I’m going to focus here on ideas that will, I hope, inspire people to look more into HPC and computational science. I won’t cover the awards, although I congratulate all recipients and recognise their contribution to the field.
Once again I’m live blogging, unedited, unvarnished, and limited by my typing speed, so please forgive any errors and confusing bits. Ask questions, talk about the issues, engage with me! I’d love to chat.
Notes from the SC16 keynote
Keynote: Dr. Katharine Frase – Cognitive Computing: How Can We Accelerate Human Decision Making, Creativity and Innovation Using Techniques from Watson and Beyond?
Introduction: How do we make sense of the exabytes of information that have accumulated? HPC is helping us to understand ourselves, our world, and even our universe. And the future will bring even more exciting discoveries.
John West from TACC. HPC community faces two key challenges – first we need to grow the workforce to meet the growing demand for HPC. 200,000 jobs in computing go unfilled each year (in the USA). HPC is just a portion of this, but recruiting HPC professionals today is already hard and going to get worse. We need more minorities, we need to cast the net wide and deep, bring in every voice, every budding expert, and every moonshot idea. We need diverse backgrounds to bring diverse ideas to the table if we want to be successful in meeting our challenges.
22% of the SC16 student class are female. Women and minorities often don’t choose computing as a profession due to the perception that computing doesn’t have a direct impact on society. This needs to change!
We could send about 450,000,000 snapchats in the next second over SCinet. Amazing bandwidth!
Dr Katharine Frase, IBM.
Think about how you got here. Most of us relied on our phones to wake us up, remind us what we’re doing today, update us on what happened overnight, track our steps, navigate with GPS, and give us traffic information. Imagine that you were here at SC06. Almost everything I just said was not true. Almost without realising it our lives have been transformed. By the use of data, algorithms, and to some degrees learning our preferences.
We need to move beyond answers, however adaptive, flexible and realtime. How do we move towards co-creation? Co-hypothesizing? How do we move towards systems that can help us ask the right questions?
The challenge of big data is not always that it’s big. People in this room have been extracting enormous insights from data for generations. But it’s not the structured data that we’re drowning in. it’s the unstructured stuff. It’s the velocity of the data. how do we separate the signal from the noise? And it’s the veracity. The truthfulness of the data.
A project with the city of Dublin to figure out how we could get better info from the GPS systems on their buses. Everything requires transfers in Dublin, you can’t get one bus. You have to keep transferring. So everytime a bus is late you jeopardise your transfer. Couldn’t we use the GPS system to track each bus and find out what the best bus route is for today, given how buses are actually moving, rather than how they are scheduled to travel?
We found there were a number of buses that thought they were at the bottom of the Liffey river or in the basement of the Parliament building. It’s noisy data. So how do we extract insight without spending all our time cleaning up the data? How do we rely on the wisdom of crowds?
We focus on the architecture and the technology, but we need to think in another way – how do humans and systems interact? The first era of computing was counting things. The second is programmable systems. We get the same reliable answer to the same question regardless of how you ask it. We decided what those questions were going to be according to the structure of the system. We learnt to speak the way a computer thinks. The computer shaped our capabilities. We’re still, even with more usable languages like Python, learning to speak the way computers think.
So what do we mean by a cognitive system? What’s different?
The first is that a cognitive system understands language the way people use it. Natural language, machine learning, etc.
The second is that the system learns rather than being programmed. So the system gets smarter every day (with the right feedback) which means we can get a different answer depending on context – on where I am, or when it is. It responds to the most recent information. Maybe what I thought was true yesterday isn’t true anymore. Or I’ve given feedback to the system so that it gives you a subtly better answer next time based on your experience and feedback of last time.
The system can understand imagery, language and unstructured data as humans do. We have always built systems to expand human capacity, but computers have constrained our capacity as well.
I’m amazed now how anybody lived not knowing for 3 weeks at a time what the reaction was to that letter I sent. Can you imagine not being able to text your kids and find out where they are?
The challenge we face now is really one of complexity. If we’re going to have systems that help us think and decide, we need a better understanding of how people actually think and decide. Daniel Kahneman “Thinking Fast and Slow” describes the two sides of our brains. The fast side gives us reactions without us even knowing how we think about it, and the slow side thinks carefully about things. The slow side usually wins, but not right away.
You’re trying to think of the name of a guy you see in the grocery store that you know from work. The name Mike pops into your head, and you know it’s wrong, but you can’t get past that name to the real one. That’s the fast side blocking the slow side. Later the slow side pops up the name, but too late.
Humans have a tendency to prefer information that reinforces what we think we already know or what we expect. It effects what we retain from what we read and how we react when somebody argues with us. And we structure our experiments to find the results that we expect. We also overestimate the chance that something will happen that has happened before, and underestimate the chances of something happening that has never happened before (Trump!).
Cognitive systems can help us with this. It can tackle our unconscious cognitive bias, and take into account that the last coin flip does not influence the next one, even though a lot of people think it does.
A cardiologist felt he was at his best when he was in the hospital and there was a resident in the ward who could summarise the patient’s situation. History, overnight events, and test results. Then the cardiologist felt he could best use his experience in support of that patient. Katharine had the opportunity to have a copy of the documentation for each patient, and she participated in the panicked 5 minutes of reading through somebody’s chart before the next person walks through. We saw 5 patients that were very similar, and when the 6th patient walked in the cardiologist replayed the same script from the first 5, even though there were significant differences from the first 5.
Cognitive systems can be a statistically fresh set of eyes and help us become aware of, and deal with, our unconscious biases.
Machine learning: Image recognition and speech recognition. There is error in how humans recognise objects – sometimes you see something and think it’s actually something else. Speech is harder. The error rate is influenced by the size of the neural network to deal with it.
We now have massive datasets for training a system for speech, images, etc. We are trying to build systems to pass the radiology exam – a difficult problem even for humans. Can we train machine learning systems with huge datasets and have them do better?
You can’t just hand them unstructured datasets and have them come up with the right answer. They’re not omniscient. But we have huge increases in computational power and huge drop in cost, which makes these things possible.
The sheer volume of expertise being thrown at AI and cognitive computing is only increasing.
Why does this matter?
Why do we need to move forward in these new relationships with systems?
Katharine used to be involved with Watson, when it played Jeopardy. How do you go really native from language in the long tail on a game show, with poor English?
We went from there and tried to tackle oncology. We plugged in medical literature and started training the system to recognise the vocabulary of health. What is a symptom? diagnosis? treatment? We needed to add new features. In health timing matters – if you get the fever before you get the rash, it makes a difference. The system needed to learn what it’s reading and how humans interact with that.
We need to get expertise out to the broader community. Most people with cancer will not see an expert in their particular variation. They’ll see whoever is local and hope they get the right treatment. So we need to get that info and expertise out to improve the quality of local care.
U of Wisconsin is doing clinical trial matching. There are nearly 14,000,000 Americans with cancer. Less than 5% are involved in clinical trials. To match a patient to a trial there can be up to 46 characteristics you need to build that cohort. Gender, ethnicity, type of tumour, genetics, etc. You need to find that patient and reach the physician to suggest the treatment. So how can we use cognitive computing to get the word out to physicians that they might want to suggest this treatment?
Dept of Vet Affairs is trying to create a cognitive system to take the 10,000 vets with cancer and increase their capacity by 30x to get the right treatment to each indivudual.
Our systems have to fit into the workflow of those people who are doing the real work. We know the syndrome of an IT system that only makes sense to the computer scientist. The bigger problem is that it doesn’t fit into the workflow. The physician has to turn his back on the patient in order to do something with the system. So however capable, the system sits on the shelf.
The workflow of the human. The practitioner. How do we get the insights of everything we know we can do surfaced to the human in a way that is timely and fits into the way they want to do their jobs.
Picture this. A young girl has a persistent cough. It’s cancer. She gets referred to an oncologist, who sees her medical record. The cognitive system can tell the oncologist how it compares to the 15,000,000 cases the system has in its records. It prompts the oncologist to ask particular questions: like her own preferences about her care. For example it is important to her to keep her hair, or because she has small children she needs to come in less frequently. The system can take the specifics and recommend treatment options with likelihood of clinical success. The system reminds the doctor to ask her if anything has changed in the last two weeks. He would probably have asked her anyway, but the system reminds him. So a new piece of information comes in that can change the treatment options significantly. The options come back ranked according to success clinically, and according to how well they align with her preferences. The oncologist and patient can then make really informed choices.
If you like and trust your doctor it significantly improves your changes of a positive outcome, and you volunteer extra information. The system can help us in these dialog driven and intangible ways.
Let’s talk about congestive heart failure. Difficult to diagnose in advance, all you can do is treat it after it shows up, to mitigate the symptoms. The law has changed so that hospitals are no longer fully compensated if they readmit a patient for the same problem too soon. So a large hospital wanted them to help predict which patients they would see again soon. It turns out the data is there and it is possible to predict it. And it has nothing to do with clinical markers. The best predictor? Did they have dementia? Addictive behaviour? Living alone? These are all social effects that predict whether they will be back or not. It’s not really a clinical answer. It’s obvious when you look at it, but not necessarily obvious when you are looking from a clinical perspective.
86,000,000 people in the USA are prediabetic. The rates of diabetes are exploding. We have 56 years of info about diabetes patients. It’s not a perfect dataset, but it contains treatment as well as lifestyle stuff. So we want to create a cognitive database of this information so that we can start producing advice about things that can be done to delay the onset of diabetes for those that are prediabetic, or delay the consequences of diabetes such as blindness and gangrene.
Yes clinical data, genomics, exogenous data like fitbits are important, but the social context and cultural context is important. We can’t make the same dietary recommendations to all cultures, for example.
As we start amassing more data and looking at it as a whole, can we start identifying new factors that allow us to intervene earlier?
Let’s look at finance. Most audits rely on the structured financial info of the firm. But the real keys are not in that data. Most audit firms can’t manage the size of the real data, so they just take a sample and hope it’s representative. So can you use a cognitive system examine the full data of the firm and surface patterns to the auditors so that they can ask the hard questions and do the higher level thinking? We want the computers to do the routine things, the data intensive things, to free up the humans to do the higher level thinking.
What about World Peace? “I would have Watson help create World Peace. However, because this is unrealistic, I would have Watson improve education around the globe.” 6th grader.
This education challenge: we have unsatisfactory numbers for HS & college graduation rates, college debt, retraining our veterans and reintegrating them into society. In the developing world the problems are different. You can’t build enough schools or train enough teachers fast enough.
Education is very data poor. There’s no such thing as a clinical trial in education. Every trial is beset by small volumes of subjects and the fear that we didn’t control all the variables. Do we really know what we think we know about learning styles? about autism? A big data, data-driven approach could change this.
What could a system do to support a teacher in the classroom? It could grab all the info that’s already in the data systems of your school that you can’t get to, to give you insights into your students. How did they perform last year? Are they gifted? Are they troubled? The system can help you with the data we already have in schools that is not surfaced to the teacher. Teachers do their planning on the computer late at night. On the planning side the teacher can get up to the minute info about her students in all their classes. If John’s grades are dropping everywhere, it’s different to if he’s only failing in yours. You can get recommendations of curated content that the school has access that are tailored to individual students. She can chat with the other teachers who teach that student. She can get that content on her tablet and send messages forward to other teachers who are about to teach that student.
The ability to informally create a coordinated care feeling around that child is the single thing teachers find most valuable. Schools put in the effort to put care around the bottom and the top students. Those of us who have fabulous kids in the great unwashed middle, technology can help us surround those kids with personal care.
The more engaged a student is, the better they learn. They become learners rather than repeaters. They become explorers and discoverers. Sesame and IBM are developing platforms and products to address early childhood education. Not just to create smarter children but kinder children. Sesame tackles the tough issues, but til now it’s been in broadcast mode. It’s the same Elmo for every child. What if it weren’t? What if a parent could help design the way in which their child experiences it? My son is crazy about dinosaurs, can we do the alphabet content with dinosaurs? My daughter is crazy about horses, let’s put the stories in that context.
How do we engage children? 40 years ago TV was magic. Now kids try to interact with the TV. So how do we use technology to engage these youngest learners? There’s a vocabulary gap by the time the kids reach the age of 4 that is frighteningly predictive of the rest of their lives. How do we change that? How do we make it not be true?
Do doctors & teachers get better when they use cognitive systems?
How do you measure betterness? Most individual physicians don’t think they need the help. 🙂 But in the aggregate quality goes up. One of the big challenges for a teacher – you’ve always bee a 4th grade teacher. You’ve been asked to teach 3rd grade. It’s a new standard. What should you be expecting your students to come in with? You don’t want other people to know that you have questions. You’re afraid it will undermine your standing in the school. But we can provide an anonymous place to get that info without feeling like you’re embarrassing yourself.
This ability to start using systems in support of what we do in the classroom is extraordinarily powerful.
Education isn’t cancer research. The financial incentive to create these tools doesn’t exist. Fighting cancer is an industry with cash, education is not.
Schools are already spending money on tech. They’re not necessarily spending it in a way that helps the teacher. What is that threshold of affordability that starts getting to adoption, and then scales to help us drop costs. Every teacher and every school has particular things they wish the system could do. How do we start from the subset that is affordable and then we can raise the bar? As in every industry business process change is harder than we as technologists every want to admit that it is. So you have to start with the individuals that really want to change.
How do we deal with the lack of diversity in HPC and machine learning, and help the systems avoid the bias that comes with biased input data?
If the system gets biased data it’s going to give biased output. But the system can actually help us identify biases in the datasets. It can surface the problems like disproportionate gender representation, missing cultural/ethnic groups etc, that we might not otherwise be aware of.
Well that was awesome. So many different fields where cognitive computing can make a difference, so many ways HPC is changing the world. If these summaries are useful to you, please share, retweet, and leave me a message!