In recent years, exercise recognition has become increasingly popular for exercise monitoring and rehabilitation for older adults. However, to identify non-learned exercises, another dataset needs to be collected for differentiation purposes. This study aims to provide an add-on technique to the encoder model to distinguish between learned and non-learned exercises without training with non-learned exercise data. First, we form a list of activation nodes based on the output of the encoder by using the k-means algorithm. Afterward, these nodes are used to compute activation scores from encoded features. These activation scores are used to differentiate non-learned exercises by a threshold value. After differentiation, the activation scores are then passed to a multi-layer perceptron (MLP) for exercise classification. Meanwhile, we proposed a unique method to compute a distinguishing score to find the optimal$k-\mathbf{nodes}$. We demonstrate the proposed method using the MM-Fit dataset, showing that it can identify fitness exercises and distinguish non-learned exercises without much performance loss. Lastly, we collected a dataset about Chair-Fitness activity to validate the proposed method's effectiveness further and welcome other researchers to utilize the dataset.