Appropriate Problems for Neural Network Learning

Unit 4 Navigation

Introduction to ANN  |  Next → Perceptrons


4.3 When Should You Use Neural Networks?

The Backpropagation algorithm is the most commonly used ANN learning technique. It is appropriate for problems with the following six characteristics:


Characteristic 1: Instances Described by Many Attribute-Value Pairs

The target function is defined over instances described by a vector of predefined features — such as pixel values in an image.

  • Input attributes may be highly correlated or independent of one another
  • Input values can be any real number (not just binary or integer)

Example: In ALVINN, each instance is described by 960 pixel intensity values.
Counter-example: Symbolic logic problems where inputs are categorical labels are less naturally suited (though ANNs can still be applied).


Characteristic 2: Output May Be Any Type

The target function may output:

  • A discrete-valued class label
  • A real-valued number
  • A vector of several real- or discrete-valued attributes

Example: In ALVINN, the output is a vector of 30 values — each a confidence score for a specific steering direction. A single network can simultaneously predict both the steering command and suggested acceleration by concatenating the two output vectors.


Characteristic 3: Training Examples May Contain Errors

ANN learning methods are robust to noise in the training data.

Because Backpropagation fits a smooth function to many examples simultaneously, individual mislabelled or noisy examples have limited influence on the final learned weights. This contrasts with methods like exact rule learning that are sensitive to noise.

Implication: ANNs are well-suited to real-world sensor data — camera images, microphone recordings, radar readings — which inherently contains measurement noise.


Characteristic 4: Long Training Times Are Acceptable

Training times can range from a few seconds to many hours, depending on:

  • Number of weights in the network
  • Number of training examples
  • Desired accuracy
  • Hardware available

Comparison: Decision tree learning is generally much faster to train. ANNs are better suited when training is done offline (once) and the trained network is deployed for repeated fast predictions.


Characteristic 5: Fast Evaluation of the Learned Function Is Required

Once trained, evaluating a neural network for a new input is very fast — just a forward pass through the layers.

Example: ALVINN applies its network several times per second to continuously update the steering command as the vehicle drives. This real-time requirement is easily met once training is complete.

Implication: ANNs have an asymmetric cost profile — slow to train, fast to predict. This is ideal for applications where training is a one-time or infrequent process.


Characteristic 6: Human Interpretability Is Not Required

The weights learned by a neural network are often difficult for humans to interpret. A trained ANN is a “black box” — it produces correct outputs but does not explain why.

Contrast with decision trees: Decision trees produce explicit human-readable rules like “if income > 50K and credit score > 700 then approve loan.” ANN weights encode the same knowledge implicitly across thousands of numbers.

Implication: Use ANNs when performance matters more than explainability. For medical diagnosis, legal decisions, or regulatory settings requiring justification, more interpretable models may be preferred.


Summary Table

CharacteristicANN Suitability
Many numeric attribute-value pairs✅ Ideal
Any output type (discrete, real, vector)✅ Flexible
Noisy training data✅ Robust
Training time can be long✅ Acceptable trade-off
Fast prediction needed after training✅ Very fast inference
Human-interpretable model not required✅ Black box is acceptable

When NOT to Use ANNs

ANNs may be a poor choice when:

  • Training data is very small — ANNs need many examples to generalise well; decision trees or Bayesian methods may work better
  • Interpretability is required — legal, medical, or regulatory contexts often require explainable decisions
  • Training time is severely constrained — faster methods like decision trees or naïve Bayes may be preferred
  • The problem is linearly separable — a simple perceptron or logistic regression is more efficient
  • The feature space is symbolic or relational — ANNs work best with numeric vectors

Comparison: ANN vs. Decision Trees

PropertyANNDecision Tree
Input typeReal-valued vectorsCategorical or real
Noise toleranceHighModerate
Training speedSlowFast
Prediction speedFastFast
InterpretabilityLowHigh
Overfitting riskHigh (without regularisation)High (without pruning)
Best forSensor data, images, speechTabular data, rule extraction

Experimental comparisons (Shavlik et al., 1991; Weiss & Kapouleas, 1989) show that ANN and decision tree learning produce results of comparable accuracy on many tasks — the choice often depends on the requirements above.


Practice Questions

  1. List the six characteristics of problems well-suited for Backpropagation. Give a real-world example that satisfies each.
  2. Why are ANNs robust to noisy training data? What property of gradient descent contributes to this?
  3. In what scenarios would you prefer a decision tree over a neural network, even if both achieve similar accuracy?
  4. Why does fast evaluation of the learned function matter for applications like autonomous driving?

References

  • Mitchell, T.M. — Machine Learning, Chapter 4, Section 4.3   📄 PDF

Introduction to ANN  |  Next → Perceptrons