chore: knn tuto

[kcelia]

In this PR, I added a new tutorial for KNN.

As you know, this model is really restrictive, as we are not able to exceed 9bits in the circuit, without having to wait > 1h for geyken + facing unstable execution with the compiler.

To apply the rounded features, we need to :

• provide the "rounding_threshold_bits (int)" parameter
• once rounding_threshold_bits parameter is provided, it will allow the instantiation of an auto-rounder object that will be used after the distance computation in 'inference'
• force the compiler to use auto_adjust_rounders to True.

I want your advice to know where rounding_threshold_bits should be set for the user:

• init:
  - naturel place for arguments
  - easy to add
  - but knn won't follow other CML patterns
• in the compiler:
  - which seems more appropriate.
  - The way I see it, the auto-rounder object will be instantiated in this function. But it won't work if we use 'predict=disable', unless if I put a condition to skip the auto-rounder. But the big problem is that we won't have same results with and without auto-rounder.
• fit, the paramater

(for Q4, we will try to add this feature it to our built-in models, so it's like an advanced conversion)

CC @zama-ai/machine-learning

---

We agreed to hide the parameter rounding_threshold_bits to the user and add a default value in base.py.
I suggest the value 5, you can see the plots in this branch 'testing/impact_of_rouning_on_knn' in use_case_example folder.

Update:
<!> The tests when we do not compile before the prediction, typically for fhe=disable are crushing, because we use the autorounder and we havent compiled the model !

New update:
I used the adjust method instead of the compilation to use the autorounder feature.
The concrete python autorounder must become a serialized object !

[andrei-stoian-zama]

Looks pretty good! But I'd still like to see some graphs for rounding bits vs accuracy first.

1. "b. Simulation: inference on non-encrypted quantized data, while simulating all FHE operations, failure probabilities and crypto-parameters. This mode of inference is recommended in the deployment phase. For further information, please consult this link"

I think you mean "development" not "deployment"

2. you mention "accuracy loss" in the last line in the notebook but there is actually a 5% gain. You could say "accuracy difference"

[RomanBredehoft]

looks good to me, thanks !

[github-actions]

Coverage passed ✅

Coverage details

---------- coverage: platform linux, python 3.8.18-final-0 -----------
Name    Stmts   Miss  Cover   Missing
-------------------------------------
TOTAL    6211      0   100%

51 files skipped due to complete coverage.

[andrei-stoian-zama]

Good job on this !