An official website of the United States government
Nowadays, one practical limitation of deep neural network (DNN) is its high degree of specialization to a single task or domain (e.g., one visual domain). It motivates researchers to develop algorithms that can adapt DNN model to multiple domains sequentially, while still performing well on the past domains, which is known as multi-domain learning. Almost all conventional methods only focus on improving accuracy with minimal parameter update, while ignoring high computing and memory cost during training, which makes it difficult to deploy multi-domain learning into more and more widely used resource-limited edge devices, like mobile phone, IoT, embedded system, etc. During our study in multi-domain training process, we observe that large memory used for activation storage is the bottleneck that largely limits the training time and cost on edge devices. To reduce training memory usage, while keeping the domain adaption accuracy performance, we propose Dynamic Additive Attention Adaption (DA3), a novel memory-efficient on-device multi-domain learning method. DA3 learns a novel additive attention adaptor module, while freezing the weights of the pre-trained backbone model for each domain. Differentiating from prior works, such module not only mitigates activation memory buffering for reducing memory usage during training, but also serves as a dynamic gating mechanism to reduce the computation cost for fast inference. We validate DA3 on multiple datasets against state-of-the-art methods, which shows great improvement in both accuracy and training time. Moreover, we deployed DA3 into the popular NIVDIA Jetson Nano edge GPU, where the measured experimental results show our proposed \mldam reduces the on-device training memory consumption by 19x-37x, and training time by 2x, in comparison to the baseline methods (e.g., standard fine-tuning, Parallel and Series Res. adaptor, and Piggyback).
more »
« less
MobileTL: On-Device Transfer Learning with Inverted Residual Blocks
Transfer learning on edge is challenging due to on-device limited resources. Existing work addresses this issue by training a subset of parameters or adding model patches. Developed with inference in mind, Inverted Residual Blocks (IRBs) split a convolutional layer into depthwise and pointwise convolutions, leading to more stacking layers, e.g., convolution, normalization, and activation layers. Though they are efficient for inference, IRBs require that additional activation maps are stored in memory for training weights for convolution layers and scales for normalization layers. As a result, their high memory cost prohibits training IRBs on resource-limited edge devices, and making them unsuitable in the context of transfer learning. To address this issue, we present MobileTL, a memory and computationally efficient on-device transfer learning method for models built with IRBs. MobileTL trains the shifts for internal normalization layers to avoid storing activation maps for the backward pass. Also, MobileTL approximates the backward computation of the activation layer (e.g., Hard-Swish and ReLU6) as a signed function which enables storing a binary mask instead of activation maps for the backward pass. MobileTL fine-tunes a few top blocks (close to output) rather than propagating the gradient through the whole network to reduce the computation cost. Our method reduces memory usage by 46% and 53% for MobileNetV2 and V3 IRBs, respectively. For MobileNetV3, we observe a 36% reduction in floating-point operations (FLOPs) when fine-tuning 5 blocks, while only incurring a 0.6% accuracy reduction on CIFAR10. Extensive experiments on multiple datasets demonstrate that our method is Pareto-optimal (best accuracy under given hardware constraints) compared to prior work in transfer learning for edge devices.
more »
« less
- Award ID(s):
- 2107085
- PAR ID:
- 10461950
- Date Published:
- Journal Name:
- Proceedings of the AAAI Conference on Artificial Intelligence
- Volume:
- 37
- Issue:
- 6
- ISSN:
- 2159-5399
- Page Range / eLocation ID:
- 7166 to 7174
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Conventionally, DNN models are trained once in the cloud and deployed in edge devices such as cars, robots, or unmanned aerial vehicles (UAVs) for real-time inference. However, there are many cases that require the models to adapt to new environments, domains, or users. In order to realize such domain adaption or personalization, the models on devices need to be continuously trained on the device. In this work, we design EF-Train, an efficient DNN training accelerator with a unified channel-level parallelism-based convolution kernel that can achieve end-to-end training on resource-limited low-power edge-level FPGAs. It is challenging to implement on-device training on resource-limited FPGAs due to the low efficiency caused by different memory access patterns among forward and backward propagation and weight update. Therefore, we developed a data reshaping approach with intra-tile continuous memory allocation and weight reuse. An analytical model is established to automatically schedule computation and memory resources to achieve high energy efficiency on edge FPGAs. The experimental results show that our design achieves 46.99 GFLOPS and 6.09 GFLOPS/W in terms of throughput and energy efficiency, respectively.more » « less
-
Due to the separate memory and computation units in traditional Von-Neumann architecture, massive data transfer dominates the overall computing system’s power and latency, known as the ‘Memory-Wall’ issue. Especially with ever-increasing deep learning-based AI model size and computing complexity, it becomes the bottleneck for state-of-the-art AI computing systems. To address this challenge, In-Memory Computing (IMC) based Neural Network accelerators have been widely investigated to support AI computing within memory. However, most of those works focus only on inference. The on-device training and continual learning have not been well explored yet. In this work, for the first time, we introduce on-device continual learning with STT-assisted-SOT (SAS) Magnetic Random Access Memory (MRAM) based IMC system. On the hardware side, we have fabricated a SAS-MRAM device prototype with 4 Magnetic Tunnel Junctions (MTJ, each at 100nm × 50nm) sharing a common heavy metal layer, achieving significantly improved memory writing and area efficiency compared to traditional SOT-MRAM. Next, we designed fully digital IMC circuits with our SAS-MRAM to support both neural network inference and on-device learning. To enable efficient on-device continual learning for new task data, we present an 8-bit integer (INT8) based continual learning algorithm that utilizes our SAS-MRAM IMC-supported bit-serial digital in-memory convolution operations to train a small parallel reprogramming Network (Rep-Net) while freezing the major backbone model. Extensive studies have been presented based on our fabricated SAS-MRAM device prototype, cross-layer device-circuit benchmarking and simulation, as well as the on-device continual learning system evaluation.more » « less
-
We propose SLOPE, a Double-Pruned Sparse Plus Lazy Low-rank Adapter Pretraining method for LLMs that improves the accuracy of sparse LLMs while accelerating their pretraining and inference and reducing their memory footprint. Sparse pretraining of LLMs reduces the accuracy of the model, to overcome this, prior work uses dense models during fine-tuning. SLOPE improves the accuracy of sparsely pretrained models by adding low-rank adapters in the final 1% iterations of pretraining without adding significant overheads to the model pretraining and inference. In addition, SLOPE uses a double-pruned backward pass formulation that prunes the transposed weight matrix using N:M sparsity structures to enable an accelerated sparse backward pass. SLOPE accelerates the training and inference of models with billions of parameters up to 1.25→ and 1.54→ respectively (OPT-33B and OPT-66B) while reducing their memory usage by up to 0.63→ and 0.61→ for training and inference respectively.more » « less
-
The growing computational demands of deep learning have driven interest in analog neural networks using resistive memory and silicon photonics. However, these technologies face inherent limitations in computing parallelism when used independently. Photonic phase-change memory (PCM), which integrates photonics with PCM, overcomes these constraints by enabling simultaneous processing of multiple inputs encoded on different wavelengths, significantly enhancing parallel computation for deep neural network (DNN) inference and training. This paper presents MERIT, a sustainable DNN accelerator that capitalizes on the non-volatility of resistive memory and the high operating speed of photonic devices. MERIT enables seamless inference and training by loading weight kernels into photonic PCM arrays and selectively supplying light encoded with input features for the forward pass and loss gradients for the backward pass. We compare MERIT with state-of-the-art digital and analog DNN accelerators including TPU, DEAP, and PTC. Simulation results demonstrate that MERIT reduces execution time by 68% and energy consumption by 64% for inference, and reduces execution time by 79% and energy consumption by 84% for training.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.1609/aaai.v37i6.25874
