Detecting when a neural sequence model does "interesting" computation is an open problem. The next token prediction loss is a poor indicator: Low loss can stem from trivially predictable sequences that are uninteresting, while high loss may reflect unpredictable but also irrelevant information that can be ignored by the model. We propose a better metric: measuring the model's ability to predict its own future hidden states. We show empirically that this metric -- in contrast to the next token prediction loss -- correlates with the intuitive interestingness of the task. To measure predictability, we introduce the architecture-agnostic "prediction of hidden states" (PHi) layer that serves as an information bottleneck on the main pathway of the network (e.g., the residual stream in Transformers). We propose a novel learned predictive prior that enables us to measure the novel information gained in each computation step, which serves as our metric. We show empirically that our metric predicts the description length of formal languages learned in-context, the complexity of mathematical reasoning problems, and the correctness of self-generated reasoning chains.

Models that rely on subword tokenization have significant drawbacks, such as sensitivity to character-level noise like spelling errors and inconsistent compression rates across different languages and scripts. While character-or byte-level models like ByT5 attempt to address these concerns, they have not gained widespread adoption--processing raw byte streams without tokenization results in significantly longer sequence lengths, making training and inference inefficient. This work introduces MrT5 (MergeT5), a more efficient variant of ByT5 that integrates a token deletion mechanism in its encoder to dynamically shorten the input sequence length. After processing through a fixed number of encoder layers, a learnt delete gate determines which tokens are to be removed and which are to be retained for subsequent layers. MrT5 effectively "merges" critical information from deleted tokens into a more compact sequence, leveraging contextual information from the remaining tokens. In continued pre-training experiments, we find that MrT5 can achieve significant gains in inference runtime with minimal effect on performance. When trained on English text, MrT5 demonstrates the capability to transfer its deletion feature zero-shot across several languages, with significant additional improvements following multilingual training. Furthermore, MrT5 shows comparable accuracy to ByT5 on downstream evaluations such as XNLI and character-level tasks while reducing sequence lengths by up to 80%. Our approach presents a solution to the practical limitations of existing byte-level models. Subword tokenization, typically via algorithms such as byte-pair encoding (Sennrich et al., 2016) or SentencePiece (Kudo & Richardson, 2018), is a fundamental text preprocessing step that has become ubiquitous in modern language models. Subword tokenizers divide text into meaningful units known as tokens, which closely resemble words or parts of words. Tokenization can be seen as a form of compression, since it reduces the sequence length of the input passed to the computeintensive Transformer (Vaswani et al., 2017). However, subword tokenizers have several drawbacks.

Previous work on Universal Transformers (UTs) has demonstrated the importance of parameter sharing across layers. By allowing recurrence in depth, UTs have advantages over standard Transformers in learning compositional generalizations, but layer-sharing comes with a practical limitation of parameter-compute ratio: it drastically reduces the parameter count compared to the non-shared model with the same dimensionality. Naively scaling up the layer size to compensate for the loss of parameters makes its computational resource requirements prohibitive. In practice, no previous work has succeeded in proposing a shared-layer Transformer design that is competitive in parameter count-dominated tasks such as language modeling. Here we propose MoEUT (pronounced "moot"), an effective mixture-of-experts (MoE)-based shared-layer Transformer architecture, which combines several recent advances in MoEs for both feedforward and attention layers of standard Transformers together with novel layer-normalization and grouping schemes that are specific and crucial to UTs. The resulting UT model, for the first time, slightly outperforms standard Transformers on language modeling tasks such as BLiMP and PIQA, while using significantly less compute and memory.

The costly self-attention layers in modern Transformers require memory and compute quadratic in sequence length. Existing approximation methods usually underperform and fail to obtain significant speedups in practice. Here we present SwitchHead--a novel method that reduces both compute and memory requirements and achieves wall-clock speedup, while matching the language modeling performance of baseline Transformers with the same parameter budget. Switch-Head uses Mixture-of-Experts (MoE) layers for the value and output projections and requires 4 to 8 times fewer attention matrices than standard Transformers. Our novel attention can also be combined with MoE MLP layers, resulting in an efficient fully-MoE "SwitchAll" Transformer model. Large language models (LLMs) have shown remarkable capabilities (Radford et al., 2019; Brown et al., 2020; OpenAI, 2022; 2023) and great versatility (Bubeck et al., 2023). However, training enormous Transformers (Vaswani et al., 2017; Schmidhuber, 1992) requires a considerable amount of computing power and memory, which is not accessible to most researchers, academic institutions, and even companies. Even running them in inference mode, which is much less resource-intensive, requires significant engineering effort (Gerganov, 2023). Accelerating big Transformers remains an important open research question. However, in these works, the parameter efficiency of MoEs has not been studied; MoE models have been typically compared to dense baselines with the same number of FLOPs but with much less parameters.

General-purpose learning systems should improve themselves in open-ended fashion in ever-changing environments. Conventional learning algorithms for neural networks, however, suffer from catastrophic forgetting (CF)--previously acquired skills are forgotten when a new task is learned. Instead of hand-crafting new algorithms for avoiding CF, we propose Automated Continual Learning (ACL) to train self-referential neural networks to meta-learn their own in-context continual (meta-)learning algorithms. Our experiments demonstrate that ACL effectively solves "in-context catastrophic forgetting"; our ACL-learned algorithms outperform hand-crafted ones, e.g., on the Split-MNIST benchmark in the replay-free setting, and enables continual learning of diverse tasks consisting of multiple few-shot and standard image classification datasets. Enemies of memories are other memories (Eagleman, 2020). Continually-learning artificial neural networks (NNs) are memory systems in which their weights store memories of task-solving skills or programs, and their learning algorithm is responsible for memory read/write operations. Conventional learning algorithms--used to train NNs in the standard scenarios where all training data is available at once--are known to be inadequate for continual learning (CL) of multiple tasks where data for each task is available sequentially and exclusively, one at a time. They suffer from "catastrophic forgetting" (CF; McCloskey & Cohen (1989); Ratcliff (1990); French (1999); McClelland et al. (1995)); the NNs forget, or rather, the learning algorithm erases, previously acquired skills, in exchange of learning to solve a new task. Naturally, a certain degree of forgetting is unavoidable when the memory capacity is limited, and the amount of things to remember exceeds such an upper bound. In general, however, capacity is not the fundamental cause of CF; typically, the same NNs, suffering from CF when trained on two tasks sequentially, can perform well on both tasks when they are jointly trained on the two tasks at once instead (see, e.g., Irie et al. (2022a)). The real root of CF lies in the learning algorithm as a memory mechanism. A "good" CL algorithm should preserve previously acquired knowledge while also leveraging previous learning experiences to improve future learning, by maximally exploiting the limited memory space of model parameters. All of this is the decision-making problem of learning algorithms.

How to reduce compute and memory requirements of neural networks (NNs) without sacrificing performance? Many recent works use sparse Mixtures of Experts (MoEs) to build resource-efficient large language models (LMs). Here we introduce several novel perspectives on MoEs, presenting a general framework that unifies various methods to approximate two-layer NNs (e.g., feedforward blocks of Transformers), including product-key memories (PKMs). Leveraging insights from this framework, we propose methods to improve both MoEs and PKMs. Unlike prior work that compares MoEs with dense baselines under the compute-equal condition, our evaluation condition is parameter-equal, which is crucial to properly evaluate LMs. We show that our MoEs are competitive with the dense Transformer-XL on both the WikiText-103 and enwiki8 datasets at two different scales, while being much more resource efficient. This demonstrates that MoEs are relevant not only to extremely large LMs but also to any-scale resource-efficient LMs. Our code is public.

Recent studies of the computational power of recurrent neural networks (RNNs) reveal a hierarchy of RNN architectures, given real-time and finite-precision assumptions. Here we study auto-regressive Transformers with linearised attention, a.k.a. linear Transformers (LTs) or Fast Weight Programmers (FWPs). LTs are special in the sense that they are equivalent to RNN-like sequence processors with a fixed-size state, while they can also be expressed as the now-popular self-attention networks. We show that many well-known results for the standard Transformer directly transfer to LTs/FWPs. Our formal language recognition experiments demonstrate how recently proposed FWP extensions such as recurrent FWPs and self-referential weight matrices successfully overcome certain limitations of the LT, e.g., allowing for generalisation on the parity problem. Our code is public.

Both Minsky's "society of mind" and Schmidhuber's "learning to think" inspire diverse societies of large multimodal neural networks (NNs) that solve problems by interviewing each other in a "mindstorm." Recent implementations of NN-based societies of minds consist of large language models (LLMs) and other NN-based experts communicating through a natural language interface. In doing so, they overcome the limitations of single LLMs, improving multimodal zero-shot reasoning. In these natural language-based societies of mind (NLSOMs), new agents -- all communicating through the same universal symbolic language -- are easily added in a modular fashion. To demonstrate the power of NLSOMs, we assemble and experiment with several of them (having up to 129 members), leveraging mindstorms in them to solve some practical AI tasks: visual question answering, image captioning, text-to-image synthesis, 3D generation, egocentric retrieval, embodied AI, and general language-based task solving. We view this as a starting point towards much larger NLSOMs with billions of agents-some of which may be humans. And with this emergence of great societies of heterogeneous minds, many new research questions have suddenly become paramount to the future of artificial intelligence. What should be the social structure of an NLSOM? What would be the (dis)advantages of having a monarchical rather than a democratic structure? How can principles of NN economies be used to maximize the total reward of a reinforcement learning NLSOM? In this work, we identify, discuss, and try to answer some of these questions.

Transformers have impressive generalization capabilities on tasks with a fixed context length. However, they fail to generalize to sequences of arbitrary length, even for seemingly simple tasks such as duplicating a string. Moreover, simply training on longer sequences is inefficient due to the quadratic computation complexity of the global attention mechanism. In this work, we demonstrate that this failure mode is linked to positional encodings being out-of-distribution for longer sequences (even for relative encodings) and introduce a novel family of positional encodings that can overcome this problem. Concretely, our randomized positional encoding scheme simulates the positions of longer sequences and randomly selects an ordered subset to fit the sequence's length. Our large-scale empirical evaluation of 6000 models across 15 algorithmic reasoning tasks shows that our method allows Transformers to generalize to sequences of unseen length (increasing test accuracy by 12.0% on average).

Unsupervised learning of discrete representations from continuous ones in neural networks (NNs) is the cornerstone of several applications today. Vector Quantisation (VQ) has become a popular method to achieve such representations, in particular in the context of generative models such as Variational Auto-Encoders (VAEs). For example, the exponential moving average-based VQ (EMA-VQ) algorithm is often used. Here we study an alternative VQ algorithm based on the learning rule of Kohonen Self-Organising Maps (KSOMs; 1982) of which EMA-VQ is a special case. In fact, KSOM is a classic VQ algorithm which is known to offer two potential benefits over the latter: empirically, KSOM is known to perform faster VQ, and discrete representations learned by KSOM form a topological structure on the grid whose nodes are the discrete symbols, resulting in an artificial version of the topographic map in the brain. We revisit these properties by using KSOM in VQ-VAEs for image processing. In particular, our experiments show that, while the speed-up compared to well-configured EMA-VQ is only observable at the beginning of training, KSOM is generally much more robust than EMA-VQ, e.g., w.r.t. the choice of initialisation schemes. Our code is public.