AI can be applied for different purposes under different parameters. Experts from prominent companies joined the “Best Practices to Get Started with Enterprise AI” panel discussion. On June 6, they shared the applications of AI in their companies and what challenges which can be solved with AI. The speakers for this panel discussion are: Alessandro Gasparini of ImmerVersion, Handol Kim of Quadrant, and Casey Yeh of CTCI; with Jesse Chen of GLORIA NCKU as the moderator.
ImmerVersion: Quality & Reliable Data is Needed
The panel started with the Executive Vice President & Chief Commercial Officer of ImmerVersion, Alessandro Gasparini. ImmerVersion is a company focusing on enabling intelligent vision on any machine. Mr. Gasparini stated that vision is a human's most important sense and also a validation of everything. For a machine, the challenge is to combine it with other sensory inputs. For this, Mr. Gasparini suggested fusing the sensors and let all the sensing be done in the pixels of images. When the sensing is dynamically synchronized with each frame of a video, the resulting information can generate less uncertainty than when used individually. The technology to realize this is “data in picture” technology or abbreviated as DIP.
The combined information from sensors, devices, users, and external data are stored together in a library, and then aggregated, compressed, before finally the information is encoded in markers in the images. When each frame of the video embeds data in picture, it can be dynamically synchronized for machine learning inference or data collection. The processing method will consume less power as all the information will not require external support from multiple devices that need to be synchronized. DIP is also contextual and supports different sensor types to gives a better idea of the environment of where the image was taken. Because DIP excels in data collection and storage, it can provide data which is both reliable and high-quality to train and run the systems.
Quadrant: Generative Learning Shines when Deep Learning Fails
The second speaker in the panel is Handol Kim, the General Manager of Quadrant. Quadrant is the generative machine learning business unit of D-Wave Systems which employs cutting edge energy based computation methods to solve previously intractable machine learning problems. According to Mr. Kim, AI is an industry designation, quite like semiconductor. The technology terminology should be Machine Learning which has many varieties; with deep learning being the most popular form. Deep learning as a system still has limitations, namely: the real world is not full of high-quality usable data; many problems are not classification problems, but different statistical problems that need to be solved; and datasets might not have the characteristics necessary for deep learning applications.
Deep learning is popular in the digital field due to how its performance and accuracy scale with the amount of data fed to the deep learning model. When the amount of data is small, only a small part of the datasets, messy, noisy, or mislabeled, generative learning can perform better. A particular situation where generative learning will be most useful is when the system is used to find anomalies, because deep learning might generalize and thereby fail in detecting said anomalies.
When users or researchers try and fit their data to a model, the obtained results may look better, but it might not reflect reality. The data could have an unknown attribute or information that the model cannot show. In this situation, using an unsupervised model to understand the structure or inherent data that exists from the latent space can yield better performance.
CTCI: Traditional Industries Have Many Opportunities for Startups
CTCI (中鼎集團) is one of the largest construction companies in Taiwan and one of the top 100 companies in the world. The speaker from CTCI is Casey Yeh, the Head of Group Research & Innovation Center. Mr. Yeh stated that construction and traditional industries in general are lagging in adopting new technologies. This means that there are plenty of opportunities for startups to offer their solutions and solve the industries' pain points. The main possible solutions to the pain points are industry 4.0 and the digital disruptions it can offer; and AI/ Machine Learning. He believes that the future of engineering and construction will be digitally connected, data oriented, and AI powered. In his speech, Mr. Yeh focused on the possible improvements in 4 areas: engineering information discovery, engineering design automation, safety & security monitoring, and inspection & progress tracking.
Engineering information discovery can use Natural Language Processing and object recognition which will read the documents; find the information scattered in multiple documents; read the diagrams to extract information; and afterwards combine the information into a single file. Engineering design automation will need to be automated with generative design ideas to not only reduce the time spent on the process, but also generate more design options. The engineers then can simply select the best option to apply to the problems. This segment still needs optimization and AI/ machine learning can help liberate the limitations of the designs. Major advances of monitoring technology are already employed for safety and security monitoring today; such as drones, robots, security cameras, and more, but there might be a more advanced application that can improve their performance. Inspection and progress tracking is often done manually, but Mr. Yeh hopes that image capture technologies can be implemented to do perform this function
The Q&A session of the panel started with a question raised by Jesse Chen, the CEO of GLORIA NCKU and the moderator of the panel before opening it to the audience.
Question: Taiwan Is Successful in Semiconductors, with TSMC as the Industry's Gold Standard. How Do We Create The TSMC Equivalent For AI In Taiwan?
Mr. Kim of Quadrant answered first and stated that government assistance is necessary. If the Taiwanese government believes that AI is a critical tech and almost a zero sum game, where whoever wins first will get most of the benefit; then AI must be made a national priority. The government then will need to treat AI the same way they did with semiconductors in the late 1970s, namely making something akin to Hsinchu Science Park for the AI industry. The government will also need to attract the Taiwanese talents who are currently working as premier machine learning researchers in leading companies such as Google Deepmind, Facebook, and more. They will need to return to Taiwan and to start their own companies. Starting companies will need funding which might not be easy to come by. Mr. Kim suggested giving government grants, tax holidays, and as many hardwares as needed for the undertaking to aggressively convince them to return and contribute.
Mr. Gasparini of ImmerVersion commented that government currently already has some initiatives to bring academia closer to the industry. A lot of the success will come from getting the initiatives of projects driven by industries and work together with academia to find solutions. Currently there are many problems that can be solved with AI/ machine learning so cooperation between academia and industry is necessary.
Mr. Yeh of CTCI said that talent is vital for any country to be a powerhouse in the AI industry. Taiwan has talent in the form of many smart people with good education that builds a great foundation as the value that Taiwan can provide. If the right direction can be found, capitals can be given to the correct people to build on AI.
Question: To Get a Total Solution for AI, There are 3 Important Components: Hardware, Software, and Cloud Platform. Which is the most important? How much focus should be given to it?
Mr. Kim of Quadrant stated that software is the most important. There are many hardware which are readily available in the forms of CPU, GPU, and TPU (Tensor Processing Unit). Aside from neuromorphic, TPU or new generation ASICs which are specifically optimized for neural networks or deep neural networks, there was not much differentiation in hardware. Enterprises also generally do not want to run their AI models in the cloud; they prefer to run it On-premise. Machine learning as a service is also difficult for startups to enter because of the heavy contenders such as Google, IBM, and Microsoft. In his opinion, the only area where innovation is possible is in software because AI logarithm is software.
Mr. Gasparini agrees on the importance of software. AI can only work if it's taught, so the system needs to learn using good & reliable data. This will help the system grow its knowledge and understanding and be more accurate in its predictions. Software is a key part of AI and many companies are working on the software side of AI. However, it is important to note that many software developers do not have reliable hardware to capture the data. This can be an area that needs to be focused on to developing more applications.
Mr. Yeh of CTCI also believes that software is the most important part of AI. To be able to use AI properly, the user will need specific training or education such as a data scientist, mathematician, software developer, or a combination of all of the above. A tool or toolset might be necessary for ordinary people to leverage the power of AI. This can be an opportunity also for the software aspect of AI.
At the end of the Q&A session, the moderator, Mr. Chen of GLORIA NCKU offered a different perspective to the question. In his opinion, the most important part of AI is defining the problems, parameters, and goals first to ensure that AI development is going to the right direction.
Watch the panel here: https://www.youtube.com/watch?v=T_BPmtHKrOg
InnoVEX will Return on May 29, 2019
InnoVEX 2018 hosted 388 startups, 17,867 visitors, and over 200 global investors; many of whom gave a positive impression of InnoVEX. After a stellar 2018, InnoVEX 2019 will be held in late May. As there will be a new exhibition hall for COMPUTEX, the scale of InnoVEX is also expected to grow even larger than this year. The registration of startups and speakers will be announced in October 2018.