However, conventional training methods based on surrogate gradients and Backpropagation Through Time (BPTT) not only lag behind Artificial Neural Networks (ANNs) in performance, but also incur significant computational and memory overheads that grow linearly with the temporal dimension. To enable high-performance SNN training under limited computational resources, we propose an enhanced self-distillation framework, jointly optimized with rate-based backpropagation. Specifically, the firing rates of intermediate SNN layers are projected onto lightweight ANN branches, and high-quality knowledge generated by the model itself is used to optimize substructures through the ANN pathways. Unlike traditional self-distillation paradigms, we observe that low-quality self-generated knowledge may hinder convergence. To address this, we decouple the teacher signal into reliable and unreliable components, ensuring that only reliable knowledge is used to guide the optimization of the model. Extensive experiments on CIFAR-10, CIFAR-100, CIFAR10-DVS, and ImageNet demonstrate that our method reduces training complexity while achieving high-performance SNN training.

Knowledge Distillation (KD) is a powerful approach for compressing a large model into a smaller, more efficient model, particularly beneficial for latency-sensitive applications like recommender systems. However, current KD research predominantly focuses on Computer Vision (CV) and NLP tasks, overlooking unique data characteristics and challenges inherent to recommender systems. This paper addresses these overlooked challenges, specifically: (1) mitigating data distribution shifts between teacher and student models, (2) efficiently identifying optimal teacher configurations within time and budgetary constraints, and (3) enabling computationally efficient and rapid sharing of teacher labels to support multiple students. We present a robust KD system developed and rigorously evaluated on multiple large-scale personalized video recommendation systems within Google. Our live experiment results demonstrate significant improvements in student model performance while ensuring consistent and reliable generation of high quality teacher labels from a continuous data stream of data.

Knowledge distillation (KD) has been widely employed to transfer knowledge from a large language model (LLM) to a specialized model in low-data regimes through pseudo label learning. However, pseudo labels generated by teacher models are usually noisy and may influence KD performance. This study delves into KD with noisy teachers and uncovers that the student model can already generate more accurate predictions than the teacher labels used to train it during KD, indicating its inherent ability to denoise noisy teacher labels. Motivated by this finding, we propose Peer-Advised KD to improve vanilla KD from noisy teachers. Experiments show that Peer-Advised KD can outperform LLM by approximately 5% with 50 human-labeled data, and even competitive to standard supervised finetuning with 750 human-labeled data.

While deep learning has achieved remarkable success, there is no clear explanation about why it works so well. In order to discuss this question quantitatively, we need a mathematical framework that explains what learning is in the first place. After several considerations, we succeeded in constructing a mathematical framework that can provide a unified understanding of all types of learning, including deep learning and learning in the brain. We call it learning principle, and it follows that all learning is equivalent to estimating the probability of input data. We not only derived this principle, but also mentioned its application to actual machine learning models. For example, we found that conventional supervised learning is equivalent to estimating conditional probabilities, and succeeded in making supervised learning more effective and generalized. We also proposed a new method of defining the values of estimated probability using differentiation, and showed that unsupervised learning can be performed on arbitrary dataset without any prior knowledge. Namely, this method is a general-purpose machine learning in the true sense. Moreover, we succeeded in describing the learning mechanism in the brain by considering the time evolution of a fully or partially connected model and applying this new method. The learning principle provides solutions to many unsolved problems in deep learning and cognitive neuroscience.

Moreover, with the ever-increasing transmission bandwidth, nonlinearity becomes even more important [1]. Various digital signal processing (DSP) techniques have been proposed to minimize nonlinear effects [2]. Due to the universal approximation capability of neural networks (NNs), the NNs have recently been intensively studied for the optical channel post-equalization, because they can approximate the inverse optical channel transfer function with good accuracy and revert the nonlinear distortions. In particular, recurrent NNs (RNN) based equalizers have shown the best capability in equalizing nonlinear impairments as compared to the feed-forward NN types [3-5]. However, since the RNN structure has a feedback loop, it is not easily parallelizable.