skill emphasizes using CNNs for their spatial inductive biases, incorporating skip connections (ResNet) for depth, and balancing depth/width to prevent vanishing gradients and degradation.

Quick Start

I'm designing a neural network for a new problem. How do I choose the right architecture and avoid over-engineering?