Decoding electroencephalography (EEG) signals with deep learning remains challenging due to strong interand intra-individual variability, low signal-to-noise ratio, and the absence of an optimal architecture. While deep neural networks have demonstrated promising performance for EEG decoding, selecting an appropriate model architecture remains a costly and subject-specific process. Neural architecture search approaches partially address this issue but are computationally prohibitive in practice. In this work, we investigate a growing neural network strategy that adapts model expressivity during training, only expanding the architecture when required by the data. Building upon a lightweight convolutional neural network (CNN) architecture commonly used for a recent Brain-Computer Interface paradigm, code modulated visual evoked potential, decoding, we propose to grow only the classification head with two different growing modes, directed acyclic graph (DAG) and growing linear layers (GLL). This design explicitly targets subject-specific variability while preserving architectural priors learned from the literature. We evaluate our approaches on a five-class EEG dataset comprising 24 participants and compare them against fixed CNN, Riemannian TS-LDA, and GREEN architectures. Results show that the growing CNN significantly outperforms fixed CNN by increasing the accuracy by 1% while having 80% less neurons in the last layers. Our findings suggest that selectively growing the final layers of a deep network provides the capacity to adapt to subject-specific signatures, thus addressing the inter-subject variability. Moreover, we find an interesting trend where participants with better accuracy need fewer additional neurons to optimize their performance.