How Do We Grade Mistakes? (Error Metrics Made Easy)¶

When a regular calculator gives the wrong answer, it gets an F. In approximate computing, we design circuits that make deliberate small mistakes. But how do we know if a mistake is "good and safe" or "catastrophic"?

We use statistical error metrics! Let's explain them using simple, everyday analogies.

🍕 1. Mean Error Distance (MED) — "The Pizza Slice Test"¶

Imagine you order a pizza with 100 slices. - If the restaurant gives you 99 slices on Monday (1 slice error) - And 101 slices on Tuesday (1 slice error) - And 100 slices on Wednesday (0 slice error)

Your average absolute mistake is: $$\text{Average Error} = \frac{1 + 1 + 0}{3} = 0.67\text{ slices}$$

In circuits, Mean Error Distance (MED) is the average difference between what the approximate chip calculated and what an exact computer would have produced:

\[\text{MED} = \frac{1}{N} \sum |Y_{\text{approx}} - Y_{\text{exact}}|\]

Rule of Thumb: A smaller MED means the circuit stays very close to the true answer on average.

📊 2. Mean Relative Error Distance (MRED) — "The Dollar Test"¶

Think about these two mistakes: 1. Being off by $10** when buying a **$20,000 car (a tiny $0.05\%$ mistake — nobody cares!). 2. Being off by $10** when buying a **$1 candy bar (a $1000\%$ mistake — huge disaster!).

Because numbers have different scales, we calculate the relative percentage error:

\[\text{Relative Error} = \frac{|\text{Mistake}|}{\text{True Value}}\]

Taking the average over all calculations gives us Mean Relative Error Distance (MRED):

\[\text{MRED} = \frac{1}{N} \sum \frac{|Y_{\text{approx}} - Y_{\text{exact}}|}{|Y_{\text{exact}}|}\]

High-School Physics Tip: In image processing and audio, keeping $\text{MRED} < 2\%$ makes the noise completely invisible to humans!

💥 3. Worst-Case Error (WCE) — "The Ceiling"¶

What is the biggest possible blunder this circuit can ever make in the worst possible scenario?

\[\text{WCE} = \max |Y_{\text{approx}} - Y_{\text{exact}}|\]

If a 16-bit multiplier has $\text{WCE} = 64$, you are guaranteed that no matter what crazy numbers you multiply, the answer will never be off by more than 64 (out of $4,294,967,295$!).

📸 4. PSNR (Peak Signal-to-Noise Ratio) — "Image Crispness Score"¶

When you filter or compress a photo, engineers use PSNR (measured in decibels, dB) to measure picture quality:

PSNR > 40 dB  :  Flawless crystal clear (Identical to original)
PSNR 30–40 dB :  Excellent quality (Instagram / YouTube HD standard)
PSNR 20–30 dB :  Noticeable grain or compression blur
PSNR < 20 dB  :  Pixelated mess

Approximate multipliers routinely achieve $\text{PSNR} > 36\text{ dB}$ while saving $60\%$ chip energy!

⚖️ Summary Cheat Sheet¶

Metric	What It Asks	Best Analogy
MED	On average, how far off is the answer?	Average distance from target
MRED	What is the average percentage mistake?	Percent error on a lab report
WCE	What is the worst mistake that can ever happen?	The safety limit / ceiling
ER (Error Rate)	What percentage of answers have any error at all?	How often the coin lands on heads
PSNR	How beautiful does the filtered picture look?	High-definition display rating