Which pixels predict the digit?
Each pixel has a correlation with the label. High correlation = that pixel helps the model classify. Drag the slider to keep only the most important pixels.
Correlation heatmap
Important pixels
Masked digit
Top 10 most predictive pixels
| Rank | Pixel | Position | Correlation |
|---|---|---|---|
| 1 | pixel_0 | row 0, col 0 | nan |
| 2 | pixel_32 | row 4, col 0 | nan |
| 3 | pixel_39 | row 4, col 7 | nan |
| 4 | pixel_52 | row 6, col 4 | 0.391 |
| 5 | pixel_27 | row 3, col 3 | 0.275 |
| 6 | pixel_35 | row 4, col 3 | 0.266 |
| 7 | pixel_12 | row 1, col 4 | 0.244 |
| 8 | pixel_28 | row 3, col 4 | 0.234 |
| 9 | pixel_33 | row 4, col 1 | 0.223 |
| 10 | pixel_29 | row 3, col 5 | 0.216 |
Loading...
Loading...
Loading...
Think Deeper
Try this:
In the Explore tab, set the correlation threshold to 0.5 and click 'Show on digit' several times. Can you still recognize the digits?
Usually yes — center pixels carry almost all the signal. This is feature selection: drop noisy columns (like TTL or source MAC) and keep the signal (entropy, byte distribution, timing patterns).
Cybersecurity tie-in: This is feature selection. In security ML:
high-correlation features are dest_port, bytes_sent, entropy.
Low-correlation features are source MAC, TTL.
Dropping useless features makes models faster and more accurate.