問91
頻出度★★☆
AIに利用されるニューラルネットワークにおける活性化関数に関する記述として, 適切なものはどれか。
令和5年度 ITパスポート試験 問91
テクノロジ系基礎理論 › 情報に関する理論
This question's explanation and your answers so far.
解説
問 91
Incorrectア
“Outputting a confidence level from the overall result of the network” is processing that corresponds to evaluating the model output or computing its likelihood, not the activation function. The activation function handles the transformation of values inside each individual neuron; it is not a function that calculates the confidence of the network's overall result.
Incorrectイ
Deciding the number of neurons needed in the intermediate (hidden) layer is a matter of the network-structure design stage, not the work of the activation function. The activation function does not output the number of neurons.
Incorrectウ
Outputting the optimal number of neurons from the connection structure of neurons is an explanation about optimizing (designing and searching) the network structure, which is something different from the activation function. The activation function is not a function that computes a structure to find the number of neurons.
- エ
正解 The activation function is a function that works inside each individual neuron (node) that makes up a neural network. Each neuron multiplies the several input values received from the previous layer by weights and sums them, passes that sum through the activation function to decide an output value, and then passes it on to the neurons of the next layer. The sigmoid function and ReLU are typical examples; they play the role of non-linearly transforming, according to the input, how strongly the signal should fire. Therefore, choice エ, which describes “within a single neuron, it calculates based on the input value and outputs the value to be passed to the next neuron”, is appropriate.
ここがポイント
- The activation function is a non-linear function (e.g., the sigmoid function, ReLU) that, inside each neuron, receives the “weighted sum of the inputs” and transforms it into an output value (the value passed to the next neuron).
- Deciding the number of neurons and the layer structure belongs to network design, and should be distinguished from the role of the activation function (the value transformation inside a neuron).
用語