Despite their nearly universal adoption for large language models, the internal workings of transformers are not well understood. We aim to better understand the impact of removing or reorganizing information throughout the layers of a pretrained transformer. Such an understanding could both yield better usage of existing models as well as to make architectural improvements to produce new variants. We present a series of empirical studies on frozen models that show that the lower and final layers of pretrained transformers differ from middle layers, but that middle layers have a surprising amount of uniformity. We further show that some classes of problems have robustness to skipping layers, running the layers in an order different from how they were trained, or running the layers in parallel. Our observations suggest that even frozen pretrained models may gracefully trade accuracy for latency by skipping layers or running layers in parallel.

Biological brains demonstrate complex neural activity, where neural dynamics are critical to how brains process information. Most artificial neural networks ignore the complexity of individual neurons. We challenge that paradigm. By incorporating neuron-level processing and synchronization, we reintroduce neural timing as a foundational element. We present the Continuous Thought Machine (CTM), a model designed to leverage neural dynamics as its core representation. The CTM has two innovations: (1) neuron-level temporal processing}, where each neuron uses unique weight parameters to process incoming histories; and (2) neural synchronization as a latent representation. The CTM aims to strike a balance between neuron abstractions and biological realism. It operates at a level of abstraction that effectively captures essential temporal dynamics while remaining computationally tractable. We demonstrate the CTM's performance and versatility across a range of tasks, including solving 2D mazes, ImageNet-1K classification, parity computation, and more. Beyond displaying rich internal representations and offering a natural avenue for interpretation owing to its internal process, the CTM is able to perform tasks that require complex sequential reasoning. The CTM can also leverage adaptive compute, where it can stop earlier for simpler tasks, or keep computing when faced with more challenging instances. The goal of this work is to share the CTM and its associated innovations, rather than pushing for new state-of-the-art results. To that end, we believe the CTM represents a significant step toward developing more biologically plausible and powerful artificial intelligence systems. We provide an accompanying [interactive online demonstration](https://pub.sakana.ai/ctm/) and an [extended technical report](https://pub.sakana.ai/ctm/paper).

Abstract If one sees the place name Houston Mercer Dog Run in New York, how does one know how to pronounce it? Assuming one knows that Houston in New York is pronounced /ˈhaʊstən/ and not like the Texas city (/ˈhjuːstən/), then one can probably guess that /ˈhaʊstən/ is also used in the name of the dog park. We present a novel architecture that learns to use the pronunciations of neighboring names in order to guess the pronunciation of a given target feature. Applied to Japanese place names, we demonstrate the utility of the model to finding and proposing corrections for errors in Google Maps. To demonstrate the utility of this approach to structurally similar problems, we also report on an application to a totally different task: Cognate reflex prediction in comparative historical linguistics. A version of the code has been open-sourced.1

Computational biology and bioinformatics provide vast data gold-mines from protein sequences, ideal for Language Models (LMs) taken from Natural Language Processing (NLP). These LMs reach for new prediction frontiers at low inference costs. Here, we trained two auto-regressive models (Transformer-XL, XLNet) and four auto-encoder models (BERT, Albert, Electra, T5) on data from UniRef and BFD containing up to 393 billion amino acids. The protein LMs (pLMs) were trained on the Summit supercomputer using 5616 GPUs and TPU Pod up-to 1024 cores. Dimensionality reduction revealed that the raw pLM-embeddings from unlabeled data captured some biophysical features of protein sequences. We validated the advantage of using the embeddings as exclusive input for several subsequent tasks: (1) a per-residue (per-token) prediction of protein secondary structure (3-state accuracy Q3=81%-87%); (2) per-protein (pooling) predictions of protein sub-cellular location (ten-state accuracy: Q10=81%) and membrane versus water-soluble (2-state accuracy Q2=91%). For secondary structure, the most informative embeddings (ProtT5) for the first time outperformed the state-of-the-art without multiple sequence alignments (MSAs) or evolutionary information thereby bypassing expensive database searches. Taken together, the results implied that pLMs learned some of the grammar of the language of life. All our models are available through https://github.com/agemagician/ProtTrans.

End-to-end (E2E) modeling is advantageous for automatic speech recognition (ASR) especially for Japanese since word-based tokenization of Japanese is not trivial, and E2E modeling is able to model character sequences directly. This paper focuses on the latest E2E modeling techniques, and investigates their performances on character-based Japanese ASR by conducting comparative experiments. The results are analyzed and discussed in order to understand the relative advantages of long short-term memory (LSTM), and Conformer models in combination with connectionist temporal classification, transducer, and attention-based loss functions. Furthermore, the paper investigates on effectivity of the recent training techniques such as data augmentation (SpecAugment), variational noise injection, and exponential moving average. The best configuration found in the paper achieved the state-of-the-art character error rates of 4.1%, 3.2%, and 3.5% for Corpus of Spontaneous Japanese (CSJ) eval1, eval2, and eval3 tasks, respectively. The system is also shown to be computationally efficient thanks to the efficiency of Conformer transducers.

LSTMs and other RNN variants have shown strong performance on character-level language modeling. These models are typically trained using truncated backpropagation through time, and it is common to assume that their success stems from their ability to remember long-term contexts. In this paper, we show that a deep (64-layer) transformer model (Vaswani et al. 2017) with fixed context outperforms RNN variants by a large margin, achieving state of the art on two popular benchmarks: 1.13 bits per character on text8 and 1.06 on enwik8. To get good results at this depth, we show that it is important to add auxiliary losses, both at intermediate network layers and intermediate sequence positions.

We present the Natural Questions corpus, a question answering data set. Questions consist of real anonymized, aggregated queries issued to the Google search engine. An annotator is presented with a question along with a Wikipedia page from the top 5 search results, and annotates a long answer (typically a paragraph) and a short answer (one or more entities) if present on the page, or marks null if no long/short answer is present. The public release consists of 307,373 training examples with single annotations; 7,830 examples with 5-way annotations for development data; and a further 7,842 examples with 5-way annotated sequestered as test data. We present experiments validating quality of the data. We also describe analysis of 25-way annotations on 302 examples, giving insights into human variability on the annotation task. We introduce robust metrics for the purposes of evaluating question answering systems; demonstrate high human upper bounds on these metrics; and establish baseline results using competitive methods drawn from related literature.

The machine reading task, where a computer reads a document and answers questions about it, is important in artificial intelligence research. Recently, many models have been proposed to address it. Word-level models, which have words as units of input and output, have proven to yield state-of-the-art results when evaluated on English datasets. However, in morphologically richer languages, many more unique words exist than in English due to highly productive prefix and suffix mechanisms. This may set back word-level models, since vocabulary sizes too big to allow for efficient computing may have to be employed. Multiple alternative input granularities have been proposed to avoid large input vocabularies, such as morphemes, character n-grams, and bytes. Bytes are advantageous as they provide a universal encoding format across languages, and allow for a small vocabulary size, which, moreover, is identical for every input language. In this work, we investigate whether bytes are suitable as input units across morphologically varied languages. To test this, we introduce two large-scale machine reading datasets in morphologically rich languages, Turkish and Russian. We implement 4 byte-level models, representing the major types of machine reading models and introduce a new seq2seq variant, called encoder-transformer-decoder. We show that, for all languages considered, there are models reading bytes outperforming the current state-of-the-art word-level baseline. Moreover, the newly introduced encoder-transformer-decoder performs best on the morphologically most involved dataset, Turkish. The large-scale Turkish and Russian machine reading datasets are released to public.

Mia Xu Chen, Orhan Firat, Ankur Bapna, Melvin Johnson, Wolfgang Macherey, George Foster, Llion Jones, Mike Schuster, Noam Shazeer, Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Zhifeng Chen, Yonghui Wu, Macduff Hughes. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2018.

We introduce a hierarchical architecture for machine reading capable of extracting precise information from long documents.The model divides the document into small, overlapping windows and encodes all windows in parallel with an RNN.It then attends over these window encodings, reducing them to a single encoding, which is decoded into an answer using a sequence decoder.This hierarchical approach allows the model to scale to longer documents without increasing the number of sequential steps.In a supervised setting, our model achieves state of the art accuracy of 76.8 on the WikiReading dataset.We also evaluate the model in a semi-supervised setting by downsampling the WikiReading training set to create increasingly smaller amounts of supervision, while leaving the full unlabeled document corpus to train a sequence autoencoder on document windows.We evaluate models that can reuse autoencoder states and outputs without finetuning their weights, allowing for more efficient training and inference.

The dominant sequence transduction models are based on complex recurrent orconvolutional neural networks in an encoder and decoder configuration. The best performing such models also connect the encoder and decoder through an attentionm echanisms. We propose a novel, simple network architecture based solely onan attention mechanism, dispensing with recurrence and convolutions entirely.Experiments on two machine translation tasks show these models to be superiorin quality while being more parallelizable and requiring significantly less timeto train. Our single model with 165 million parameters, achieves 27.5 BLEU onEnglish-to-German translation, improving over the existing best ensemble result by over 1 BLEU. On English-to-French translation, we outperform the previoussingle state-of-the-art with model by 0.7 BLEU, achieving a BLEU score of 41.1.

Daniel Hewlett, Alexandre Lacoste, Llion Jones, Illia Polosukhin, Andrew Fandrianto, Jay Han, Matthew Kelcey, David Berthelot. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2016.