Question at position 65 Why is argmax considered unsuitable for backpropagation in classification tasks?Why is argmax considered unsuitable for backpropagation in classification tasks?Its derivative is effectively zero, preventing gradient-based parameter updatesIt introduces negative probabilities for less likely classesIts probabilities often exceed 1, making the network unstable单项选择题
A
Its derivative is effectively zero, preventing gradient-based parameter updates
B
It introduces negative probabilities for less likely classes
C
Its probabilities often exceed 1, making the network unstable

Question

Question at position 65 Why is argmax considered unsuitable for backpropagation in classification tasks?Why is argmax considered unsuitable for backpropagation in classification tasks?Its derivative is effectively zero, preventing gradient-based parameter updatesIt introduces negative probabilities for less likely classesIts probabilities often exceed 1, making the network unstable单项选择题

A

Its derivative is effectively zero, preventing gradient-based parameter updates

B

It introduces negative probabilities for less likely classes

C

Its probabilities often exceed 1, making the network unstable

BlackTom AI · Accepted Answer

The question asks why argmax is considered unsuitable for backpropagation in classification tasks. We will evaluate each option in turn.
Option 1: Its derivative is effectively zero, preventing gradient-based parameter updates. This is the correct intuition. The argmax operation is a non-differentiable, piecewise-constant function: small changes in input do not change the argmax until a threshold is crossed, and at the decision boundary the derivative does not provide meaningful gradient information. In practice, this means that using argmax directly in the loss or as part of the computation graph yields near-zero or undefined gradients, which prevents effective parameter updates via gradient descent.
Option 2: It introduces negative probabilities for less likely classes. Argmax does not assign any probabilities at all; it selects a single class as the maximum. It does not produce negative probabilities, because probabilities are not produced by argmax itself. Therefore, this claim is inaccurate—argmax is a discrete decision, not a probabilistic output.
Option 3: Its probabilities often exceed 1, making the network unstable. This statement is incorrect because argmax does not output probabilities at all. It returns an index corresponding to the largest value among inputs; there is no probability value produced by argmax that could exceed 1. The notion of probabilities exceeding 1 does not apply to the argmax operation.
In summary, the central issue with argmax for backpropagation is its non-differentiability and, more importantly, the absence of useful gradients, which blocks gradient-based optimization. The other options mischaracterize what argmax outputs or how it behaves, since it neither produces probabilities nor values outside the [0,1] range, and it does not inherently introduce negative probabilities.

类似问题

以下哪一项正确描述了反向传播算法的步骤？ Which steps correctly describe the backpropagation algorithm?

Which of the following options is the key factor that led the return of neural networks?

[Multiple choice] Which of the following options is the key factor that led the return of neural networks?

Question at position 2 What does backpropagation refer to in neural networks? Computing gradients for all parametersMapping inputs to outputsNormalizing dataUpdating weights in the forward direction

What is back propagation?

In the connectionist model, the weights vary as a function of training, in which feedback as to the correctness or incorrectness leads to a mathematical adjustment of the weights. A commonly used training procedure for this weight adjustment is known as ___.

What is backpropagation in the context of training a language model?

更多留学生实用工具

智能学习助手

风格化写作助手

论文查重助手

文献引用助手

课堂转译助手

课堂笔记助手

Quiz搜索助手

学校历年真题

智能刷题助手

智能匹配练习

希望你的学习变得更简单