Few-Shot Domain Adaptation For End-to-End Communication

Authors: Jayaram Raghuram, Yijing Zeng, Dolores Garcia, Rafael Ruiz, Somesh Jha, Joerg Widmer, Suman Banerjee

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on many simulated distribution changes common to the wireless setting, and a real mm Wave FPGA testbed demonstrate the effectiveness of our method at adaptation using very few target domain samples.
Researcher Affiliation Collaboration Jayaram Raghuram 1, Yijing Zeng 1, Dolores García Martí 2, Rafael Ruiz Ortiz 2, Somesh Jha 1,3, Joerg Widmer 2, Suman Banerjee 1 1 University of Wisconsin Madison 2 IMDEA Networks Institute, Madrid 3 Xai Pient
Pseudocode Yes Algorithm 1 End-to-end training of the autoencoder with an MDN channel
Open Source Code Yes Code for our work: https://github.com/jayaram-r/domain-adaptation-autoencoder
Open Datasets No The paper describes generating simulated data based on common channel models and collecting data from a custom FPGA testbed. It does not provide access information (link, DOI, citation) to a publicly available or open dataset used for training.
Dataset Splits Yes We create a random class-stratified 50-50 train-test split (each of size 300,000) for data from both the source and target domains. Performance on both domains is always evaluated on the held-out test split. The train split from the target domain dataset is sub-sampled to create adaptation datasets of different sizes, specifically with 5, 10, 20, 30, 40, and 50 samples per class (symbol).
Hardware Specification Yes All the experiments were run on a Macbook Pro with 16 GB memory and 8 CPU cores. ... This ultra-wide-band mm-wave transceiver baseband memory-based design is developed on top of an ZCU111 RFSo C FPGA. This FPGA is equipped with 8 8 AD/DA converters with Giga-sampling capabilities, which make it ideal for RF system development; the 4 GB DDR4 memories contain RF-ADCs with up to 4 GSPS of sampling rate, and RF-DACs with up to 6.544 GSPS. This board also includes a quad-core ARM Cortex-A53 and a dual-core ARM Cortex-R5 real-time processor.
Software Dependencies No The paper mentions "Python using Tensor Flow (Abadi et al., 2015) and Tensor Flow Probability" but does not specify version numbers for these software components or any other libraries.
Experiment Setup Yes We used the following setting in our experiments. The size of the message set m is fixed to 16, corresponding to 4 bits. The dimension of the encoding (output of the encoder) d is set to 2, and the number of mixture components k is set to 5. ... The regularization constant λ in the adaptation objective was varied over 8 equally-spaced values on the log-scale with range 10 5 to 100, specifically {10 5, 10 4, 10 3, 10 2, 0.1, 1, 10, 100}. ... We used the Adam optimizer (Kingma & Ba, 2015) with a fixed learning rate of 0.001, batch size of 128, and 100 epochs for training the MDN. For adaptation of the MDN using the baseline methods Finetune and Finetune last, we used Adam with the same learning rate for 200 epochs. The batch size is set as b = max{10, 0.1 N t}, where N t is number of adaptation samples in the target dataset. For training the autoencoder using Algorithm 1, we found that stochastic gradient descent (SGD) with Nesterov momentum (constant 0.9), and an exponential learning rate schedule between 0.1 and 0.005 works better than Adam.