Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations

Large-scale recommendation systems are characterized by their reliance on high cardinality, heterogeneous features and the need to handle tens of billions of user actions on a daily basis. Despite being trained on huge volume of data with thousands of features, most Deep Learning Recommendation Models (DLRMs) in industry fail to scale with compute. Inspired by success achieved by Transformers in language and vision domains, we revisit fundamental design choices in recommendation systems. We reformulate recommendation problems as sequential transduction tasks within a generative modeling framework (``Generative Recommenders''), and propose a new architecture, HSTU, designed for high cardinality, non-stationary streaming recommendation data. HSTU outperforms baselines over synthetic and public datasets by up to 65.8\% in NDCG, and is 5.3x to 15.2x faster than FlashAttention2-based Transformers on 8192 length sequences. HSTU-based Generative Recommenders, with 1.5 trillion parameters, improve metrics in online A/B tests by 12.4\% and have been deployed on multiple surfaces of a large internet platform with billions of users. More importantly, the model quality of Generative Recommenders empirically scales as a power-law of training compute across three orders of magnitude, up to GPT-3/LLaMa-2 scale, which reduces carbon footprint needed for future model developments, and further paves the way for the first foundational models in recommendations.

MindGPT: Interpreting What You See with Non-invasive Brain Recordings

Decoding of seen visual contents with non-invasive brain recordings has important scientific and practical values. Efforts have been made to recover the seen images from brain signals. However, most existing approaches cannot faithfully reflect the visual contents due to insufficient image quality or semantic mismatches. Compared with reconstructing pixel-level visual images, speaking is a more efficient and effective way to explain visual information. Here we introduce a non-invasive neural decoder, termed as MindGPT, which interprets perceived visual stimuli into natural languages from fMRI signals. Specifically, our model builds upon a visually guided neural encoder with a cross-attention mechanism, which permits us to guide latent neural representations towards a desired language semantic direction in an end-to-end manner by the collaborative use of the large language model GPT. By doing so, we found that the neural representations of the MindGPT are explainable, which can be used to evaluate the contributions of visual properties to language semantics. Our experiments show that the generated word sequences truthfully represented the visual information (with essential details) conveyed in the seen stimuli. The results also suggested that with respect to language decoding tasks, the higher visual cortex (HVC) is more semantically informative than the lower visual cortex (LVC), and using only the HVC can recover most of the semantic information. The code of the MindGPT model will be publicly available at https://github.com/JxuanC/MindGPT.

A Human-Machine Joint Learning Framework to Boost Endogenous BCI Training

Brain-computer interfaces (BCIs) provide a direct pathway from the brain to external devices and have demonstrated great potential for assistive and rehabilitation technologies. Endogenous BCIs based on electroencephalogram (EEG) signals, such as motor imagery (MI) BCIs, can provide some level of control. However, mastering spontaneous BCI control requires the users to generate discriminative and stable brain signal patterns by imagery, which is challenging and is usually achieved over a very long training time (weeks/months). Here, we propose a human-machine joint learning framework to boost the learning process in endogenous BCIs, by guiding the user to generate brain signals towards an optimal distribution estimated by the decoder, given the historical brain signals of the user. To this end, we firstly model the human-machine joint learning process in a uniform formulation. Then a human-machine joint learning framework is proposed: 1) for the human side, we model the learning process in a sequential trial-and-error scenario and propose a novel ``copy/new'' feedback paradigm to help shape the signal generation of the subject toward the optimal distribution; 2) for the machine side, we propose a novel adaptive learning algorithm to learn an optimal signal distribution along with the subject's learning process. Specifically, the decoder reweighs the brain signals generated by the subject to focus more on ``good'' samples to cope with the learning process of the subject. Online and psuedo-online BCI experiments with 18 healthy subjects demonstrated the advantages of the proposed joint learning process over co-adaptive approaches in both learning efficiency and effectiveness.

Revisiting Neural Retrieval on Accelerators

Retrieval finds a small number of relevant candidates from a large corpus for information retrieval and recommendation applications. A key component of retrieval is to model (user, item) similarity, which is commonly represented as the dot product of two learned embeddings. This formulation permits efficient inference, commonly known as Maximum Inner Product Search (MIPS). Despite its popularity, dot products cannot capture complex user-item interactions, which are multifaceted and likely high rank. We hence examine non-dot-product retrieval settings on accelerators, and propose \textit{mixture of logits} (MoL), which models (user, item) similarity as an adaptive composition of elementary similarity functions. This new formulation is expressive, capable of modeling high rank (user, item) interactions, and further generalizes to the long tail. When combined with a hierarchical retrieval strategy, \textit{h-indexer}, we are able to scale up MoL to 100M corpus on a single GPU with latency comparable to MIPS baselines. On public datasets, our approach leads to uplifts of up to 77.3\% in hit rate (HR). Experiments on a large recommendation surface at Meta showed strong metric gains and reduced popularity bias, validating the proposed approach's performance and improved generalization.

ESL-SNNs: An Evolutionary Structure Learning Strategy for Spiking Neural Networks

Spiking neural networks (SNNs) have manifested remarkable advantages in power consumption and event-driven property during the inference process. To take full advantage of low power consumption and improve the efficiency of these models further, the pruning methods have been explored to find sparse SNNs without redundancy connections after training. However, parameter redundancy still hinders the efficiency of SNNs during training. In the human brain, the rewiring process of neural networks is highly dynamic, while synaptic connections maintain relatively sparse during brain development. Inspired by this, here we propose an efficient evolutionary structure learning (ESL) framework for SNNs, named ESL-SNNs, to implement the sparse SNN training from scratch. The pruning and regeneration of synaptic connections in SNNs evolve dynamically during learning, yet keep the structural sparsity at a certain level. As a result, the ESL-SNNs can search for optimal sparse connectivity by exploring all possible parameters across time. Our experiments show that the proposed ESL-SNNs framework is able to learn SNNs with sparse structures effectively while reducing the limited accuracy. The ESL-SNNs achieve merely 0.28% accuracy loss with 10% connection density on the DVS-Cifar10 dataset. Our work presents a brand-new approach for sparse training of SNNs from scratch with biologically plausible evolutionary mechanisms, closing the gap in the expressibility between sparse training and dense training. Hence, it has great potential for SNN lightweight training and inference with low power consumption and small memory usage.

LaSNN: Layer-wise ANN-to-SNN Distillation for Effective and Efficient Training in Deep Spiking Neural Networks

Spiking Neural Networks (SNNs) are biologically realistic and practically promising in low-power computation because of their event-driven mechanism. Usually, the training of SNNs suffers accuracy loss on various tasks, yielding an inferior performance compared with ANNs. A conversion scheme is proposed to obtain competitive accuracy by mapping trained ANNs' parameters to SNNs with the same structures. However, an enormous number of time steps are required for these converted SNNs, thus losing the energy-efficient benefit. Utilizing both the accuracy advantages of ANNs and the computing efficiency of SNNs, a novel SNN training framework is proposed, namely layer-wise ANN-to-SNN knowledge distillation (LaSNN). In order to achieve competitive accuracy and reduced inference latency, LaSNN transfers the learning from a well-trained ANN to a small SNN by distilling the knowledge other than converting the parameters of ANN. The information gap between heterogeneous ANN and SNN is bridged by introducing the attention scheme, the knowledge in an ANN is effectively compressed and then efficiently transferred by utilizing our layer-wise distillation paradigm. We conduct detailed experiments to demonstrate the effectiveness, efficacy, and scalability of LaSNN on three benchmark data sets (CIFAR-10, CIFAR-100, and Tiny ImageNet). We achieve competitive top-1 accuracy compared to ANNs and 20x faster inference than converted SNNs with similar performance. More importantly, LaSNN is dexterous and extensible that can be effortlessly developed for SNNs with different architectures/depths and input encoding methods, contributing to their potential development.

Dynamic Ensemble Bayesian Filter for Robust Control of a Human Brain-machine Interface

Objective: Brain-machine interfaces (BMIs) aim to provide direct brain control of devices such as prostheses and computer cursors, which have demonstrated great potential for mobility restoration. One major limitation of current BMIs lies in the unstable performance in online control due to the variability of neural signals, which seriously hinders the clinical availability of BMIs. Method: To deal with the neural variability in online BMI control, we propose a dynamic ensemble Bayesian filter (DyEnsemble). DyEnsemble extends Bayesian filters with a dynamic measurement model, which adjusts its parameters in time adaptively with neural changes. This is achieved by learning a pool of candidate functions and dynamically weighting and assembling them according to neural signals. In this way, DyEnsemble copes with variability in signals and improves the robustness of online control. Results: Online BMI experiments with a human participant demonstrate that, compared with the velocity Kalman filter, DyEnsemble significantly improves the control accuracy (increases the success rate by 13.9% and reduces the reach time by 13.5% in the random target pursuit task) and robustness (performs more stably over different experiment days). Conclusion: Our results demonstrate the superiority of DyEnsemble in online BMI control. Significance: DyEnsemble frames a novel and flexible framework for robust neural decoding, which is beneficial to different neural decoding applications.

Dynamic Ensemble Modeling Approach to Nonstationary Neural Decoding in Brain-Computer Interfaces

Brain-computer interfaces (BCIs) have enabled prosthetic device control by decoding motor movements from neural activities. Neural signals recorded from cortex exhibit nonstationary property due to abrupt noises and neuroplastic changes in brain activities during motor control. Current state-of-the-art neural signal decoders such as Kalman filter assume fixed relationship between neural activities and motor movements, thus will fail if this assumption is not satisfied. We propose a dynamic ensemble modeling (DyEnsemble) approach that is capable of adapting to changes in neural signals by employing a proper combination of decoding functions. The DyEnsemble method firstly learns a set of diverse candidate models. Then, it dynamically selects and combines these models online according to Bayesian updating mechanism. Our method can mitigate the effect of noises and cope with different task behaviors by automatic model switching, thus gives more accurate predictions. Experiments with neural data demonstrate that the DyEnsemble method outperforms Kalman filters remarkably, and its advantage is more obvious with noisy signals.

Spiking Deep Residual Network

Recently, spiking neural network (SNN) has received significant attentions for its biological plausibility. SNN theoretically has at least the same computational power as traditional artificial neural networks (ANNs), and it has the potential to achieve revolutionary energy-efficiency. However, at current stage, it is still a big challenge to train a very deep SNN. In this paper, we propose an efficient approach to build a spiking version of deep residual network (ResNet), which represents the state-of-the-art convolutional neural networks (CNNs). We employ the idea of converting a trained ResNet to a network of spiking neurons named Spiking ResNet. To address the conversion problem, we propose a shortcut normalisation mechanism to appropriately scale continuous-valued activations to match firing rates in SNN, and a layer-wise error compensation approach to reduce the error caused by discretisation. Experimental results on MNIST, CIFAR-10, and CIFAR-100 demonstrate that the proposed Spiking ResNet yields the state-of-the-art performance of SNNs.

Spectral Sparse Representation for Clustering: Evolved from PCA, K-means, Laplacian Eigenmap, and Ratio Cut

Dimensionality reduction, cluster analysis, and sparse representation are basic components in machine learning. However, their relationships have not yet been fully investigated. In this paper, we find that the spectral graph theory underlies a series of these elementary methods and can unify them into a complete framework. The methods include PCA, K-means, Laplacian eigenmap (LE), ratio cut (Rcut), and a new sparse representation method developed by us, called spectral sparse representation (SSR). Further, extended relations to conventional over-complete sparse representations (e.g., method of optimal directions, KSVD), manifold learning (e.g., kernel PCA, multidimensional scaling, Isomap, locally linear embedding), and subspace clustering (e.g., sparse subspace clustering, low-rank representation) are incorporated. We show that, under an ideal condition from the spectral graph theory, PCA, K-means, LE, and Rcut are unified together. And when the condition is relaxed, the unification evolves to SSR, which lies in the intermediate between PCA/LE and K-mean/Rcut. An efficient algorithm, NSCrt, is developed to solve the sparse codes of SSR. SSR combines merits of both sides: its sparse codes reduce dimensionality of data meanwhile revealing cluster structure. For its inherent relation to cluster analysis, the codes of SSR can be directly used for clustering. Scut, a clustering approach derived from SSR reaches the state-of-the-art performance in the spectral clustering family. The one-shot solution obtained by Scut is comparable to the optimal result of K-means that are run many times. Experiments on various data sets demonstrate the properties and strengths of SSR, NSCrt, and Scut.