Resource-efficient Medical Image Analysis with Self-adapting Forward-Forward Networks

TL;DR

The paper tackles the high energy and resource demands of backpropagation-based training in medical imaging by introducing the Convolutional Forward-Forward Algorithm (CFFA) and the self-adapting SaFF-Net, a BP-free framework with automatic hyperparameter and architectural adaptation during warm-up and training. It extends the Forward-Forward Algorithm to CNNs, using a goodness function $g(x)$ and contrastive loss $F(x^{+},x^{-})$ with a trainable threshold $ heta$, and emphasizes on-device viability through early stopping, pruning, and normalization techniques. Empirical results on MNIST, MedMNIST, and VinDr-CXR demonstrate competitive performance with far fewer parameters and function evaluations, especially in one-shot and large-batch regimes, with ablations validating the impact of key components. The work offers a path toward energy-efficient, accessible medical imaging models suitable for deployment in low-resource clinical settings, potentially reducing disparities in healthcare technology access.

Abstract

We introduce a fast Self-adapting Forward-Forward Network (SaFF-Net) for medical imaging analysis, mitigating power consumption and resource limitations, which currently primarily stem from the prevalent reliance on back-propagation for model training and fine-tuning. Building upon the recently proposed Forward-Forward Algorithm (FFA), we introduce the Convolutional Forward-Forward Algorithm (CFFA), a parameter-efficient reformulation that is suitable for advanced image analysis and overcomes the speed and generalisation constraints of the original FFA. To address hyper-parameter sensitivity of FFAs we are also introducing a self-adapting framework SaFF-Net fine-tuning parameters during warmup and training in parallel. Our approach enables more effective model training and eliminates the previously essential requirement for an arbitrarily chosen Goodness function in FFA. We evaluate our approach on several benchmarking datasets in comparison with standard Back-Propagation (BP) neural networks showing that FFA-based networks with notably fewer parameters and function evaluations can compete with standard models, especially, in one-shot scenarios and large batch sizes. The code will be available at the time of the conference.

Figures (3)

• Figure 1: The Forward-Forward Multi-Layer Perceptron (left), the Forward-Forward Convolutional Neural Network (right). The networks are optimised layer-wise. The positive and negative samples are fed into the first layer and via layer normalisation, we obtain orientation and length of the activation vector. The orientation is forwarded to the next layer as its input. The length is used for the computation of the goodness. Each layer is optimised so that positive samples have high goodness (> threshold) and negative samples have low goodness (< threshold). For inference, the sum of the goodness of all layers, excluding the first layer, needs to be determined for every possible label.
• Figure 2: Classification on MNIST. ACC - Accuracy, AUC - Area Under the Receiver Operating characteristic, mAP - Mean Average Precision vs. Number of Parameters Comparison for MLP and FFA (top) and CNN and CFFA (bottom) with maximum batch size. Ours in orange.
• Figure 3: Self-adapting SaFF-Net framework. Key features of the training data and the hardware components are used for self-configuration. The fixed parameters for the pipeline are given by default or via an experiment file. After self-configuration, the SaFF-Net selects the best network configuration and starts training. Inference, postprocessing, calibration and pruning can be enabled.