Loss Spike Detection

Problem

Flag loss spikes using a z-score over the whole loss list: return the indices where (loss - mean) / std exceeds z_threshold, using the population standard deviation. If the standard deviation is zero, return an empty list.

The signature is loss_spike_detect(losses, z_threshold).

Hint

A spike is an outlier relative to the training run's own loss distribution.

Test cases

case 1

in: [[1,1.1,0.9,1,5],1.5]

expected: [4]

case 2

in: [[1,1,1],2]

expected: []

case 3

in: [[10,1,1,1],1]

expected: [0]

case 4

in: [[0.5,0.6,0.55,2],1]

expected: [3]

Starter codePython

import math
def loss_spike_detect(losses, z_threshold):
    # Your code here
    pass

4 test cases · run your code against them in the editor