Predictive coding fine-tuning enables computationally efficient domain adaptation for deep neural networks

Deep neural networks often struggle to maintain accuracy when deploying in real settings, as changes to input data that require variations in lighting and sensor drift, and continuous model adaptation. Matteo Cardoni and Sam Leroux of Ghent University and IMEC will address this challenge by developing new training methodologies that combine two established technologies, backpropagation and predictive coding strengths. Their approach initially uses backpropagation to train the network for strong initial performance, employing prediction code to adapt the model online, and recovering the accuracy lost due to shifting input data. This hybrid strategy provides a computationally efficient solution for continuous learning, demonstrating a critical step towards robust and reliable artificial intelligence systems in dynamic environments, particularly valuable for resource-constrained devices and future hardware accelerators.

This work proposes a hybrid training methodology that combines backpropagation with predictive coding to enable efficient on-device domain adaptation. This method starts with a deep neural network trained offline using backpropagation, achieves high initial performance and uses predictive coding for online adaptation to allow the model to recover the accuracy lost due to shifting input data distributions. This approach leverages the robustness of backpropagation for initial representation learning and the computational efficiency of predictive coding for subsequent improvements, ultimately allowing effective adaptation to changes in environmental conditions.

Predictive coding for image domain adaptation

Scientists investigated whether predictive coding could effectively adapt image classification models to new conditions and compared their performance with traditional backpropagation. This study focused on training the model on one dataset and testing it with a slightly modified version, simulating the actual scenario in which the input data is altered. The team experimented with a variety of network architectures, including simplified versions of VGG networks and fully connected multi-layer perceptrons, introducing different types of noise in their test images, including color inversion, rotating images, adding random noise, and assessing the robustness of the model. The researchers carefully adjusted key parameters for both backpropagation and predictive coding, such as learning rates, weight loss for normalization, and specific parameters that control the rate of updates in predictions.

They employed a systematic search to ensure stable training and identify the optimal settings that would prevent a significant drop in accuracy. The model was trained on 10 epochs with a batch size of 128 and applied data normalization and augmentation techniques to improve performance. The team normalized the CIFAR-10 dataset and applied techniques such as zero padding, random tripping, and horizontal inversion to increase the diversity of the training data. A detailed analysis revealed specific hyperparameters used for training with backpropagation and predictive coding under various noise conditions. This study provides valuable insight into the possibilities of predictive coding as a viable alternative to backpropagation, particularly when addressing the challenges posed by data distribution changes.

Hybrid Training adapts neural networks online

Scientists have developed a new hybrid training methodology that combines backpropagation and predictive coding to enable efficient on-device adaptation of deep neural networks. This task addresses the challenge of maintaining model performance in dynamic environments where input data distributions are shifted due to factors such as sensor drift and changes in lighting conditions. The team trained deep neural networks offline using backpropagation to achieve high initial accuracy, and adopted predictive coding for online adaptation, allowing the model to recover lost performance due to these data shifts. Experiments on the MNIST and CIFAR-10 datasets demonstrate the effectiveness of this approach.

Researchers achieved important results by leveraging the robustness of backpropagation for early learning and the computational efficiency of predictive coding for continuous adaptation. This combination allows for updates on devices without the need for extensive computing resources or communication with the cloud. This method relies on local computation and error-driven updates, and is well matched with the distributed nature of the new neural architecture. During the predictive coding phase, the input sample is applied to the first layer, and the last layer is set to the desired output.

The system then updates the layer activity, minimizing the energy function, which is the sum of the layer-wise prediction errors. This process allows the model to adapt to changes in the data distribution by minimizing discrepancies between predicted and actual activities at each layer. The results demonstrate a promising solution for maintaining model performance in real-time applications on energy-constrained edge devices, paving the way for more robust and adaptive artificial intelligence systems.

Hybrid learning adapts to data shifts

This study demonstrates efficient ways to combine backpropagation and predictive coding to adapt deep neural networks to changing environments. The team successfully demonstrated that models initially trained with backpropagation can effectively utilize online tweak predictive coding when faced with shifts in input data. This approach allows for continuous accuracy without the need for complete retraining, providing promising solutions for resource-constrained devices and future computing architectures. Experiments on standard datasets including MNIST and CIFAR-10 confirm the effectiveness of this hybrid strategy in maintaining performance under varying conditions.

This study acknowledges that adapting deeper predictive coding-based networks can present greater challenges, and that current implementations rely on monitored learning. Future work will focus on assessing the training time of embedded neural hardware, and extending this approach to more complex network architectures. Researchers also plan to investigate unsupervised, self-supervised learning approaches to broaden the applicability of this technique in real-world scenarios where labeled data is limited. This ongoing research aims to advance computationally efficient domain adaptation, particularly for deployment on specialized hardware platforms.