Hongliang Li

Shaoxu Cheng, Kanglei Geng, Chiyuan He et al.

Continual Learning (CL) aims to enable Deep Neural Networks (DNNs) to learn new data without forgetting previously learned knowledge. The key to achieving this goal is to avoid confusion at the feature level, i.e., avoiding confusion within old tasks and between new and old tasks. Previous prototype-based CL methods generate pseudo features for old knowledge replay by adding Gaussian noise to the centroids of old classes. However, the distribution in the feature space exhibits anisotropy during the incremental process, which prevents the pseudo features from faithfully reproducing the distribution of old knowledge in the feature space, leading to confusion in classification boundaries within old tasks. To address this issue, we propose the Distribution-Level Memory Recall (DMR) method, which uses a Gaussian mixture model to precisely fit the feature distribution of old knowledge at the distribution level and generate pseudo features in the next stage. Furthermore, resistance to confusion at the distribution level is also crucial for multimodal learning, as the problem of multimodal imbalance results in significant differences in feature responses between different modalities, exacerbating confusion within old tasks in prototype-based CL methods. Therefore, we mitigate the multi-modal imbalance problem by using the Inter-modal Guidance and Intra-modal Mining (IGIM) method to guide weaker modalities with prior information from dominant modalities and further explore useful information within modalities. For the second key, We propose the Confusion Index to quantitatively describe a model's ability to distinguish between new and old tasks, and we use the Incremental Mixup Feature Enhancement (IMFE) method to enhance pseudo features with new sample features, alleviating classification confusion between new and old knowledge.

Ahmad Rashid, Ruotian Wu, Rongqi Fan et al.

Reward-guided text generation (RGTG) has emerged as a viable alternative to offline reinforcement learning from human feedback (RLHF). RGTG methods can align baseline language models to human preferences without further training like in standard RLHF methods. However, they rely on a reward model to score each candidate token generated by the language model at inference, incurring significant test-time overhead. Additionally, the reward model is usually only trained to score full sequences, which can lead to sub-optimal choices for partial sequences. In this work, we present a novel reward model architecture that is trained, using a Bradley-Terry loss, to prefer the optimal expansion of a sequence with just a \emph{single call} to the reward model at each step of the generation process. That is, a score for all possible candidate tokens is generated simultaneously, leading to efficient inference. We theoretically analyze various RGTG reward models and demonstrate that prior techniques prefer sub-optimal sequences compared to our method during inference. Empirically, our reward model leads to significantly faster inference than other RGTG methods. It requires fewer calls to the reward model and performs competitively compared to previous RGTG and offline RLHF methods.