Question: Discuss why relying on manually designed intermediate representations can limit the best possible performance of a machine learning system. Use MFCC features and phonemes in speech recognition to illustrate the two main ways this can happen.

Sample answer: Hand-designed preprocessing can hold back a model in two different ways. First, it may compress the input so much that useful details disappear before the learner ever sees them. MFCCs are an example: they turn a speech waveform into a compact feature set that is convenient for modeling, but some information in the original sound signal is intentionally discarded. Second, a pipeline may force the data through an intermediate representation that is only an approximation of the real process. Phonemes are useful linguistic units, but actual speech is continuous, context-dependent, and often does not break cleanly into those categories. If the chosen intermediate unit does not match the true structure of the signal, it becomes a bottleneck and limits how well the full system can perform.

Key points:

• Hand-crafted components can limit performance by removing information or imposing an imperfect representation.
• MFCCs reduce raw audio to a smaller feature set and lose some signal detail.
• Phonemes are a simplified linguistic abstraction, not a perfect model of speech acoustics.
• A poor intermediate representation can become the system’s performance bottleneck.

Rubric: A strong response should identify both limitations clearly and connect them to MFCCs and phonemes as examples of information loss and imperfect intermediate representations.