Low-rank adaptation with gating mechanisms in large language models, an improved method for fine-tuning: G-LoRA

The gating mechanism in typical models such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) effectively controls the flow of information in neural network structures. The Low-Rank Adaptation (LoRA) method involves incorporating two structures, A and B, alongside a pre-trained model. Typically, at the beginning of training, the parameters of these structures are initialized with Gaussian distribution and zeros, respectively. The output dimension of A and the input dimension of B are much smaller than the original model's input and output dimensions. Considering the characteristics of the gating mechanism, A, and B structures, a fusion is performed to achieve control over the information in A and B structures. Experimental results on large Transformer-based models show that, for the same hyperparameters, the LoRA structure with gating mechanism (G-LoRA) provides significant improvements in certain tasks.