Degradation
Degradation is a decline in a system’s or model’s performance, quality, or effectiveness relative to a previous or expected level.
Sutskever's List
Model degradation is the worsening of training or evaluation performance as network depth increases despite added representational capacity.
- How it is used
- Frames worsening performance as the engineering obstacle that motivated residual networks.
- How deep it goes
- Explains degradation through forward-signal distortion and backward-gradient loss, then compares measured errors without a mathematical derivation.
- Examples
- Compares 20- and 56-layer plain networks on CIFAR-10 and 18- and 34-layer plain and residual networks on ImageNet.
- What you can do after
- Diagnose/debug with it why adding layers can worsen training error and why residual shortcuts can help.
Math and Architectures of Deep Learning
In residual networks, the phenomenon where adding more layers makes training and test performance worse instead of better.
- How it is used
- Explains why simply stacking layers can worsen optimization and motivates residual networks.
- How deep it goes
- Defines the phenomenon, compares errors, diagnoses likely causes, and derives skip connections without detailing optimization internals.
- Examples
- Compares 20-layer and 56-layer CIFAR-10 networks and motivates residual blocks implemented with PyTorch.
- What you can do after
- Evaluate depth-related error increases and understand why residual skip connections address them.
Read how Math and Architectures of Deep Learning explains degradation →
LLMs in Production
Performance decay is a decline in a model's output quality over time despite the model continuing to operate.
- How it is used
- Uses degradation to motivate ML monitoring that detects quality loss beyond ordinary service failures.
- How deep it goes
- Defines performance decay and gives monitoring procedures, but does not explain model-internal causes or remediation algorithms.
- Examples
- Demonstrates rolling whylogs profiles with LangKit metrics for readability, complexity, toxicity, and prompt-injection similarity.
- What you can do after
- Diagnose silent LLM quality decline by monitoring text profiles and evaluating production outputs.
Designing Deep Learning Systems: A guide for software engineers
A decline in a deployed machine learning model's predictive or operational performance over time.
- How it is used
- Supports production monitoring by connecting serving telemetry and evaluation metrics to declining model performance.
- How deep it goes
- Defines degradation, identifies changing target distributions as a cause, and names monitoring signals without detailing remediation.
- Examples
- Uses response latency, prediction miss rate, evaluation metrics, and changing slogans affecting voice recognition.
- What you can do after
- Diagnose degradation with serving execution data, performance metrics, and lineage linking models to training runs.
Read how Designing Deep Learning Systems: A guide for software engineers explains degradation →
Building Reliable AI Systems
A decline in the quality of a system's outputs compared with an earlier or expected level.
- How it is used
- Detects declining LLM response quality in production before subtle failures erode user satisfaction or trust.
- How deep it goes
- Details golden-dataset checks, automated scoring, and asynchronous shadow testing without specifying a complete evaluation system.
- Examples
- Uses password-reset response checks and parallel candidate-model comparisons to illustrate quality changes.
- What you can do after
- Diagnose emerging quality changes by testing response patterns and comparing candidate outputs with production responses.
Read how Building Reliable AI Systems explains degradation →
Related concepts
- data driftproduced by · Data drift is a change over time in the statistical distribution or characteristics of data relative to a reference distribution or set of characteristics.
- throughputuses · The amount of work, data, or output a system processes or completes per unit of time.
- benchmarkmeasured by · A benchmark is a defined task, dataset, workload, procedure, or reference measure used to assess or compare performance or capability under specified criteria or conditions.
- problemkind of · A problem is a difficulty, obstacle, or undesirable condition requiring investigation, mitigation, or resolution.
- atomicityproduced by · Atomicity is the property that an operation or group of operations takes effect as an indivisible, all-or-nothing unit, without observable partial effects.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.