A denoising autoencoder is being trained on greyscale image strips eight pixels long, with brightness running from (black) to (white).
For one training example, three pixels were knocked out (set to ) before the strip reached the encoder. Here is everything involved in that one example:
| position | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
|---|---|---|---|---|---|---|---|---|
| original strip | 200 | 160 | 120 | 80 | 80 | 120 | 160 | 200 |
| corrupted strip, fed to the encoder | 200 | 0 | 120 | 80 | 0 | 120 | 0 | 200 |
| network output | 190 | 140 | 120 | 90 | 70 | 130 | 160 | 180 |
The reconstruction loss is the usual one: subtract, square, then average over every position.
What reconstruction loss does training compute for this example?
Round your answer to 1 decimal place.