Transformers with multi-head self-attention have achieved remarkable success in sequence modeling and beyond. However, they suffer from high computational and memory complexities for computing the attention matrix at each head. Recently, it has been shown that those attention matrices lie on a low-dimensional manifold and, thus, are redundant. We propose the Transformer with a Finite Admixture of Shared Heads (FiSHformers), a novel class of efficient and flexible transformers that allow the sharing of attention matrices between attention heads. At the core of FiSHformer is a novel finite admixture model of shared heads (FiSH) that samples attention matrices from a set of global attention matrices. The number of global attention matrices is much smaller than the number of local attention matrices generated. FiSHformers directly learn these global attention matrices rather than the local ones as in other transformers, thus significantly improving the computational and memory efficiency of the model. We empirically verify the advantages of the FiSHformer over the baseline transformers in a wide range of practical applications including language modeling, machine translation, and image classification. On the WikiText-103, IWSLT'14 De-En and WMT'14 En-De, FiSHformers use much fewer floating-point operations per second (FLOPs), memory, and parameters compared to the baseline transformers.

Equivariance is a powerful prior for learning physical dynamics, yet exact group equivariance can degrade performance if the symmetries are broken. We propose object-centric world models built with geometric algebra neural networks, providing a soft geometric inductive bias. Our models are evaluated using simulated environments of 2d rigid body dynamics with static obstacles, where we train for next-step predictions autoregressively. For long-horizon rollouts we show that the soft inductive bias of our models results in better performance in terms of physical fidelity compared to non-equivariant baseline models. The approach complements recent soft-equivariance ideas and aligns with the view that simple, well-chosen priors can yield robust generalization. These results suggest that geometric algebra offers an effective middle ground between hand-crafted physics and unstructured deep nets, delivering sample-efficient dynamics models for multi-object scenes.

Multi-head attention empowers the recent success of transformers, the state-of-the-art models that have achieved remarkable success in sequence modeling and beyond.

Transformers with multi-head self-attention have achieved remarkable success in sequence modeling and beyond. However, they suffer from high computational and memory complexities for computing the attention matrix at each head. Recently, it has been shown that those attention matrices lie on a low-dimensional manifold and, thus, are redundant. We propose the Transformer with a Finite Admixture of Shared Heads (FiSHformers), a novel class of efficient and flexible transformers that allow the sharing of attention matrices between attention heads. At the core of FiSHformer is a novel finite admixture model of shared heads (FiSH) that samples attention matrices from a set of global attention matrices.

Given large data sets and sufficient compute, is it beneficial to design neural architectures for the structure and symmetries of each problem? Or is it more efficient to learn them from data? We study empirically how equivariant and non-equivariant networks scale with compute and training samples. Focusing on a benchmark problem of rigid-body interactions and on general-purpose transformer architectures, we perform a series of experiments, varying the model size, training steps, and dataset size. We find evidence for three conclusions. First, equivariance improves data efficiency, but training non-equivariant models with data augmentation can close this gap given sufficient epochs. Second, scaling with compute follows a power law, with equivariant models outperforming non-equivariant ones at each tested compute budget. Finally, the optimal allocation of a compute budget onto model size and training duration differs between equivariant and non-equivariant models.

Can we preserve the accuracy of neural models while also providing faithful explanations? We present wrapper boxes, a general approach to generate faithful, example-based explanations for model predictions while maintaining predictive performance. After training a neural model as usual, its learned feature representation is input to a classic, interpretable model to perform the actual prediction. This simple strategy is surprisingly effective, with results largely comparable to those of the original neural model, as shown across three large pre-trained language models, two datasets of varying scale, four classic models, and four evaluation metrics. Moreover, because these classic models are interpretable by design, the subset of training examples that determine classic model predictions can be shown directly to users.

To predict the next token, autoregressive models However, those NP-hard distributions are artificial. For naturally ordinarily examine the past. Could they also benefit occurring sequences, why might one expect lookahead from also examining hypothetical futures? We to help autoregressive modeling? We argue that when the consider a novel Transformer-based autoregressive sequences represent an agent's behavior, an autoregressive architecture that estimates the next-token distribution parameterization is not always the simplest description. If by extrapolating multiple continuations the behavior is goal-directed--for example, an agent trying of the past, according to some proposal distribution, to achieve high reward in a Markov Decision Process--then and attending to these extended strings. This the simplest description may include a characterization of architecture draws insights from classical AI systems the agent's environment and goals. Even if the agent explicitly such as board game players: when making consults an autoregressive policy p(action | state) a local decision, a policy may benefit from exploring at each step, that policy is not arbitrary: while it may appear possible future trajectories and analyzing complex, it was shaped by reinforcement learning or them. On multiple tasks including morphological by natural selection so as to achieve high-reward trajectories.

In recent years, a great deal of attention has been paid to the Transformer network for speech recognition tasks due to its excellent model performance. However, the Transformer network always involves heavy computation and large number of parameters, causing serious deployment problems in devices with limited computation sources or storage memory. In this paper, a new lightweight model called Sim-T has been proposed to expand the generality of the Transformer model. Under the help of the newly developed multiplexing technique, the Sim-T can efficiently compress the model with negligible sacrifice on its performance. To be more precise, the proposed technique includes two parts, that are, module weight multiplexing and attention score multiplexing. Moreover, a novel decoder structure has been proposed to facilitate the attention score multiplexing. Extensive experiments have been conducted to validate the effectiveness of Sim-T. In Aishell-1 dataset, when the proposed Sim-T is 48% parameter less than the baseline Transformer, 0.4% CER improvement can be obtained. Alternatively, 69% parameter reduction can be achieved if the Sim-T gives the same performance as the baseline Transformer. With regard to the HKUST and WSJ eval92 datasets, CER and WER will be improved by 0.3% and 0.2%, respectively, when parameters in Sim-T are 40% less than the baseline Transformer.