Pre-training with self-supervised models, such as Hidden-unit BERT (HuBERT) and wav2vec 2.0, has brought significant improvements in automatic speech recognition (ASR). However, these models usually require an expensive computational cost to achieve outstanding performance, slowing down the inference speed. To improve the model efficiency, we introduce an early exit scheme for ASR, namely HuBERT-EE, that allows the model to stop the inference dynamically. In HuBERT-EE, multiple early exit branches are added at the intermediate layers. When the intermediate prediction of the early exit branch is confident, the model stops the inference, and the corresponding result can be returned early. We investigate the proper early exiting criterion and fine-tuning strategy to effectively perform early exiting. Experimental results on the LibriSpeech show that HuBERT-EE can accelerate the inference of the HuBERT while simultaneously balancing the trade-off between the performance and the latency.

Self-supervised speech models have shown to be useful for various tasks, but their large size limits the use in devices with low computing power and memory. In this work, we explore early exit, an approach for reducing latency by exiting the forward process of a network early. Most approaches of early exit need a separate early exit model for each task, with some even requiring fine-tuning of the entire pretrained model. We introduce Data Adaptive Self-Supervised Early Exit (DAISY), an approach that decides when to exit based on the self-supervised loss, eliminating the need for multiple round of training and fine-tuning. DAISY matches the performance of HuBERT on the MiniSUPERB benchmark, but with much faster inference times. Our analysis on the adaptivity of DAISY shows that the model exits early (using fewer layers) on clean data while exits late (using more layers) on noisy data, dynamically adjusting the computational cost of inference based on the noise level of each sample.

Batteryless systems frequently face power failures, requiring extra runtime buffers to maintain inference progress and leaving only a memory space for storing ultra-tiny deep neural networks (DNNs). Besides, making these models responsive to stochastic energy harvesting dynamics during inference requires a balance between inference accuracy, latency, and energy overhead. Recent works on compression mostly focus on time and memory, but often ignore energy dynamics or significantly reduce the accuracy of pre-trained DNNs. Existing energy-adaptive inference works modify the architecture of pre-trained models and have significant memory overhead. Thus, energy-adaptive and accurate inference of pre-trained DNNs on batteryless devices with extreme memory constraints is more challenging than traditional microcontrollers. We combat these issues by proposing FreeML, a framework to optimize pre-trained DNN models for memory-efficient and energy-adaptive inference on batteryless systems. FreeML comprises (1) a novel compression technique to reduce the model footprint and runtime memory requirements simultaneously, making them executable on extremely memory-constrained batteryless platforms; and (2) the first early exit mechanism that uses a single exit branch for all exit points to terminate inference at any time, making models energy-adaptive with minimal memory overhead. Our experiments showed that FreeML reduces the model sizes by up to $95 \times$, supports adaptive inference with a $2.03-19.65 \times$ less memory overhead, and provides significant time and energy benefits with only a negligible accuracy drop compared to the state-of-the-art.

Everyday models get heavier and heavier (in terms of learnable parameters). For example, LEMON_large has 200M parameters and GPT-3 has over 175 billion parameters! Though they give State-of-the-Art Performance, how well are they deployed today? This calls for an efficient and faster method for training and inferring. So, we explore various methods through which we can speed up compute-intensive networks while preserving accuracy!