Can Mamba Learn How to Learn? A Comparative Study on In-Context Learning Tasks

State-space models (SSMs), such as Mamba Gu & Dao (2034), have been proposed as alternatives to Transformer networks in language modeling, by incorporating gating, convolutions, and input-dependent token selection to mitigate the quadratic cost of multi-head attention. Although SSMs exhibit competitive performance, their in-context learning (ICL) capabilities, a remarkable emergent property of modern language models that enables task execution without parameter optimization, remain underexplored compared to Transformers. In this study, we evaluate the ICL performance of SSMs, focusing on Mamba, against Transformer models across various tasks. Our results show that SSMs perform comparably to Transformers in standard regression ICL tasks, while outperforming them in tasks like sparse parity learning. However, SSMs fall short in tasks involving non-standard retrieval functionality. To address these limitations, we introduce a hybrid model, \variant, that combines Mamba with attention blocks, surpassing individual models in tasks where they struggle independently. Our findings suggest that hybrid architectures offer promising avenues for enhancing ICL in language models.

Strong Lottery Ticket Hypothesis with $\varepsilon$--perturbation

The strong Lottery Ticket Hypothesis (LTH) claims the existence of a subnetwork in a sufficiently large, randomly initialized neural network that approximates some target neural network without the need of training. We extend the theoretical guarantee of the strong LTH literature to a scenario more similar to the original LTH, by generalizing the weight change in the pre-training step to some perturbation around initialization. In particular, we focus on the following open questions: By allowing an $\varepsilon$-scale perturbation on the random initial weights, can we reduce the over-parameterization requirement for the candidate network in the strong LTH? Furthermore, does the weight change by SGD coincide with a good set of such perturbation? We answer the first question by first extending the theoretical result on subset sum to allow perturbation on the candidates. Applying this result to the neural network setting, we show that such $\varepsilon$-perturbation reduces the over-parameterization requirement of the strong LTH. To answer the second question, we show via experiments that the perturbed weight achieved by the projected SGD shows better performance under the strong LTH pruning.

Frame Difference-Based Temporal Loss for Video Stylization

Neural style transfer models have been used to stylize an ordinary video to specific styles. To ensure temporal inconsistency between the frames of the stylized video, a common approach is to estimate the optic flow of the pixels in the original video and make the generated pixels match the estimated optical flow. This is achieved by minimizing an optical flow-based (OFB) loss during model training. However, optical flow estimation is itself a challenging task, particularly in complex scenes. In addition, it incurs a high computational cost. We propose a much simpler temporal loss called the frame difference-based (FDB) loss to solve the temporal inconsistency problem. It is defined as the distance between the difference between the stylized frames and the difference between the original frames. The differences between the two frames are measured in both the pixel space and the feature space specified by the convolutional neural networks. A set of human behavior experiments involving 62 subjects with 25,600 votes showed that the performance of the proposed FDB loss matched that of the OFB loss. The performance was measured by subjective evaluation of stability and stylization quality of the generated videos on two typical video stylization models. The results suggest that the proposed FDB loss is a strong alternative to the commonly used OFB loss for video stylization.