Key Takeaways from AI Conference in San Francisco 2017 – Day 1
Highlights and key takeaways from day 1 of AI Conference San Francisco 2017, including current state review, future trends, and top recommendations for AI initiatives.
Abu Qader, CTO, GliaLab had a fireside chat with Peter Norvig, Director of Research, Google on “Deep learning to fight cancer”. At an early age of 15, Abu was fascinated by AI technology and it’s potential. He started doing the MOOCs (including Andrew Ng’s ML class on Coursera) and playing with the open-source code on Github to understand how AI works. Soon, he started applying AI to real-life uses cases. One such endeavor led to the development of a breast cancer diagnosis software that works for more accurately and more efficiently than radiologists.
Talking about challenges, Abu mentioned that getting training data was indeed the biggest challenge as most of medical data is locked up due to regulatory and privacy concerns. Thus, he was limited to the open-source datasets to train his model on detecting and diagnosing anomalies in mammograms.
Abu emphasized that in healthcare it is not sufficient to focus merely on accuracy; sensitivity and specificity are equally important. A small 1% change in model accuracy can literally impact the lives of thousands of people. Thus, one needs to design the data model accordingly.
Vijay Pande, Andreessen Horowitz gave an insightful keynote on “How AI is ushering in a new era of healthcare”. He highlighted that the AI revolution is partially triggered from the exponential decrease in the cost of technology (both compute and storage) – from impossible, to possible, to free (almost!) in next 20 years.
The new sources of data are playing an important role too. The wearables and smartphones are producing a large amount of healthcare data that can be analyzed for insights and even, real-time feedback. The cost of sensors is radically decreasing. This is leading to new possibilities. As an example, he talked about a company called Cardiogram that uses pulse data to predict heart issues.
Genomics is another data source for healthcare data. Genomics data has been increasing at a faster rate than Moore’s law, thanks to sharp drop in the sequencing costs. He emphasized the importance of studying DNA as it can provide real-time insight into our bodies as of now. DNA data can be the test bed for a large varieties of disease diagnosis.
Biology is super-complicated; maybe it’s too complicated for human brains to comprehend. Thus, AI provides an unprecedented opportunity to augment human brain capabilities and making it possible to understand various biological phenomenon deeply interlinked with our health.
Vijay suggested that instead of obsessively searching for the cure of cancer (as if there would be some magic pill for it), we should divert our attention to early detection of cancer so that they can be prevented altogether or treated in a better way. He mentioned that more than 80% of breast/ovarian/prostate/lung cancer deaths are entirely preventable if detected early.
Vijay concluded that the common theme behind successful AI companies is a strategic understanding and deployment of “data network effects”.
Andrew Ng, Co-founder and Co-chairman, Coursera gave a highly thought-provoking keynote on “AI is the new electricity”. As the invention and adoption of electricity revolutionized every industry about a century ago, similarly AI has a clear path now to transform every industry. We are building an AI-empowered society.
As of today, almost all the economic value of AI comes from supervised learning, which is critically dependent on good quality labeled data. Supervised learning is no more limited to a binary output (such as approve or reject loan application) to assist decision making; rather it is also giving output as text (converted from speech or a picture) and audio (converted from text).
With an increasing amount of data, the neural networks perform much better than the traditional ML methods. He joked that Deep Learning is a great brand for neural networks.
Besides supervised learning, we are also seeing growth in the implementation of transfer learning. Unsupervised learning is still in a very early phase of practical applications despite enormous potential. Reinforcement learning is great for applications such as gaming and robotics, but it also has the limitation of heavy dependency on the “reward” data – this dependency can be even worse than that for supervised learning.
Another significant trend in AI is the move from structured data to unstructured data. The former is very machine friendly, such as database tables, lists, etc. The latter is human friendly, such as speech and vision. Structured data can lead to vertical-specific insights, whereas unstructured data can lead to a wide-range of insights and innovations.
From a business strategy perspective, he highlighted that the barrier to entry is no more the technology, rather it’s the data. So, a key part of the strategy of an AI company should be the virtuous cycle of: Data -> Product -> Users -> Data. This data accumulation loop helps you create strategic advantage compared to your competition.
It is important to understand that AI is just not merely machine learning, it also includes methods such as graphical models and knowledge graphs. However, in recent years the research and applications of machine learning have just exploded relative to other methods. The machine learning (and deep learning) methods are best suited for situations where there is a large amount of data and relatively, small amount of human expertise – for example, online advertising. On the other hand, for situations with small amount of data and large amount of human expertise, other approaches such as graphical models are more helpful.
He suggested that the biggest untapped opportunities today are in areas where there is a large amount of data available, to which latest deep learning methods can be applied.
In his advice to engineers, he stressed upon the importance of reading academic papers and implementing those approaches to replicate the results. That will not only lead to a deeper understanding of the advanced concepts, but also help you in coming up with new ideas.
As the previous 2 decades were dominated by the rise of internet companies, the near future would be dominated by the rise of AI companies. Andrew highlighted that one does not become an internet company just by having a website, rather the core characteristics of an internet company are A/B testing, short cycle times, and data-driven bottom-up decision making. Similarly, the core characteristics of an AI company are strategic data acquisition, centralized data warehouses, pervasive automation, and new job descriptions.