PEMA: Plug-in External Memory Adaptation for Language Models

Pre-trained language models (PLMs) have demonstrated impressive performance across various downstream NLP tasks. Nevertheless, the resource requirements of pre-training large language models in terms of memory and training compute pose significant challenges. Furthermore, due to the substantial resources required, many PLM weights are confidential. Consequently, users are compelled to share their data with model owners for fine-tuning on specific tasks. To overcome the limitations, we introduce Plug-in External Memory Adaptation (PEMA), a Parameter-Efficient Fine-Tuning (PEFT) approach designed for fine-tuning PLMs without the need for all weights. PEMA can be integrated into the context representation of test data during inference to execute downstream tasks. It leverages an external memory to store context representations generated by a PLM, mapped with the desired target word. Our method entails training LoRA-based weight matrices within the final layer of the PLM for enhanced efficiency. The probability is then interpolated with the next-word distribution from the PLM to perform downstream tasks. To improve the generation quality, we propose a novel interpolation strategy named Gradual Unrolling. To demonstrate the effectiveness of our proposed method, we conduct experiments to demonstrate the efficacy of PEMA with a syntactic dataset and assess its performance on machine translation and style transfer tasks using real datasets. PEMA outperforms other PEFT methods in terms of memory and latency efficiency for training and inference. Furthermore, it outperforms other baselines in preserving the meaning of sentences while generating appropriate language and styles.

CTMQ: Cyclic Training of Convolutional Neural Networks with Multiple Quantization Steps

This paper proposes a training method having multiple cyclic training for achieving enhanced performance in low-bit quantized convolutional neural networks (CNNs). Quantization is a popular method for obtaining lightweight CNNs, where the initialization with a pretrained model is widely used to overcome degraded performance in low-resolution quantization. However, large quantization errors between real values and their low-bit quantized ones cause difficulties in achieving acceptable performance for complex networks and large datasets. The proposed training method softly delivers the knowledge of pretrained models to low-bit quantized models in multiple quantization steps. In each quantization step, the trained weights of a model are used to initialize the weights of the next model with the quantization bit depth reduced by one. With small change of the quantization bit depth, the performance gap can be bridged, thus providing better weight initialization. In cyclic training, after training a low-bit quantized model, its trained weights are used in the initialization of its accurate model to be trained. By using better training ability of the accurate model in an iterative manner, the proposed method can produce enhanced trained weights for the low-bit quantized model in each cycle. Notably, the training method can advance Top-1 and Top-5 accuracies of the binarized ResNet-18 on the ImageNet dataset by 5.80% and 6.85%, respectively.

PLAM: a Posit Logarithm-Approximate Multiplier for Power Efficient Posit-based DNNs

The Posit Number System was introduced in 2017 as a replacement for floating-point numbers. Since then, the community has explored its application in Neural Network related tasks and produced some unit designs which are still far from being competitive with their floating-point counterparts. This paper proposes a Posit Logarithm-Approximate Multiplication (PLAM) scheme to significantly reduce the complexity of posit multipliers, the most power-hungry units within Deep Neural Network architectures. When comparing with state-of-the-art posit multipliers, experiments show that the proposed technique reduces the area, power, and delay of hardware multipliers up to 72.86%, 81.79%, and 17.01%, respectively, without accuracy degradation.

Effects of Approximate Multiplication on Convolutional Neural Networks

This paper analyzes the effects of approximate multiplication when performing inferences on deep convolutional neural networks (CNNs). The approximate multiplication can reduce the cost of underlying circuits so that CNN inferences can be performed more efficiently in hardware accelerators. The study identifies the critical factors in the convolution, fully-connected, and batch normalization layers that allow more accurate CNN predictions despite the errors from approximate multiplication. The same factors also provide an arithmetic explanation of why bfloat16 multiplication performs well on CNNs. The experiments are performed with recognized network architectures to show that the approximate multipliers can produce predictions that are nearly as accurate as the FP32 references, without additional training. For example, the ResNet and Inception-v4 models with Mitch-$w$6 multiplication produces Top-5 errors that are within 0.2% compared to the FP32 references. A brief cost comparison of Mitch-$w$6 against bfloat16 is presented, where a MAC operation saves up to 80% of energy compared to the bfloat16 arithmetic. The most far-reaching contribution of this paper is the analytical justification that multiplications can be approximated while additions need to be exact in CNN MAC operations.