We develop hybrid memory architectures for general-purpose sequence processing neural networks, that combine key-value memory using softmax attention (KV-memory) with fast weight memory through dynamic synaptic modulation (FW-memory)---the core principles of quadratic and linear transformers, respectively. These two memory systems have complementary but individually limited properties: KV-memory offers precise retrieval but is constrained by quadratic complexity in sequence length, while FW-memory supports arbitrarily long sequences and enables more expressive computation but sacrifices precise recall. We propose and compare three methods to blend these two systems into a single memory system, differing in how and when input information is delivered to each system, to leverage the strengths of both. We conduct experiments on general language modeling and retrieval tasks by training 340M- and 1.3B-parameter models from scratch, as well as on synthetic algorithmic tasks designed to precisely illustrate the benefits of certain hybrid methods over others. We also evaluate our hybrid memory systems on reinforcement learning in partially observable environments. Overall, we demonstrate how a well-designed hybrid can overcome the limitations of its individual components, offering new insights into the design principle of neural memory systems.

Inspired by Minsky's Society of Mind, Schmidhuber's Learning to Think, and other more recent works, this paper proposes and advocates for the concept of natural language-based societies of mind (NLSOMs). We imagine these societies as consisting of a collection of multimodal neural networks, including large language models, which engage in a “mindstorm” to solve problems using a shared natural language interface. Here, we work to identify and discuss key questions about the social structure, governance, and economic principles for NLSOMs, emphasizing their impact on the future of AI. Our demonstrations with NLSOMs-which feature up to 129 agents-show their effectiveness in various tasks, including visual question answering, image captioning, and prompt generation for text-to-image synthesis.

Real-time recurrent learning (RTRL) for sequence-processing recurrent neural networks (RNNs) offers certain conceptual advantages over backpropagation through time (BPTT). RTRL requires neither caching past activations nor truncating context, and enables online learning. However, RTRL's time and space complexity make it impractical. To overcome this problem, most recent work on RTRL focuses on approximation theories, while experiments are often limited to diagnostic settings. Here we explore the practical promise of RTRL in more realistic settings. We study actor-critic methods that combine RTRL and policy gradients, and test them in several subsets of DMLab-30, ProcGen, and Atari-2600 environments. On DMLab memory tasks, our system trained on fewer than 1.2B environmental frames is competitive with or outperforms well-known IMPALA and R2D2 baselines trained on 10B frames. To scale to such challenging tasks, we focus on certain well-known neural architectures with element-wise recurrence, allowing for tractable RTRL without approximation. Importantly, we also discuss rarely addressed limitations of RTRL in real-world applications, such as its complexity in the multi-layer case.

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.

Despite many recent works on Mixture of Experts (MoEs) for resource-efficient Transformer language models, existing methods mostly focus on MoEs for feedforward layers. Previous attempts at extending MoE to the self-attention layer fail to match the performance of the parameter-matched baseline. Our novel SwitchHead is an effective MoE method for the attention layer that successfully reduces both the compute and memory requirements, achieving wall-clock speedup, while matching the language modeling performance of the baseline Transformer. Our novel MoE mechanism allows SwitchHead to compute up to 8 times fewer attention matrices than the standard Transformer. SwitchHead can also be combined with MoE feedforward layers, resulting in fully-MoE "SwitchAll" Transformers. For our 262M parameter model trained on C4, SwitchHead matches the perplexity of standard models with only 44% compute and 27% memory usage. Zero-shot experiments on downstream tasks confirm the performance of SwitchHead, e.g., achieving more than 3.5% absolute improvements on BliMP compared to the baseline with an equal compute resource.

Despite the remarkable empirical performance of Transformers, their theoretical understanding remains elusive. Here, we consider a deep multi-head self-attention network, that is closely related to Transformers yet analytically tractable. We develop a statistical mechanics theory of Bayesian learning in this model, deriving exact equations for the network's predictor statistics under the finite-width thermodynamic limit, i.e., $N,P\rightarrow\infty$, $P/N=\mathcal{O}(1)$, where $N$ is the network width and $P$ is the number of training examples. Our theory shows that the predictor statistics are expressed as a sum of independent kernels, each one pairing different "attention paths", defined as information pathways through different attention heads across layers. The kernels are weighted according to a "task-relevant kernel combination" mechanism that aligns the total kernel with the task labels. As a consequence, this interplay between attention paths enhances generalization performance. Experiments confirm our findings on both synthetic and real-world sequence classification tasks. Finally, our theory explicitly relates the kernel combination mechanism to properties of the learned weights, allowing for a qualitative transfer of its insights to models trained via gradient descent. As an illustration, we demonstrate an efficient size reduction of the network, by pruning those attention heads that are deemed less relevant by our theory.

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.1

Current state-of-the-art object-centric models use slots and attention-based routing for binding. However, this class of models has several conceptual limitations: the number of slots is hardwired; all slots have equal capacity; training has high computational cost; there are no object-level relational factors within slots. Synchrony-based models in principle can address these limitations by using complex-valued activations which store binding information in their phase components. However, working examples of such synchrony-based models have been developed only very recently, and are still limited to toy grayscale datasets and simultaneous storage of less than three objects in practice. Here we introduce architectural modifications and a novel contrastive learning method that greatly improve the state-of-the-art synchrony-based model. For the first time, we obtain a class of synchrony-based models capable of discovering objects in an unsupervised manner in multi-object color datasets and simultaneously representing more than three objects.

The discovery of reusable subroutines simplifies decision making and planning in complex reinforcement learning problems. Previous approaches propose to learn such temporal abstractions in an unsupervised fashion through observing state-action trajectories gathered from executing a policy. However, a current limitation is that they process each trajectory in an entirely sequential manner, which prevents them from revising earlier decisions about subroutine boundary points in light of new incoming information. In this work, we propose slot-based transformer for temporal abstraction (SloTTAr), a fully parallel approach that integrates sequence processing transformers with a slot attention module to discover subroutines in an unsupervised fashion while leveraging adaptive computation for learning about the number of such subroutines solely based on their empirical distribution. We demonstrate how SloTTAr is capable of outperforming strong baselines in terms of boundary point discovery, even for sequences containing variable amounts of subroutines, while being up to seven times faster to train on existing benchmarks.

Well-designed diagnostic tasks have played a key role in studying the failure of neural nets (NNs) to generalize systematically.Famous examples include SCAN and Compositional Table Lookup (CTL).Here we introduce CTL++, a new diagnostic dataset based on compositions of unary symbolic functions.While the original CTL is used to test length generalization or productivity, CTL++ is designed to test systematicity of NNs, that is, their capability to generalize to unseen compositions of known functions.CTL++ splits functions into groups and tests performance on group elements composed in a way not seen during training.We show that recent CTL-solving Transformer variants fail on CTL++.The simplicity of the task design allows for fine-grained control of task difficulty, as well as many insightful analyses.For example, we measure how much overlap between groups is needed by tested NNs for learning to compose.We also visualize how learned symbol representations in outputs of functions from different groups are compatible in case of success but not in case of failure.These results provide insights into failure cases reported on more complex compositions in the natural language domain.Our code is public.1

Linear layers in neural networks (NNs) trained by gradient descent can be expressed as a key-value memory system which stores all training datapoints and the initial weights, and produces outputs using unnormalised dot attention over the entire training experience. While this has been technically known since the 1960s, no prior work has effectively studied the operations of NNs in such a form, presumably due to prohibitive time and space complexities and impractical model sizes, all of them growing linearly with the number of training patterns which may get very large. However, this dual formulation offers a possibility of directly visualising how an NN makes use of training patterns at test time, by examining the corresponding attention weights. We conduct experiments on small scale supervised image classification tasks in single-task, multi-task, and continual learning settings, as well as language modelling, and discuss potentials and limits of this view for better understanding and interpreting how NNs exploit training patterns. Our code is public.

The weight matrix (WM) of a neural network (NN) is its program. The programs of many traditional NNs are learned through gradient descent in some error function, then remain fixed. The WM of a self-referential NN, however, can keep rapidly modifying all of itself during runtime. In principle, such NNs can meta-learn to learn, and meta-meta-learn to meta-learn to learn, and so on, in the sense of recursive self-improvement. While NN architectures potentially capable of implementing such behaviour have been proposed since the ’90s, there have been few if any practical studies. Here we revisit such NNs, building upon recent successes of fast weight programmers and closely related linear Transformers. We propose a scalable self-referential WM (SRWM) that learns to use outer products and the delta update rule to modify itself. We evaluate our SRWM in supervised few-shot learning and in multi-task reinforcement learning with procedurally generated game environments. Our experiments demonstrate both practical applicability and competitive performance of the proposed SRWM. Our code is public.

Neural ordinary differential equations (ODEs) have attracted much attention as continuous-time counterparts of deep residual neural networks (NNs), and numerous extensions for recurrent NNs have been proposed. Since the 1980s, ODEs have also been used to derive theoretical results for NN learning rules, e.g., the famous connection between Oja's rule and principal component analysis. Such rules are typically expressed as additive iterative update processes which have straightforward ODE counterparts. Here we introduce a novel combination of learning rules and Neural ODEs to build continuous-time sequence processing nets that learn to manipulate short-term memory in rapidly changing synaptic connections of other nets. This yields continuous-time counterparts of Fast Weight Programmers and linear Transformers. Our novel models outperform the best existing Neural Controlled Differential Equation based models on various time series classification tasks, while also addressing their fundamental scalability limitations. Our code is public.

Recently, many datasets have been proposed to test the systematic generalization ability of neural networks.The companion baseline Transformers, typically trained with default hyper-parameters from standard tasks, are shown to fail dramatically.Here we demonstrate that by revisiting model configurations as basic as scaling of embeddings, early stopping, relative positional embedding, and Universal Transformer variants, we can drastically improve the performance of Transformers on systematic generalization.We report improvements on five popular datasets: SCAN, CFQ, PCFG, COGS, and Mathematics dataset.Our models improve accuracy from 50% to 85% on the PCFG productivity split, and from 35% to 81% on COGS.On SCAN, relative positional embedding largely mitigates the EOS decision problem (Newman et al., 2020), yielding 100% accuracy on the length split with a cutoff at 26. Importantly, performance differences between these models are typically invisible on the IID data split.This calls for proper generalization validation sets for developing neural networks that generalize systematically.We publicly release the code to reproduce our results 1 .

We show the formal equivalence of linearised self-attention mechanisms and fast weight controllers from the early ’90s, where a slow neural net learns by gradient descent to program the fast weights of another net through sequences of elementary programming instructions which are additive outer products of self-invented activation patterns (today called keys and values). Such Fast Weight Programmers (FWPs) learn to manipulate the contents of a finite memory and dynamically interact with it. We infer a memory capacity limitation of recent linearised softmax attention variants, and replace the purely additive outer products by a delta rule-like programming instruction, such that the FWP can more easily learn to correct the current mapping from keys to values. The FWP also learns to compute dynamically changing learning rates. We also propose a new kernel function to linearise attention which balances simplicity and effectiveness. We conduct experiments on synthetic retrieval problems as well as standard machine translation and language modelling tasks which demonstrate the benefits of our methods.

Transformers with linearised attention (''linear Transformers'') have demonstrated the practical scalability and effectiveness of outer product-based Fast Weight Programmers (FWPs) from the '90s. However, the original FWP formulation is more general than the one of linear Transformers: a slow neural network (NN) continually reprograms the weights of a fast NN with arbitrary architecture. In existing linear Transformers, both NNs are feedforward and consist of a single layer. Here we explore new variations by adding recurrence to the slow and fast nets. We evaluate our novel recurrent FWPs (RFWPs) on two synthetic algorithmic tasks (code execution and sequential ListOps), Wikitext-103 language models, and on the Atari 2600 2D game environment. Our models exhibit properties of Transformers and RNNs. In the reinforcement learning setting, we report large improvements over LSTM in several Atari games. Our code is public.

Despite advances in neural language modeling, obtaining a good model on a large scale multi-domain dataset still remains a difficult task. We propose training methods for building neural language models for such a task, which are not only domain robust, but reasonable in model size and fast for evaluation. We combine knowledge distillation from pretrained domain expert language models with the noise contrastive estimation (NCE) loss. Knowledge distillation allows to train a single student model which is both compact and domain robust, while the use of NCE loss makes the model self-normalized, which enables fast evaluation. We conduct experiments on a large English multi-domain speech recognition dataset provided by AppTek. The resulting student model is of the size of one domain expert, while it gives similar perplexities as various teacher models on their expert domain; the model is self-normalized, allowing for 30% faster first pass decoding than the naive models which require the full soft- max computation, and finally it gives improvements of more than 8% relative in terms of word error rate over a large multidomain 4-gram count model trained on more than 10 B words.

We propose simple architectural modifications in the standard Transformer with the goal to reduce its total state size (defined as the number of self-attention layers times the sum of the key and value dimensions, times position) without loss of performance. Large scale Transformer language models have been empirically proved to give very good performance. However, scaling up results in a model that needs to store large states at evaluation time. This can increase the memory requirement dramatically for search e.g., in speech recognition (first pass decoding, lattice rescoring, or shallow fusion). In order to efficiently increase the model capacity without increasing the state size, we replace the single-layer feed-forward module in the Transformer layer by a deeper network, and decrease the total number of layers. In addition, we also evaluate the effect of key-value tying which directly divides the state size in half. On TED-LIUM 2, we obtain a model of state size 4 times smaller than the standard Transformer, with only 2% relative loss in terms of perplexity, which makes the deployment of Transformer language models more convenient.

We present a complete training pipeline to build a state-of-the-art hybrid HMM-based ASR system on the 2nd release of the TED-LIUM corpus. Data augmentation using SpecAugment is successfully applied to improve performance on top of our best SAT model using i-vectors. By investigating the effect of different maskings, we achieve improvements from SpecAugment on hybrid HMM models without increasing model size and training time. A subsequent sMBR training is applied to fine-tune the final acoustic model, and both LSTM and Transformer language models are trained and evaluated. Our best system achieves a 5.6% WER on the test set, which outperforms the previous state-of-the-art by 27% relative.

While recurrent neural networks can motivate cross-sentence language modeling and its application to automatic speech recognition (ASR), corresponding modifications of the training method for that end are rarely discussed. In fact, even more generally, the impact of training sequence construction strategy in language modeling for different evaluation conditions is typically ignored. In this work, we revisit this basic but fundamental question. We train language models based on long short-term memory recurrent neural networks and Transformers using various types of training sequences and study their robustness with respect to different evaluation modes. Our experiments on 300h Switchboard and Quaero English datasets show that models trained with back-propagation over sequences consisting of concatenation of multiple sentences with state carry-over across sequences effectively outperform those trained with the sentence-level training, both in terms of perplexity and word error rates for cross-utterance ASR.

We present competitive results using a Transformer encoder-decoder-attention model for end-to-end speech recognition needing less training time compared to a similarly performing LSTM model. We observe that the Transformer training is in general more stable compared to the LSTM, although it also seems to overfit more, and thus shows more problems with generalization. We also find that two initial LSTM layers in the Transformer encoder provide a much better positional encoding. Data-augmentation, a variant of SpecAugment, helps to improve both the Transformer by 33% and the LSTM by 15% relative. We analyze several pretraining and scheduling schemes, which is crucial for both the Transformer and the LSTM models. We improve our LSTM model by additional convolutional layers. We perform our experiments on Lib-riSpeech 1000h, Switchboard 300h and TED-LIUM-v2 200h, and we show state-of-the-art performance on TED-LIUM-v2 for attention based end-to-end models. We deliberately limit the training on LibriSpeech to 12.5 epochs of the training data for comparisons, to keep the results of practical interest, although we show that longer training time still improves more. We publish all the code and setups to run our experiments.

In conventional speech recognition, phoneme-based models outperform grapheme-based models for non-phonetic languages such as English. The performance gap between the two typically reduces as the amount of training data is increased. In this work, we examine the impact of the choice of modeling unit for attention-based encoder-decoder models. We conduct experiments on the LibriSpeech 100hr, 460hr, and 960hr tasks, using various target units (phoneme, grapheme, and word-piece); across all tasks, we find that grapheme or word-piece models consistently outperform phoneme-based models, even though they are evaluated without a lexicon or an external language model. We also investigate model complementarity: we find that we can improve WERs by up to 9% relative by rescoring N-best lists generated from a strong word-piece based baseline with either the phoneme or the grapheme model. Rescoring an N-best list generated by the phonemic system, however, provides limited improvements. Further analysis shows that the word-piece-based models produce more diverse N-best hypotheses, and thus lower oracle WERs, than phonemic models.

We explore deep autoregressive Transformer models in language modeling for speech recognition. We focus on two aspects. First, we revisit Transformer model configurations specifically for language modeling. We show that well configured Transformer models outperform our baseline models based on the shallow stack of LSTM recurrent neural network layers. We carry out experiments on the open-source LibriSpeech 960hr task, for both 200K vocabulary word-level and 10K byte-pair encoding subword-level language modeling. We apply our word-level models to conventional hybrid speech recognition by lattice rescoring, and the subword-level models to attention based encoder-decoder models by shallow fusion. Second, we show that deep Transformer language models do not require positional encoding. The positional encoding is an essential augmentation for the self-attention mechanism which is invariant to sequence ordering. However, in autoregressive setup, as is the case for language modeling, the amount of information increases along the position dimension, which is a positional signal by its own. The analysis of attention weights shows that deep autoregressive self-attention models can automatically make use of such positional information. We find that removing the positional encoding even slightly improves the performance of these models.

We present state-of-the-art automatic speech recognition (ASR) systems employing a standard hybrid DNN/HMM architecture compared to an attention-based encoder-decoder design for the LibriSpeech task. Detailed descriptions of the system development, including model design, pretraining schemes, training schedules, and optimization approaches are provided for both system architectures. Both hybrid DNN/HMM and attention-based systems employ bi-directional LSTMs for acoustic modeling/encoding. For language modeling, we employ both LSTM and Transformer based architectures. All our systems are built using RWTHs open-source toolkits RASR and RETURNN. To the best knowledge of the authors, the results obtained when training on the full LibriSpeech training set, are the best published currently, both for the hybrid DNN/HMM and the attention-based systems. Our single hybrid system even outperforms previous results obtained from combining eight single systems. Our comparison shows that on the LibriSpeech 960h task, the hybrid DNN/HMM system outperforms the attention-based system by 15% relative on the clean and 40% relative on the other test sets in terms of word error rate. Moreover, experiments on a reduced 100h-subset of the LibriSpeech training corpus even show a more pronounced margin between the hybrid DNN/HMM and attention-based architectures.

We present a new architecture and a training strategy for an adaptive mixture of experts with applications to domain robust language modeling. The proposed model is designed to benefit from the scenario where the training data are available in diverse domains as is the case for YouTube speech recognition. The two core components of our model are an ensemble of parallel long short-term memory (LSTM) expert layers for each domain and another LSTM based network which generates state dependent mixture weights for combining expert LSTM states by linear interpolation. The resulting model is a recurrent adaptive mixture model (RADMM) of domain experts. We train our model on 4.4B words from YouTube speech recognition data. We report results on the YouTube speech recognition test set. Compared with a background LSTM model, we obtain up to 12% relative improvement in perplexity and an improvement in word error rate from 12.3% to 12.1 % while using a lattice rescoring with strong pruning.

Long short-term memory (LSTM) recurrent neural network language models compress the full context of variable lengths into a fixed size vector. In this work, we investigate the task of predicting the LSTM hidden representation of the full context from a truncated n-gram context as a subtask for training an n-gram feedforward language model. Since this approach is a form of knowledge distillation, we compare two methods. First, we investigate the standard transfer based on the Kullback-Leibler divergence of the output distribution of the feedforward model from that of the LSTM. Second, we minimize the mean squared error between the hidden state of the LSTM and that of the n-gram feedforward model. We carry out experiments on different subsets of the Switchboard speech recognition dataset for feedforward models with a short (5-gram) and a medium (10-gram) context length. We show that we get improvements in perplexity and word error rate of up to 8% and 4% relative for the medium model, while the improvements are only marginal for the short model.

Sequence-to-sequence attention-based models on subword units allow simple open-vocabulary end-to-end speech recognition. In this work, we show that such models can achieve competitive results on the Switchboard 300h and LibriSpeech 1000h tasks. In particular, we report the state-of-the-art word error rates (WER) of 3.54% on the dev-clean and 3.82% on the test-clean evaluation subsets of LibriSpeech. We introduce a new pretraining scheme by starting with a high time reduction factor and lowering it during training, which is crucial both for convergence and final performance. In some experiments, we also use an auxiliary CTC loss function to help the convergence. In addition, we train long short-term memory (LSTM) language models on subword units. By shallow fusion, we report up to 27% relative improvements in WER over the attention baseline without a language model.

In this paper, we present an investigation on technical details of the byte-level convolutional layer which replaces the conventional linear word projection layer in the neural language model. In particular, we discuss and compare the effective filter configurations, pooling types and the use of bytes instead of characters. We carry out experiments on language packs released by the IARPA Babel project and measure the performance in terms of perplexity and word error rate. Introducing a convolutional layer consistently improves the results on all languages. Also, there is no degradation from using raw bytes instead of proper Unicode characters, even on syllabic alphabets like Amharic. In addition, we report improvements in word error rate from rescoring lattices and evaluate keyword search performance on several languages.

Inspired by the success of multi-task training in acoustic modeling, this paper investigates a new architecture for a multi-domain neural network based language model (NNLM). The proposed model has several shared hidden layers and domain-specific output layers. As will be shown, the log-linear interpolation of the multi-domain outputs and the optimization of interpolation weights fit naturally in the framework of NNLM. The resulting model can be expressed as a single NNLM. As an initial study of such an architecture, this paper focuses on deep feed-forward neural networks (DNNs). We also re-investigate the potential of long context up to 30-grams, and depth up to 5 hidden layers in DNN-LM. Our final feed-forward multidomain NNLM is trained on 3.1B running words across 11 domains for English broadcast news and conversations large vocabulary continuous speech recognition task. After log-linear interpolation and fine-tuning, we measured improvements in terms of perplexity and word error rate over the models trained on 50M running words of in-domain news resources. The final multi-domain feed-forward LM outperformed our previous best LSTM-RNN LM trained on the 50M in-domain corpus, even after linear interpolation with large count models.

Popularized by the long short-term memory (LSTM), multiplicative gates have become a standard means to design artificial neural networks with intentionally organized information flow.Notable examples of such architectures include gated recurrent units (GRU) and highway networks.In this work, we first focus on the evaluation of each of the classical gated architectures for language modeling for large vocabulary speech recognition.Namely, we evaluate the highway network, lateral network, LSTM and GRU.Furthermore, the motivation underlying the highway network also applies to LSTM and GRU.An extension specific to the LSTM has been recently proposed with an additional highway connection between the memory cells of adjacent LSTM layers.In contrast, we investigate an approach which can be used with both LSTM and GRU: a highway network in which the LSTM or GRU is used as the transformation function.We found that the highway connections enable both standalone feedforward and recurrent neural language models to benefit better from the deep structure and provide a slight improvement of recognition accuracy after interpolation with count models.To complete the overview, we include our initial investigations on the use of the attention mechanism for learning word triggers.

In this paper, we investigated various word clustering methods, by studying two clustering algorithms: Brown clustering and exchange algorithm, and three objective functions derived from different class-based language models (CBLM): two-sided, predictive and conditional models. In particular, we focused on the implementation of the exchange algorithm with improved speed. In total, we compared six clustering methods in terms of runtime and perplexity (PP) of the CBLM on a French corpus, and show that our accelerated implementation of exchange algorithm is up to 114 times faster than the original and around 6 times faster than the best implementation of Brown clustering we could find, while performing about the same (slightly better) in PP. In addition, we conducted a keyword search experiment on the Babel Lithuanian task (IARPA-babel304b-v1.0b), which showed that CBLM improves the word error rate (WER) but not the keyword search performance. Furthermore, we used these clustering techniques for the output layer of a recurrent neural network (RNN) language model (LM) and we show that in terms of PP of the RNN LM, word classes trained under the predictive model perform slightly better than those trained under other criteria we considered. Index Terms: word clustering, language modeling, neural network based language model, recurrent neural network, long short-term memory

In most of previous works on neural network based language models (NNLMs), the words are represented as 1-of-N encoded feature vectors. In this paper we investigate an alternative encoding of the word history, known as bag-of-words (BOW) representation of a word sequence, and use it as an additional input feature to the NNLM. Both the feedforward neural network (FFNN) and the long short-term memory recurrent neural network (LSTM-RNN) language models (LMs) with additional BOW input are evaluated on an English large vocabulary automatic speech recognition (ASR) task. We show that the BOW features significantly improve both the perplexity (PP) and the word error rate (WER) of a standard FFNN LM. In contrast, the LSTM-RNN LM does not benefit from such an explicit long context feature. Therefore the performance gap between feedforward and recurrent architectures for language modeling is reduced. In addition, we revisit the cache based LM, a seeming analog of the BOW for the count based LM, which was unsuccessful for ASR in the past. Although the cache is able to improve the perplexity, we only observe a very small reduction in WER. Index Terms: language modeling, speech recognition, bag-ofwords, feedforward neural networks, recurrent neural networks, long short-term memory, cache language model

In this paper, we describe the RWTH speech recognition system for English lectures developed within the Translectures project. A difficulty in the development of an English lectures recognition system, is the high ratio of non-native speakers. We address this problem by using very effective deep bottleneck features trained on multilingual data. The acoustic model is trained on large amounts of data from different domains and with different dialects. Large improvements are obtained from unsupervised acoustic adaptation. Another challenge is the frequent use of technical terms and the wide range of topics. In our recognition system, slides, which are attached to most lectures, are used for improving lexical coverage and language model adaptation.