Be Like Water: Adaptive Floating Point for Machine Learning

Authors: Thomas Yeh, Max Sterner, Zerlina Lai, Brandon Chuang, Alexander Ihler

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate AFP on a spectrum of representative models in computer vision and NLP, and show that our technique enables ultra-low precision inference of deep learning models while providing accuracy comparable to full precision inference. We build a simulation infrastructure in Tensorflow to accurately model the numerical effects of applying AFP to the weights and layer outputs of ML models. We perform comprehensive simulations of AFP on a wide range of robust CNN and Transformer models.
Researcher Affiliation Academia 1Computer Science Department, Pomona College, Claremont, CA, USA 2Computer Science Department, Occidental College, Los Angeles, CA, USA 3Computer Science Department, University of California, Santa Cruz, CA, USA 4Department of Computer Science, University of California, Irvine, CA, USA.
Pseudocode No The paper describes the design of AFP and its hardware implementation but does not include formal pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statements or links indicating that the source code for the methodology is openly available.
Open Datasets Yes For image classification, the test dataset of Image Net V2 (Recht et al., 2019) was used to measure model accuracy.
Dataset Splits No The paper mentions using a "test dataset" but does not provide specific details on training, validation, and test splits (e.g., percentages, sample counts, or explicit references to predefined splits for all sets).
Hardware Specification No The paper refers to types of ML accelerators (e.g., "Google s TPU", "Nvidia Tensor Cores") in a general context but does not specify the exact hardware (e.g., GPU/CPU models, memory) used for their experiments.
Software Dependencies No To simulate AFP in hardware, several DNN inference models were employed in Tensorflow using the Keras (Chollet et al., 2015) and Hugging Face (Wolf et al., 2019) libraries. However, specific version numbers for these software components are not provided.
Experiment Setup Yes Using a custom round function, all types of layers weights and outputs were rounded with AFP, such as Conv2D, Batch Normalization, and Dense layers. To properly simulate inference using AFP in hardware, all the weights were rounded when instantiating the model and all layer outputs were rounded between every layer, before being input into the next layer. Examples from data sets were individually input into each model and a block size of 16 was used for most experiments.