Rethinking the Bias of Foundation Model under Long-tailed Distribution

Long-tailed learning has garnered increasing attention due to its practical significance. Among the various approaches, the fine-tuning paradigm has gained considerable interest with the advent of foundation models. However, most existing methods primarily focus on leveraging knowledge from these models, overlooking the inherent biases introduced by the imbalanced training data they rely on. In this paper, we examine how such imbalances from pre-training affect long-tailed downstream tasks. Specifically, we find the imbalance biases inherited in foundation models on downstream task as parameter imbalance and data imbalance. During fine-tuning, we observe that parameter imbalance plays a more critical role, while data imbalance can be mitigated using existing re-balancing strategies. Moreover, we find that parameter imbalance cannot be effectively addressed by current re-balancing techniques, such as adjusting the logits, during training, unlike data imbalance. To tackle both imbalances simultaneously, we build our method on causal learning and view the incomplete semantic factor as the confounder, which brings spurious correlations between input samples and labels. To resolve the negative effects of this, we propose a novel backdoor adjustment method that learns the true causal effect between input samples and labels, rather than merely fitting the correlations in the data. Notably, we achieve an average performance increase of about 1.67% on each dataset.

Many real-world datasets are imbalanced — some categories have far more examples than others. This makes it harder for machine learning models to perform well on the less common, or "tail," categories. A popular way to address this is to start with a large, pre-trained “foundation model” and fine-tune it for a specific task. But these foundation models themselves are often trained on imbalanced data, introducing hidden biases that current methods fail to fully address.In our study, we investigate how these pre-existing imbalances affect performance during fine-tuning. We identify two types of inherited bias: parameter imbalance (how the model’s internal settings are skewed) and data imbalance (how the training data is distributed). We find that while current methods can partially fix data imbalance, they struggle to handle parameter imbalance effectively.To solve this, we use a technique called causal learning. Our method treats missing or misleading information as a “confounder” and uses a strategy known as backdoor adjustment to learn the real cause-effect relationships — not just surface-level patterns. Our approach improves performance across several datasets by an average of 1.67%.