Figure 8: The average values during training of the two components used in the criteria for neuron importance in the input layer: the absolute gradient of the loss with respect to the reconstructed samples and the sum of the absolute weights connected to a neuron. A.1 Implementation Details For all datasets, we used standard normalization that scales the features to have zero mean and standard deviation of one. The architecture of the autoencoder consists of one hidden layer with sigmoid activation. A linear activation is used for the output layer. We use a hidden layer of 200 neurons for all datasets.

Training models with longer in-context lengths is a significant challenge for multimodal machine learning due to substantial GPU memory and computational costs. This exploratory study does not present state-of-the-art models; rather, it introduces an innovative method designed to increase in-context text length in multi-modality large language models (MLLMs) efficiently. We present \ModelFullName (\ModelName), which processes long in-context text using visual tokens. This technique significantly reduces GPU memory usage and floating point operations (FLOPs). For instance, our method expands the pre-training in-context length from 256 to 2048 tokens with fewer FLOPs for a 56 billion parameter MOE model. Experimental results demonstrate that \ModelName enhances OCR capabilities and delivers superior performance on common downstream benchmarks for in-context few-shot evaluation. Additionally, \ModelName proves effective for long context inference, achieving results comparable to full text input while maintaining computational efficiency.

State-of-the-art results in large language models (LLMs) often rely on scale, whichbecomes computationally expensive. This has sparked a research agenda to reducethese models' parameter counts and computational costs without significantlyimpacting their performance. Our study focuses on transformer-based LLMs,specifically targeting the computationally intensive feedforward networks (FFNs),which are less studied than attention blocks. We consider three structured linearparameterizations of the FFN using efficient low-rank and block-diagonal matrices.In contrast to many previous works that examined these approximations, our studyi) explores these structures from a training-from-scratch perspective, ii) scales upto 1.3B parameters, and iii) is conducted within recent Transformer-based LLMsrather than convolutional architectures. We demonstrate that these structures canlead to actual computational gains in various scenarios, including online decodingwhen using a pre-merge technique. Additionally, we propose a novel trainingregime, called self-guided training, aimed at improving the poor training dynamicsthat these approximations exhibit when used from initialization. Interestingly,the scaling performance of structured matrices is explored, revealing steepercurves in scaling training FLOPs, along with a favorable scaling trend in theovertraining regime. Specifically, we show that wide and structured networkscan utilize training FLOPs more efficiently, with fewer parameters and lowerloss than dense models at their optimal trade-off.

By combining a caching algorithm, namely Greedy Dual Size with Frequency (GDSF) or Least Expected Cost (LEC), with a model multiplexer, we achieve optimal rates in both offline and online settings.