In this paper, we present an in-depth study on online attention mechanisms and distillation techniques for dual-mode (i.e., joint online and offline) ASR using the Conformer Transducer.In the dual-mode Conformer Transducer model, layers can function in online or offline mode while sharing parameters, and in-place knowledge distillation from offline to online mode is applied in training to improve online accuracy.In our study, we first demonstrate accuracy improvements from using chunked attention in the Conformer encoder compared to autoregressive attention with and without lookahead.Furthermore, we explore the efficient KLD and 1-best KLD losses with different shifts between online and offline outputs in the knowledge distillation.Finally, we show that a simplified dual-mode Conformer that only has mode-specific self-attention performs equally well as the one also having mode-specific convolutions and normalization.Our experiments are based on two very different datasets: the Librispeech task and an internal corpus of medical conversations.Results show that the proposed dual-mode system using chunked attention yields 5 % and 4 % relative WER improvement on the Librispeech and medical tasks, compared to the dual-mode system using autoregressive attention with similar average lookahead.

End-to-end (E2E) multi-channel ASR systems show state-of-the-art performance in far-field ASR tasks by joint training of a multi-channel front-end along with the ASR model. The main limitation of such systems is that they are usually trained with data from a fixed array geometry, which can lead to degradation in accuracy when a different array is used in testing. This makes it challenging to deploy these systems in practice, as it is costly to retrain and deploy different models for various array configurations. To address this, we present a simple and effective data augmentation technique, which is based on randomly dropping channels in the multi-channel audio input during training, in order to improve the robustness to various array configurations at test time. We call this technique ChannelAugment, in contrast to SpecAugment (SA) which drops time and/or frequency components of a single channel input au-dio. We apply ChannelAugment to the Spatial Filtering (SF) and Minimum Variance Distortionless Response (MVDR) neural beam-forming approaches. For SF, we observe 10.6 % WER improvement across various array configurations employing different numbers of microphones. For MVDR, we achieve a 74 % reduction in training time without causing degradation of recognition accuracy.

In this paper, we propose a dual-encoder ASR architecture for joint modeling of close-talk (CT) and far-talk (FT) speech, in order to combine the advantages of CT and FT devices for better accuracy. The key idea is to add an encoder selection network to choose the optimal input source (CT or FT) and the corresponding encoder. We use a single-channel encoder for CT speech and a multi-channel encoder with Spatial Filtering neural beamforming for FT speech, which are jointly trained with the encoder selection. We validate our approach on both attention-based and RNN Transducer end-to-end ASR systems. The experiments are done with conversational speech from a medical use case, which is recorded simultaneously with a CT device and a microphone array. Our results show that the proposed dual-encoder architecture obtains up to 9% relative WER reduction when using both CT and FT input, compared to the best single-encoder system trained and tested in matched condition.

The compaction of test programs for processor-based systems is of utmost practical importance: Software-Based Self-Test (SBST) is nowadays increasingly adopted, especially for in-field test of safety-critical applications, and both the size and the execution time of the test are critical parameters. However, while compacting the size of binary test sequences has been thoroughly studied over the years, the reduction of the execution time of test programs is still a rather unexplored area of research. This paper describes a family of algorithms able to automatically enhance an existing test program, reducing the time required to run it and, as a side effect, its size. The proposed solutions are based on instruction removal and restoration, which is shown to be computationally more efficient than instruction removal alone. Experimental results demonstrate the compaction capabilities, and allow analyzing computational costs and effectiveness of the different algorithms.

The iterated prisoner's dilemma is a famous model of cooperation and conflict in game theory. Its origin can be traced back to the Cold War, and countless strategies for playing it have been proposed so far, either designed by hand or automatically generated by computers. In the 2000s, scholars started focusing on adaptive players, that is, able to classify their opponent's behavior and adopt an effective counter-strategy. The player presented in this paper, pushes such idea even further: it builds a model of the current adversary from scratch, without relying on any pre-defined archetypes, and tweaks it as the game develops using an evolutionary algorithm; at the same time, it exploits the model to lead the game into the most favorable continuation. Models are compact nondeterministic finite state machines; they are extremely efficient in predicting opponents' replies, without being completely correct by necessity. Experimental results show that such a player is able to win several one-to-one games against strong opponents taken from the literature, and that it consistently prevails in round-robin tournaments of different sizes.

While compaction of binary test sequences for generic sequential circuits has been widely explored, the compaction of test programs for processor-based systems is still an open area of research. Test program compaction is practically important because there are several scenarios in which Software-based Self-Test (SBST) is adopted, and the size of the test program is often a critical parameter. This paper is among the first to propose algorithms able to automatically compact an existing test program. The proposed solution is based on instruction removal and restoration, which is shown to significantly reduce the computational cost compared with instruction removal alone. Experimental results are reported, showing the compaction capabilities and computational costs of the proposed algorithms.

Evolutionary algorithms greatly benefit from an optimal application of the different genetic operators during the optimization process: thus, it is not surprising that several research lines in literature deal with the self-adapting of activation probabilities for operators. The current state of the art revolves around the use of the Multi-Armed Bandit (MAB) and Dynamic Multi-Armed bandit (D-MAB) paradigms, that modify the selection mechanism based on the rewards of the different operators. Such methodologies, however, update the probabilities after each operator's application, creating possible issues with positive feedbacks and impairing parallel evaluations, one of the strongest advantages of evolutionary computation in an industrial perspective. Moreover, D-MAB techniques often rely upon measurements of population diversity, that might not be applicable to all real-world scenarios. In this paper, we propose a generalization of the D-MAB approach, paired with a simple mechanism for operator management, that aims at removing several limitations of other D-MAB strategies, allowing for parallel evaluations and self-adaptive parameter tuning. Experimental results show that the approach is particularly effective with frameworks containing many different operators, even when some of them are ill-suited for the problem at hand, or are sporadically failing, as it commonly happens in the real world.

The iterated prisoner's dilemma is a widely known model in game theory, fundamental to many theories of cooperation and trust among self-interested beings. There are many works in literature about developing efficient strategies for this problem, both inside and outside the machine learning community. This paper shift the focus from finding a “good strategy” in absolute terms, to dynamically adapting and optimizing the strategy against the current opponent. Turan evolves competitive non-deterministic models of the current opponent, and exploit them to predict its moves and maximize the payoff as the game develops. Experimental results show that the proposed approach is able to obtain good performances against different kind of opponent, whether their strategies can or cannot be implemented as finite state machines.

Branch Prediction Units (BPUs) are widely used to reduce the performance penalties caused by branch instructions in pipelined processors. BPUs may be implemented in different forms: the Branch History Table (BHT) is an effective solution when the goal is predicting the result of conditional branches. In this paper we propose a method to generate test programs able to detect faults affecting the memory existing within a BHT implementing the correlated predictors approach. Our method is particularly suited to be used for the in-field test of a processor and allows detecting any stuck-at fault in the BPU memory. The method does not require the detailed knowledge of the BPU implementation, but only relies on the key parameters of its architecture. We gathered experimental results using the SimpleScalar environment.

Defining a distance measure over the individuals in the population of an Evolutionary Algorithm can be exploited for several applications, ranging from diversity preservation to balancing exploration and exploitation. When individuals are encoded as strings of bits or sets of real values, computing the distance between any two can be a straightforward process; when individuals are represented as trees or linear graphs, however, quite often the user must resort to phenotype-level problem-specific distance metrics. This paper presents a generic genotype-level distance metric for Linear Genetic Programming: the information contained by an individual is represented as a set of symbols, using n-grams to capture significant recurring structures inside the genome. The difference in information between two individuals is evaluated resorting to a symmetric difference. Experimental evaluations show that the proposed metric has a strong correlation with phenotype-level problem-specific distance measures in two problems where individuals represent string of bits and Assembly-language programs, respectively.