Skip to main content

Limitations

Updated May 04, 2023 ·

Overview​

Machine learning has impressive capabilities, but it also has limitations:

  • Data quality
  • Explainability

Data Quality​

The quality of input data is important for machine learning models.

  • "Garbage in, garbage out" means poor data leads to poor results
  • Bad data can produce inaccurate, incomplete, or incoherent outputs

Importance of scrutinizing your model's output.

  • Never blindly trust your model
  • Awareness of data's role is important for machine learning projects
  • The model's performance is only as good as the input data quality

Ensuring high-quality data involves several steps.

  • Data analysis: Examining characteristics, distribution, source, and relevance
  • Reviewing outliers and suspicious data
  • Involving domain experts to explain unexpected patterns
  • Documenting processes to ensure transparency and repeatability

Explainability​

Another significant limitation of machine learning is explainability.

  • Machine learning models are often viewed as black boxes
  • Transparency in AI reasoning is necessary for trust and understanding

Sometimes there is a need for AI systems to be transparent about the reasoning it uses, to increase trust, clarity, and understanding.

  • Business adoption: Explaining models to customers
  • Legal compliance: Adhering to data regulations
  • Bias detection: Faster and more accurate identification of biases

Despite its accuracy, Deep learning often lacks explainability.

  • Deep learning models can make precise predictions without clear reasons
  • Explainable AI methods help us understand prediction factors

Examples​

Explainable AI in Healthcare​

Explainable AI can provide valuable insights, such as in a hospital setting.

  • Prediction: A traditional model can predict Type 2 diabetes onset
  • Inference: Highlights important features, factors like blood pressure

Inexplicable AI in Handwriting Recognition​

In some cases, like handwriting recognition, explainability is less critical.

  • Recognizing letters accurately is more important than understanding why
  • Deep learning is ideal for such tasks due to its high accuracy without needing explanations