Cochlear implant (CI) is currently the only medical treatment available to partially restore hearing to patients with profound-to-severe hearing loss, and in recent years, CI has evolved into one of the most profound advances in modern medicine.

In this tutorial, we will first review fundamentals of CI speech perception. Specially, a vocoder model will be introduced to CI model speech perception, which helps to understand the perceptual impacts of important acoustic factors. In addition, combined HA and CI based speech perception will be reviewed, and CI speech perception and coding strategies for tonal languages will be presented. Objective methods to evaluate CI speech perception performance in noise will be introduced, including intrusive measures and non-intrusive measures.

The second part of this tutorial will present existing system architecture and fundamental theories of deep learning based speech enhancement (SE) approaches, and the reinforcement learning and generative adversarial network (GAN)-based SE methods. We will introduce speech enhancement studies to improve CI speech understanding in noise. Due to hearing loss, CI users normally have a reduced hearing range, which requires CI speech processor to compress the dynamic range of temporal envelope. This tutorial will also report some works that conduct adaptive envelope dynamic compression to improve CI speech perception in noise and in reverberation.

Experimental methods remain at the forefront to tackle challenging research problems. The rapid advancements and availability of computing resources encouraged a wide adoption of computer simulations to study mathematical models of various systems across many scientific and engineering disciplines. In biology and medicine, computer simulations are referred to as in silico experiments. They can greatly reduce the need to carry out time and cost demanding in vivo and in vitro experiments involving living subjects. However, unlike laboratory experiments, the in silico experiments are less understood how they should be designed to enable scientific discoveries and confirm existing knowledge that can be validated, explained and reproduced. This tutorial will review the traditional experiment design strategies, and how they can be adopted for the hypothesis testing and hypothesis generation by Monte Carlo simulations. It requires decisions on how long and how many times to run simulations, how to configure every simulation run, how to integrate simulation outputs, what data need to be collected and how they should be processed, how to choose appropriate statistical tests, and how to assess the quality of the results obtained. The in silico experiment design involves different statistical methods such as parameter inference, causal reasoning, hypothesis testing, sensitivity analysis, and event detection in order to improve the information gain and efficiency of Monte Carlo simulations.

Recently, we have witnessed a significant increase in fake news and media content, which has caused considerable unrest globally. Fake news has also been implanted to influence the financial market, geopolitical issues and affect democratic structures. Deep learning-based generative models have recently increased the complexity of fake news by creating ultra-realistic phony media content. These synthesized fake media contents are almost impossible to distinguish, even with state-of-the-art algorithms/software. It makes a threat to forensic set-ups, where videos and images are considered as indisputable evidence to provide justice. Therefore, many organizations are now focusing on deep fake detection research.

In this tutorial, we are going to discuss the recent development in deep fake research. Here we will demonstrate some of the state-of-the-art architecture to detect synthesized components of media.

The tutorial planning is as follows:

• First, we will discuss the most recent architectures for generating fake media content. Such as (a) DeepFake, (b) Face to Face.
• Next, we will discuss the state of art detection methods for detecting the generative components in the media. We will start with the famous FaceForensics++ architecture, and then we will discuss the following architectures:
  • FakeCatcher
  • MesoNet
  • Photoresponse non-uniformity (PRNU)
  • Temporal-aware pipeline to automatically detect deep fake videos
• We will discuss the deep fake detection challenge (https://ai.facebook.com/datasets/dfdc/ ).
  Also about the ongoing competition on Kaggle: https://www.kaggle.com/c/deepfake-detection-challenge/discussion/121223
• Finally, incorporate analytical models for better understanding the possible intention and impact of such fake images/videos on the security, economy, and social fabric of a nation.

The process of spoken language acquisition has been one of the topics that have attracted the most significant interest of linguists and human scientists for decades. One prominent and widely accepted explanation, proposed by Skinner in 1957, states that children acquire language based on behaviorist reinforcement principles by associating words with meanings. Along with the advances in Artificial Intelligence (AI) and Machine Learning technologies, we now have the means to construct and investigate engineering models on computers. This tutorial aims to review related works scattered in several areas and depict the roadmap to make robots efficiently acquire any language from scratch through spoken interactions just as human babies. We first survey researches in cognitive science and add some consideration from engineering viewpoints. We classify existing studies of language learning systems, clarifying what they can and can not. We then give lectures on unsupervised and reinforcement learning algorithms that form the basis of developing language acquisition systems. Finally, we introduce a toolkit that provides the research platform for automatic spoken language acquisition.

Federated learning (FL) has emerged as a communication-efficient distributed machine learning that preserves user privacy. This tutorial explains basics of FL at first, and then introduces its applications such as mobile keyboard prediction, Internet of things (IoT) security, mobile edge computing, and vehicular communications. Finally, we present recent works aiming at communication-efficient, fully-distributed, and model-free FLs. One of the key ideas is knowledge distillation, which is a powerful model training technique.

This tutorial introduces the prior-based image restoration methods. The problem of image restoration is to reconstruct an image from an observed coruppted image, such as noise, blurring, sampling, and pixel missing. Since the image restoration problem based only on the observation model is often an ill-posed problem, some methods based on the assumption of idmage priors are used.
This tutorial is divided into three parts. The first explains the basics of the image restoration problem and introduces the classical methods that have been used so far. The second introduces an image restoration method based on tensor representation. Finally, we will introduce an image restoration method that uses neural networks.