9 October 2020In some other circumstances, different participants own different data and such data is not shared and creates multiple "data islands". Without joining the different datasets it is hard to get its real value. We are exploring cutting edge technologies like Federated Learning to break such barriers in the data islands to build an ecosystem for the industry while protecting data privacy and data ownership for the participants. Computational Resource is not cheapWith the development of new technologies like deep learning, computational resource has become critical for using the state-of-art algorithms. Just several years ago as a hobby I could run some machine learning models with some personal GPU cards. Now those pretrained models for computer vision and natural language processing usually contains millions of parameters and is way beyond standard for personal computing. Democracy for machine learning is in past tense and we have to pay more to get the enough computing resource for state-of-art algorithms, probably from cloud services like GPU on AWS or even TPU on Google Cloud Service.Product Development is not cheapOpen source covers the technology under the hood, but whether a company or a project can be successful totally depends on the product we provide. If customers don't buy your product, any fancy technologies behind it will make no sense. developing a cool product will require lots of work. A beautiful and simple user interface is usually the key.For example, we have seen a lot of open source autoML projects where data scientists don't need to tune the model anymore and let machine to decide which model best fits your data. A nice autoML product like Data robot will push that even further and let users build models without writing a line of code. Instead users just drag and drop the data, pick the prediction target and click run. Similar examples can be also seen such as Rasa for building chatbots and Prodigy for data annotation. In those products, a beautiful and simple user interface is usually the key.The same delicate product development goes for insurance products that are backed up by machine learning models. Risk thresholds need to be carefully tuned, privacy and fairness need to be thoroughly considered, claim frauds need to have low false positives, and personal recommendations need to be accurate and emotional.Domain Knowledge is not cheapI often get questions from my actuary and underwriting colleagues asking whether what we are doing are stealing their jobs. The truth is, technology is not stealing jobs, but to revolutionize them. As we mentioned before, label data is the most important part for training a supervised machine learning model. For industries like insurance, people cannot do the right annotations without domain knowledge. Besides, the predictions made by machines should always be audited by domain experts to ensure its correctness and fairness. AI is not only Artificial Intelligence, but more as Augmented Intelligence, where its goal is to make the job of professionals like underwriters and actuaries easier.One thing I find amazing is how people with domain knowledge like actuaries desire to learn data science. In the near future they will be the new data scientist with unparalleled domain knowledge.And Data Scientists are not (yet) cheapTo put everything above together, we still need the right people. Data science is still a new subject without a rigorous education system. It doesn't matter if data scientists appear either as PhDs from best universities in the world or just hold Coursera certificates as long as they can spot opportunities, solve problems and generate value. Companies are reaching out to those data science professionals and they all want to recruit the best, who are still quite expensive at this moment. Guan Wang
< Page 8 | Page 10 >