Although transformers are state-of-the-art models for natural language tasks, obtaining reasonable performance still often requires large transformers which are expensive to train and deploy. Fortunately, there are techniques to increase the size of transformers without extra computing costs. One such technique is sparsity. However, it remains unclear whether sparse architecture is intrinsically more efficient than its dense counterpart. In this paper, we investigate whether replacing the feedforward networks in small transformers with sparse alternatives results in better predictions and faster inference. We found that although inference speed does not increase due to software and hardware limitations, certain sparse alternatives do result in better language understanding. Our research contributes to smarter architectural decision making when designing small language models.

The rapid expansion of large language models (LLMs) driven by the transformer architecture has raised concerns about the lack of high-quality train ing data. This study investigates the role of acti vation functions in smaller-scale language models, specifically those with approximately 10M param eters, to ensure sustained progress in LLM devel opment despite data limitations. Activation func tions, crucial for neural network performance, have evolved significantly, but comprehensive compar isons under consistent conditions remain scarce, especially for smaller parameter count models. This research systematically evaluates traditional and novel activation functions, including learnable variants, and introduces the Kolmogorov-Arnold Network (KAN) to language modeling. Using Hugging Face implementations of GPT-Neo and RoBERTa models, performance impacts were as sessed through the BabyLM evaluation pipeline. The results indicate that activation functions do not significantly impact the performance of these models. Additionally, the model with the KAN network underperformed compared to models with traditional architectures in the context of this study. These findings suggest that optimizing activation functions may not be crucial for smaller language models, emphasizing the need for further research to explore other architectural improvements.