Q1. Can you please briefly introduce yourself, SRIN, and the kind of work that goes on there? What project are you working on?

SRIN is the largest technology R&D center in Indonesia. I joined in 2014 as the Head of Software R&D Group. In that time, my team got assigned to develop applications for content and services, such as Samsung Link, ChatON, real-time communication for global services, Galaxy Gift, S LIME, and Salaam for local services in Indonesia. Along the way, SRIN has evolved naturally in terms of technology R&D competencies, especially in the areas of Internet of Things (IoT) and cloud and data management. Since 2019, the institute has increased global project contributions in SmartThings, Samsung’s private cloud and data management. Thanks to the collaboration of all involved Samsung stakeholders, SRIN teams are now working on many exciting projects that will improve technology creation and business productivity. In IoT, the institute has developed the Virtual Home plug-in for consumers, and Virtual Device and Bluetooth Low Energy (BLE) SmartTag (SmartThings Find) verification tools for developers. We also contribute to data management to give business decision-making insights in Samsung’s mobile and home appliance divisions, including local sales and marketing subsidiaries. In the cloud area, SRIN has developed middleware, such as Function-as-a-Service (FaaS) and object storage service (OSS), and various cloud applications, including data storage and Web Real-Time Communication (WebRTC). I am super excited to be part of the institute’s journey in achieving software development competencies.

Aside from software development, SRIN has also started research incubation for business analytics and artificial intelligence (AI) since 2016. Along the way, the institute has contributed various analytic projects for business data, especially on customer understanding and revenue growth prediction use cases. We also started to utilize the latest R&D advancements in deep learning for Indonesian speech and natural-language understanding (NLU) domains. We have implemented some features, such as automatic speech recognition (ASR), text-to-speech (TTS), and named-entity recognition (NER), for local services. SRIN will continue to strengthen the areas of business analytics and AI in the future, especially on applications of generative and reinforcement learning models in IoT and the cloud. We see a promising future for AI as a unique differentiator for all Samsung devices and services.


Q2. Please tell me the importance of your research field or technology. 

Despite the rapid progress in AI research and technologies, we believe that fundamental exploration has just started. Most initial AI models work within discriminative perspectives, usually to learn a direct map from inputs to the class labels or the posterior probability. However, classification can also be viewed from a generative perspective where an AI model learns the joint probability of inputs and labels and then performs classification using Bayes’ rules. To make AI more human, we foresee a multiparadigm future that combines both discriminative and generative models into a reinforcement paradigm, especially to make AI learn from environmental feedback. At SRIN, we are curious about the combination of discriminative, generative, and reinforcement learning models and how to apply them to develop useful technologies on Samsung devices and services (IoT and cloud and data management). 

In Samsung, discriminative and generative models are already applied in many AI domains, such as vision and natural language processing (NLP) and NLU. The discriminative models are used in many features of Samsung’s smartphones, such as Bixby, keyboards, biometric and security, and image/video enhancement. The generative models are mainly used to generate synthetic data with common use cases, such as realistic image generation, TTS, image super resolution, and speech and image restoration/reconstruction. The fast-growing use cases are supported by the rapid progress of deep learning in the field. We see this as an opportunity to enhance our current technology and focus on IoT and cloud and data management. 

SRIN’s goal is to improve productivity for both business and technology users of IoT and the cloud. To do that, we see the importance of adding reinforcement capabilities into current AI models. This will enable AI to learn from environment feedbacks and continuously improve its intelligence. Hence, we currently envision new features for IoT devices and cloud services by combining the latest Samsung R&D achievements on discriminative, generative, and reinforcement learning models.


Q3. Can you tell me about the main achievement and a rewarding moment in your research?

Being part of SRIN since 2014, the experience of rapid progress in human development is very rewarding for me. Since its inception, SRIN established a software engineering culture called FAST, which stands for Focus, Alignment, Scientific, and Talent. We hire local talents from top universities in Indonesia and develop their core competencies using FAST. It is amazing to see that the institute can contribute to the cutting edge of technological advancement. Starting from app development in 2012, SRIN engineers are now developing more advanced technologies, such as IoT device virtualization, private cloud middleware and applications, low code platform, and AI/Data Science. These bring firm confidence to the whole SRIN team and encourage them to contribute more. We are now able to create more patents for Samsung on AI and IoT. Also, we are giving back to the academic community by writing technical papers. For me as a professional, being part of SRIN’s journey is a valuable experience.

In terms of AI research, SRIN’s current focus are generative adversarial networks (GANs). Since 2014, many GANs have been defined using probability-based divergences and trained using stochastic gradient methods. Large-scale GANs have problems with training stability and consume huge graphics processing unit (GPU) resources when trained. At SRIN, we want to discover better divergence formulation and optimization techniques that can train stable GANs. It was a very exciting moment during my Ph.D. program when I found out that GANs can be trained using optimal transport (OT)–based divergence, such as Wasserstein and Sinkhorn that are known to have better stability, generalization, and sample complexity properties compared to common probabilistic divergences, such as Kullback-Leibler (KL) and Jensen-Shannon (JS). It opens an opportunity to develop a stable GAN that can be trained faster using fewer data. Other than divergence formulation, we are also interested in studying Bayesian techniques, such as Langevin dynamics, to solve minimax optimization in GANs. Some of our works are already published in the IEEE, International Conference on Machine Learning (ICML), and Association for Computing Machinery (ACM). More papers related to reinforcement learning models will be published soon.


Q4. What is your vision for the future and the goal you want to achieve?

Currently, SRIN is the number one technology R&D center in Indonesia. This fact gives competitive benefits for Samsung as Indonesia has been forecasted as the largest digital society and economy in Southeast Asia (SEA). At SRIN, we aim to contribute more to local, regional, and global business and technology developments. To do that, the institute is now preparing to closely collaborate with top universities in Indonesia not only for hiring top local talents but also for research. SRIN’s goal is to become a center of excellence in commercializing Samsung’s latest R&D achievements on IoT and cloud and data management differentiated with AI-based features. 

Specifically for SRIN AI incubation research, the short-term goal is to discover more efficient generative and reinforcement models that can be trained using fewer data and GPU resources. Despite the rapid progress in neural architectures, such as convolution, attention, and transformer, current models still require a large amount of data and computation. If we can discover a more efficient model, broader use cases can be possible. An example is the learning policy in federated imitation learning. Hence, in the long term, we want to monitor and gain a deep understanding of various generative and reinforcement models to design and implement useful AI features on IoT and cloud technologies. I believe SRIN is on the right track to realize that.