ISCAS 2015 Disruptive Engineering and Societal Impact



welcome everybody is my pleasure to be here today so my name is David attends i'm professor of EPFL in electric engineering and the most important role that i have here today and vice president of conferences of i triple e consider authorities animation at the policy de for many people so our keynote speaker today dr. Steve eggless who is affiliated with the Stanford University's Stevie's technology stand executive industry executive for more than 30 years he has been working on data science energy clean tech material optoelectronics so it's a very large and broad spectrum very much interesting in the industry to how to link industrial work and actually new research that is going to do is done in hindu in academia in this part he is going to be one of the very innovative speakers that we have in this conference in the area design automation he is living at Stanford right now three different main topics one is the Stanford data science initiative he's also the executive director of the artificial intelligence lab and he's also managing the secure Internet of Things project a very large effort in stand for all of them are going to be the focus of today's keynote where he's going to try to link the uptick technologies in engineering and how to make an impact in society and I think a very important hospital he has had while doing this a performing a great speaker is how to make able how to make people really be aware of these are being able to really cope all these things together and for doing that one thing that maybe not too many people know and is interesting is that he is a semi professional soccer or football referee so he's already dealing everyday almost with problems on the field and how to deal with very conflicting behavior of people so I think it is one of the main things we hopefully we're going to understand today how he does that and I think he will propose in this way but innovative ways how to link all these things together to make a compromise so that people are happy about the solution so please join me to welcome Steve on the podium thank you very much d for being here today thanks David it's great to be here as David said I'm going to talk about disruptive engineering and societal impact when I talk about disruptive engineering or disruptive innovation what I mean is a substantial and abrupt change in performance where we're moving off of a smooth increase in performance over time through some new invention some new development that causes a discontinuous and significant change in improvement this is something we all strive to do in our own work and how do we do it it requires invention and innovation it requires creativity and other things second thing I want to talk about this afternoon is how to achieve impact at scale how to achieve large-scale impact scale matters if you want to change the world scale matters if you want to reshape an entire industry and scale matters if you want to create new billion dollar class businesses for one of the largest corporations on the planet how does one get to impact its scale it certainly requires an awareness of real-world opportunities challenges and constraints this is something that corporations can often contribute it requires invention and innovation this is something that universities and national labs are often particularly good at it requires sufficient resources and it requires access to the market how do we get social impact out of the engineering and invention we do of course there are many different kinds of social impact social impact can be helping some of the neediest and poorest people on the planet or helping some of the wealthiest and most well-off people social impact can be good or bad and we'll touch a bit on both kinds of social impact this afternoon it's interesting to me that the technologies that lead to social impact are sometimes some of the hardest and most challenging problems around the sorts of things that any academic working to get tenure might want to go after but not always sometimes what's required isn't high tech but the right tech and we'll take a look at examples of both types we'll also think about the different classes of customers indeed customers can be some of the poorest people or some of the best off and the opportunities amongst the poorest people aren't always opportunities that require philanthropic work sometimes the fortune at the bottom of the pyramid can actually involve selling products and services to some of the poorest people and we'll see an example of that this afternoon as well with my coworker Sarah risk we've taken a look at examples of particularly successful and impactful innovation and we found that they tend to fall into one or more of four categories first of all improved solutions too familiar problems these are problems that have been around often for a very long time in many cases people had given up on finding innovative solutions but the innovator in the entrepreneur is able to come along and identify a new and different approach to a problem that's been around for a long time secondly basic research research in fields like mathematics or chemistry can often lead to the kind of breakthrough innovation and understanding that can allow us to be successful with disruptive innovation at scale the arc of technology evolution we're all involved in technologies that are getting progressively better over time perhaps one of the best examples in modern society is circuits and systems over the last few decades often successful innovators are able to realize that this arc of technology innovation has reached a level where whole new things are possible that weren't possible before in the past we can imagine that leading to things like cell phones and then smartphones and then fourthly business model innovation in this case the innovation isn't so much the technology but finding new business models new ways to define who the customer is new financing and payment schemes and we'll see examples of all of these another concept that's useful and important in disruptive innovation is when two different fields are combined to create significant and nonlinear opportunities an obvious example to all of us in the room is the explosion of data science capability today with the incredible capabilities of modern circuits and systems in data science we have new sources of data and more sources of data than ever before due to the ubiquity of sensors what we all call the Internet of Things new people collecting data and so forth we have new algorithms and software techniques leading to machine learning deep learning data mining and so on thanks to the work of many of you in the room we have incredibly powerful compute and storage is getting much cheaper as well and we have the communications bandwidth to tie all of this together and make it possible collectively this is what a lot of people are calling Big Data alongside of this we have the incredible capability of circuits and systems we now have processors sensors and ad hoc wireless communications virtually everywhere we have high-performance processors with incredibly low power consumption increasingly with reconfigure fpgas at the front end providing more powerful computation capability in locations where we didn't have it previously the result of all of this is we can now add compute and communications to virtually everything for nearly zero cost combining this explosion of circuits and systems capabilities with big data can lead to lots of new opportunities one of those opportunities is what I think of as the future of machine learning we're the combination of massive amounts of labeled data new more powerful compute and powerful machine learning algorithms means to a greater degree than ever before we can answer questions and predict the future I'm going to talk about these concepts through five examples how to create a secure and private Internet of Things how to enable an agile and flexible electric grid then two examples from the social sciences how to detect and identify human trafficking on the web and in the criminal justice system how to help judges make more informed decisions about defendants in front of them and then finally from the developing world how to replace kerosene lanterns with photovoltaic panels batteries and LEDs so let's talk first about creating a secure and private Internet of Things the work that I'm talking about here is from levis ohne Horowitz and co-workers at Stanford we all know what the Internet of Things is we have all kinds of devices smartphones wearable medical devices thermostats and so on they can communicate occasionally with each other more often with the web sometimes using a smartphone as a gateway in the web data from these devices can be combined with other devices perhaps from your friends or colleagues or neighbors data that's available and then that data can be sent back to you and other end users cool no well yes and no because we know that this is also big threat to security and privacy in fact a recent study by hewlett-packard found a wide range of security vulnerabilities for example they found that seventy percent of the Internet of Things devices and systems they looked at didn't even require secure updates to the operating system so a bad actor could get in and take control of your device through an update why is this why is security of these Internet of Things devices and systems so lousy well it's because it's a really hard problem for one thing these are complex distributed systems we have sensors at the edge we have gateway devices we have cloud computing and end-user devices each of these typically has a different operating system programmed in a different language there are vast differences in the resources available in terms of compute power the communications methodologies being used and so forth there are also massive streams of real-time data the volume of data is large the net result of all of this is a very large and a varied attack surface traditionally security has been worried about after the system has been designed those of you who have designed and implemented systems like this yourself know that you typically do the design finish the design and then ask how security can be added as a sort of bandage or blanket we also know that traditionally as engineers and designers we have to choose between convenience and usability on the one hand or security on the other and under the pressure of trying to get a product out or trying to get a PhD finished we all know how that trade-offs going to work what's needed the goal is to try to break that need to choose between usability and secure and so the question that these researchers put in front of themselves is if it's possible to design an Internet of Things system from the ground up that can be both highly secure and usable and they believe it can and the path to doing that they believe is a complete new design that includes data security and system security I won't talk about all aspects of this this afternoon but it involves new cryptographic systems it announced new operating systems that can carve out safe regimes for untrusted applications alongside of trusted applications a critical part of this is the concept of end-to-end security enable perhaps by homomorphic encryption and that's what I want to talk about for a minute when we say end-to-end security the concept is that the devices at the edge where the data is being correct collected through sensors or some kind of user interface will encrypt the data and then as I show in the diagram here the data will stay encrypted through the Gateway device a smartphone or whatever it might be up to the cloud we're computation statistics trends and so on are calculated and back down to the end-user the data only finally decrypted when it gets back to the end user application if this can be made to work it's a huge help to the problem of security because now even if the cloud is compromised the data is encrypted the key to doing something like this maybe homomorphic encryption as many of you know this is an encryption technology that enables computations to be formed on encrypted data today this is still not yet practical although it's theoretically possible but the computational burden can be several orders of magnitude compared to doing the same operation on unencrypted data but increasingly developers have found techniques to make home morphic encryption practical for certain specific operations that are identified in advance and so one path to a secure internet of things may be to further develop the encryption algorithms to enable us to perform unencrypted data in the cloud the kinds of calculations that are going to be needed for these Internet of Things applications the kind of things that will be necessary to determine averages trends outliers and that sort of thing so let's summarize this example we see the potential for an improved solution to a familiar problem the problem of security and privacy on the Internet of Things has been around for a while but perhaps a complete new system design and the use of new encryption and cryptography technologies might make this possible we see basic research in cryptography and mathematics playing a role here and we see this solution exploiting the arc of technology innovation through the sorts of micro processors at the edge that are powerful enough and sufficiently low power to enable this kind of end-to-end security what are the disruptive innovations here new encryption methods and these low power high performance chips now let's take a look at the modern electric grid you all understand the problem and the challenge of the modern electric grid we have no energy storage so generation and load need to be matched instantaneously and continuously on the electric grid this is a couple of really unfortunate consequences one is in any modern society where the goal is to avoid brownouts and blackouts the generation capacity on the grid needs to be sufficient to meet those few times of peak energy demand in many parts of the world on the hottest summer days when everyone's running air conditioners and all the factories and data centers are running in many cases those last few power generating plants are only run a few hours a year they're just there for those times of peak use which is an incredibly inefficient use of capital in addition generators need to be ramped up and down to meet variations in load as you can imagine ramping something like a large gas turbine up and down in output quickly can shorten the life of and be hard on the generator interestingly that's a problem that's made worse by all of the renewable wind and solar generation that's being put on the grid increasingly in regions that have lots of solar production demand is relatively low during a sunny afternoon but as the Sun sets late in the day the load is not dropping off so quickly and so the net load the total load minus the solar generation can increase at a dramatic rate that puts a huge strain on the generators in addition solar and wind are inherently intermittent clouds can get in front of the Sun wind can gust or fall off and in the modern grid all this needs to be taken up by adjusting generators the result is capital in efficiency energy and efficiency greater air pollution and greater greenhouse gas emissions many people believe that solution to this is to exploit demand-side flexibility to not put all the burden on adjusting generators but to have the ability to adjust load we have the technical capability to do this today we have sensors throughout the transmission and distribution networks we have the sorts of low-cost communications that allow generators and grid operators to communicate with the loads and we have new Smart Grid technologies to tie all of this together and make it possible including smart meters with real-time two-way communications the challenge in doing this is to predict and quantify how this would actually work to understand why this is a problem of prediction and quantification imagine for a minute that you're an electric utility and you want the ability to curtail load during a time of peak demand you might send a signal out to turn off some loads but if you get it wrong and you turn off too many loads you've lost revenue if you get it wrong in the other direction and don't curtail enough demand you're going to have brownouts or blackouts anyway so the problem is one of prediction the ability to predict an executed demand response system comes from understanding the load shape or the load profile if you average the load profile across a large area you typically see a double peaked shape like this particularly in areas where there's a lot of residential demand and people have long understood this as being a result of our normal daily cycle we wake up in the morning we use a lot of electricity and we go to work and our homes use less although of course the factories and businesses use more during the day and then we come home at the end of the day and again we use a lot of electricity previously efforts to design demand response systems assumed that every user had this double peaked profile in fact as we'll see in a minute that's far from true and this is one of the themes of big data to be able to get away from averages and to be able to look at individual users or at least small populations of similar users in many cases this is obvious if you're in the public health area you clearly don't want to average everybody's wait the heavy people will cancel out the light people it'll look like everyone's the right weight when in fact you have some people who need to lose weight and some people who need to gain weight and a similar thing is going on here but it has everything to do with size prior studies involved about a hundred customers about 9,000 load shapes in the study i'm about to tell described to you rajagopal at all looked at more than 200,000 customers and 66 million load shapes and the challenge was to segment these customers in these load shapes into a meaningful set of groups indeed one of those groups was this familiar double peaked profile that you can see on the left but as you see here many many users had load shapes that were very different from the double peaked load shape and by understanding this heterogeneity in the users where jag appalled at all were able to design a demand response system that targeted the right users the users that had the potential to actually cause a reduction in load and then they worked with the utilities to design an economic incentive so that the people who agreed to curtail load would get compensated for their inconvenience and the result was a demand response system that's now being used widely and is much more predictable than any that had come previously so how does this work how do they get a flexible and predictive demand response system to identify the right customers to design the program around those customers and offer economic incentives and the result in its predictability once you've done that you can either have a passive system where a utility or a grid operator sends a signal that perhaps postpones running the compressor on your refrigerator for a few minutes or perhaps automatically adjusts your air conditioner to a slightly higher temperature setpoint or they can be active systems where the users have to actively opt in most people think that residential systems will increasingly be passive and systems at major factories and businesses where they have a full-time energy management engineer might be active if we take a look at those classes of inquiry we see an improved solution to an old problem we see exploiting the arc of technology innovation in segmentation algorithms and smart grids and it turns out we also see business model innovation companies are springing up over the last few years exploiting this sort of thing either to offer services to utilities they go to utilities saying we will design and operate a demand response or energy management system for you or increasingly these startup companies are trying to compete with utilities in many parts of the world utilities no longer have a legislated monopoly and so we see startup companies acting like utilities but doing it in a smarter more intelligent agile and presumably more economically efficient way I don't want to switch to the social sciences and take a look at the problem of human trafficking on the web and what's being done about it this is work by Chris Rea kafir Ella and co-workers at Stanford the University of Michigan and a company called giant oak what is human trafficking human trafficking is the use of physical financial or cultural forces to try to coerce behavior out of someone it's found in manual labor it's found in the sex trade and it can be found in other areas human trafficking is proven extremely difficult to quantify and extremely difficult for law enforcement officials to find in fact the Internet has become the friend to the criminals who engage in human trafficking because it allows them to communicate and advertise widely but also hide in anonymity what's the problem well first of all the problem is the human trafficking itself this is a human tragedy second of all as I mentioned the internet makes it easier for the people doing this to operate and to hide the challenge in cracking down is multi-dimensional but one big part of it is that there are Lee activities taking place alongside of the illegal ones there are legal sex workers there are legal people providing laborers in the construction industry and so any effort to crack down on human trafficking has to somehow avoid capturing the legal workers in its net and we'll see that's possible using data mining data analytics using very large data sets and the work I'm going to tell you about is already now successfully providing law enforcement officials with actionable information let's take a look at how this is done it's really done through creating a database and here i show schematically how that works the database starts with the advertisements that the human traffickers are using on the web to advertise these forced laborers and sex trade workers the humans then step in to create a database schema what are the parameters that we want to try to extract from every advertisement the services offered the phone numbers the prices the names of the people the cities will see that the geographical information is quite important then the hardest part of this is populating that empty database with the information contained in the advertisements this is done using techniques of natural language processing and also a dramatically powerful new computer program called deep dive that I'll tell you about in a moment once the database is populated you can then start to do statistics on it and see if you can identify the bad guys and provide actionable information to law enforcement deep dive is a scalable and high-performance inference and learning engine it uses machine learning and natural language processing to collect and combine three kinds of data unstructured data sort of thing that you've all is already today when you use something like Google search to look through text also highly structured data the sort of thing that you can look at in a relational database and most challenging of all semi-structured data the sort of data in tables and graphs and charts and figures and diagrams in many of our journal articles that are the hardest of all for different kinds of machine based algorithms to get at deep dive has already been used to build more than a dozen major applications it's indexed several massive databases including pubmed central biomed central the entire google patents database wikipedia numerous libraries and web crawls and you can check out deep dive for yourself and use it yourself the software is available online what was the size of the database in this human trafficking problem they looked at more than twenty seven point four million advertisements on the web they extracted more than 11 million names of individuals more than 23 million phone numbers and similarly large numbers of services prices geographical locations and so on back to this question of how do you distinguish the bad guys from the legal operators one is by looking at geography trafficked individuals are more likely than legal individuals to move from one location to another and so if the same people are identified sequentially in multiple locations it was found that that correlates with being trafficked in addition the operators of human trafficking will often go by several cell phones at once and have a set of sequential phone numbers also people engaging in human trafficking might use the same phone number a quoi across a wide range of advertisements and most intriguing of all the price structure legal operators whether it's in labor or the sex trade tend to charge priceless for activities that depend on the hazard and abuse risk of the activity there was found to be a statistical pattern to the prices charged the trafficked operators and individuals find it much harder to do that as strange as it is to talk about this when we're talking about human lives they exist in a competitive economic environment and for a variety of reasons they can't compete effectively with the legal operators and so they end up charging less for the same activities here I've shown a plot of the frequency or occurrence of a particular price in dollars per hour with as a parameter the unemployment rate in those areas each of these ads has a geographical area there's a number of things that are interesting about this first of all just the notion of combining the data that came off of these twenty seven point four million web advertisements with data from another source in this case government collected publicly available data on unemployment rates so this notion of collecting and combining different kinds of data can be very powerful second of all what the data itself is showing us that in regions with low unemployment the trafficked workers are charging more and in regions of high unemployment the traffic workers are charging less so we see macroeconomics at work even in human trafficking not sure I would have expected that prior to seeing the result so the result is by combining all those techniques they're able to do a good job of identifying human trafficking advertisements but they can then provide to law enforcement officials is a set of scores that rank the likelihood of somebody being an operator of human trafficking from the most likely to the least likely it's geographically specific information so they can provide to local law enforcement officials likely examples of trafficking in their area and of course there's lots of other data other than a single score so law enforcement officials can combine this information with other information that they have to figure out who to go after with their scarce resources the data is proven reliable enough that it's even being used in prosecution are there unintended consequences here as we've discussed you want to avoid the legal or non trafficked economy and there are public policy questions too there are potentials for doing bad but at net most people seem to feel that this is doing a lot more good than it is raising questions of social manipulation or invasions of privacy and in fact the techniques that I've talked about here are increasingly being used widely at least in the United States and increasingly being explored for things other than trafficking other examples of crime that can be a tech detected using similar techniques to me I think this is another example of an improved solution to a familiar problem the problem of human trafficking has been around for a while this is an example of using the web and data analytics the very tool that was aiding them against them and clearly exploiting the arc of technology evolution with new more powerful data mining and machine learning algorithms I think the disruptive innovation here is machine learning on web-based data the next example is making smarter decisions about bail in the criminal justice system this is work by less kovitch and coworkers at Stanford the University of Chicago and Harvard bail as many of you know is an amount of money collected by the courts to be held while they release a defendant pending trial or pending appeal bail and the ability to release a defendant awaiting trial is something that happens in many countries around the world the goal of the bail money that's collected is to provide an incentive for the defendant to return for trial to hopefully not commit a crime in the meantime and in some countries although not in the US the money can also be used to compensate victims the challenges that judges face in making a bail decision on a defendant are substantial there's many factors they need to consider they only have a few moments to make each decision it's not the kind of thing where they can go back to their chambers and think about it for a while it's the kind of thing where they have 30 seconds or a minute to make a decision before they move on to the next case and there's no systematic way that the judges get feedback on how their decisions turn out if they release somebody and that person never returns for trial or commits a crime in the meantime they might hear about it while reading the newspaper over their coffee in the morning but that's about it that data has not typically been analyzed and presented to the judges in a way that helps them to learn from their past decisions so the challenge that these workers put in front of themselves is whether the data can be used to help judges make better decisions what's the experimental setup here it begins with training to collect the data and create a machine learning system and then test the model and compare it against the results of the judges what's the size of the data one particular state in the United States that they used in this experiment Kentucky provided more than 300,000 data points and across the u.s. they had more than 1 million data points in this data approximately three-fourths of all defendants were released on bail and about seventeen percent of them either failed to appear for their trial or committed some sort of violent or nonviolent crime in the meantime this slide shows the variables that were collected for each one of those data points there's a lot of things here age at first arrest the crime that they're accused of a variety of historical things like what's their history of criminal activity what's not on this list is the variables that at least in the US are not legal to collect on people and use for this sort of decision process like race and gender let's pause for a minute before we look at the results and think about what we're doing here we're using a machine model to compare against human judgment between humans and machines we have an information asymmetry the machines only have the information that we're providing that list of variables I showed you a moment ago the judges have all kinds of additional information they're looking at the defendant they see how he or she is dressed they see the defendants behavior they see whether the defendant appears sorry for what they did there may be different kinds of bias in the human decision-making this could be intentional bias on the part of the judge or it could be unintentional and subconscious there's an additional difficult problem related to false positives and false negatives unique to this particular problem the machine learning model can only look at the consequences of the defendants that are released on bail the judges decide who to release and then we can keep track of whether they commit a crime or not but we get no information at all on the defendants that are not released the ones that are locked up and aren't given bail I was a huge hole in the machine learning model development and the way the researchers partly dealt with that was to recognize that at least in the US cases are dafont assigned to judges randomly but some judges are systematically more lenient more likely to release the defendant and some judges are systematically stricter so by looking at the lenient judges it was possible to at least partially probe that population of defendants that are often not released by average judges so here we're plotting crime rate versus release rate so for the judges that's a single data point that's not a curve the judges released seventy-three percent of the defendants and seventeen percent of them either failed to appear or committed a crime prior to their trial date the hypothetical machine learning model did much better and showed that at that same seventy-three percent release rate the crime rate would be significantly less why might that be why should the model be able to do better than the judges the researchers came up with a number of possible reasons one is the judges non optimum use of the observable variables of those variables that were available to the machine learning algorithm perhaps the machine learning algorithm is doing a better job of dealing with variables like the history of this defendant committing crimes it might also be that the judges are non optimally using the variables that were unavailable to the machine learning algorithm like the defendants appearance and behavior it's also possible that the judges intuition and instinct and emotion is leading them to get to a higher crime rate than the machine learning model so what comes out of this clearly algorithms have the potential to support better human decisions they can identify different kinds of bias that might help us learn about ourselves in our own behavior these kinds of approaches could be applied elsewhere perhaps they'll be applied in the biomedicine area to help design the most useful diagnostic or therapeutic procedures for patients perhaps we could use these in educational settings to identify at-risk students and then offer appropriate interventions but I think for most of us these sorts of things are complicated ethically this notion that we're going to learn things based on statistics and based on past behavior has the potential for abuse and these are not currently in use anywhere outside of the research environment that I'm aware of so we have the potential for an improved solution too familiar problems they're exploiting the arc of technology innovation and in this case the disruptive innovation is machine learning based on massive data on human systems and I want to talk about efforts to use photovoltaic panels batteries and LEDs to replace kerosene lanterns in the developing world the problem as you know is that 1.5 billion people lack reliable access to electricity they use kerosene lanterns for lighting they have to pay small amounts of money and often walk a couple of kilometres to charge their smartphone that are increasingly a necessary element of the Commerce and the business that they're engaged in and they have to use biomass for cooking and heating kerosene lanterns are public health disaster they produce a flickering light that's difficult to read study and work by they produce indoor air pollution the leads to respiratory diseases and they cause burns and fires why do people use them well the kerosene lantern itself can be built with scavenged parts from a garbage dump and each day or each week people buy just as much kerosene as they can afford if it's a good week they'll buy a little more if it's a bad week they'll buy a little less but in fact integrated over an entire year some of the poorest people on this planet are spending more than a hundred dollars equivalent on kerosene ten percent or more of their entire annual income in addition they have no electricity for other things like irrigation pumps or sewing machines or computers of course better solutions exist photovoltaic panels plus batteries plus LED lights address many of the disadvantages of kerosene lanterns interestingly the payback period for this kind of solar Lantern as they're called is less than a year so this is actually cheaper than the integrated amount spent on kerosene over the course of a year the problem is this is all upfront capital and the poor people we're talking about don't have 40 or 60 or 80 dollars lying around to buy one of these well here in the developed world we know how to solve this problem it's the same solution that all of us use to make major purchases like cars or homes it's called finance and on Gaza design and many other companies are taking that approach to providing solar lanterns in the developing world it's often called pay-as-you-go solar the model is that the company buys the solar lantern hardware the PV panel battery and LED light system and then a payment scheme is put in place where the user can unlock only as many minutes of lighting or as many kilowatt hours of electricity as they're able to pay for that day or that week sometimes scratch off cards are used for payment sometimes smartphones through secure links and the idea is each day or each week you're paying less than you would have paid for the kerosene and after a period of several months you now own the equipment out right at which point it's either unlocked and you can now just use it or some people will choose to upgrade to a larger system so that they have more lights or the ability to power other things the impact that this has on people in the developing world can be huge in one example a woman who raised chickens to sell the eggs found that by putting lights in her chicken coop for a few additional hours of lighting per day egg production increased and she was able to make more money with her egg business the component parts of actually implementing a working system are fairly complicated not so much from the technology point of view but how to integrate it into a viable business model in the case of this one company on Gaza design there's a data transfer system that allows data to move between the end-user the solar Lantern and the central business operator keeping track of payments the central business operator has a variety of data panels to keep track of and then of course most important there needs to be a way for either the smartphone or the cloud to communicate with the solar lantern and unlock the right number of hours of lighting the reason this works is the cost of the solar Lantern is less than what would be spent on kerosene over a relatively period of small period of time so there's net economic value and as entrepreneurs know once you have net economic value you have the potential of creating a viable business model as I mentioned it's it's interesting that this is an example where a cutting-edge technology PV panels batteries LEDs is actually more easily economically advantageous in the developing world than in the developed world as we all know in the developed world solar panels and batteries and LEDs are still struggling to compete with the grid but in the developing world it competes more easily against kerosene this pay-as-you-go concept can probably be applied to other things not just more electricity but perhaps clean water systems at the village level or other things like that here we see taking advantage of the arc of technology evolution the communications channels the secure links the ability to remotely turn a solar lantern on and off but most important business model innovation figuring out a way to apply existing finance models to a different and old problem in the developing world so we've seen how disruptive innovation can provide an opportunity for entrepreneurship and for social impact we've seen how this can come from the conceptual combination of two different fields in some of our examples combining big data with modern circuits and systems particularly high performance low power processors but sometimes the synergies can be surprising in between two things more different than less obviously connected than big data and processors and we've seen how sometimes the technology involved is very sophisticated and other times not so sophisticated we've taken a look at a number of disruptive innovations in many cases from the world of analytics in many cases from the world of circuits and systems and sometimes from the world of finance we've seen how these classes of inquiry can help us to think about and conceptualize and categorize these examples of really impactful innovation and it might be that thinking about these classes of inquiry can help all of us to be more innovative and impactful in our own work or help our students and our colleagues and our co-workers to be more impactful or maybe even help entire institutions and organizations to become more innovative and impactful and we've seen a variety of kinds of social impact mostly good sometimes troublesome we've seen the path what may be the path to security and privacy in the Internet of Things we've seen the potential for an electric grid that's more environmentally benign and creates new businesses we've seen help for the social science problems of human trafficking and perhaps how analytics can be used in the criminal justice system to make smarter better informed decisions and we've seen a path to eliminating kerosene lanterns with the humanitarian public health and environmental benefits that go along with that as we think about applying all of this in the future I find it helpful to think about our own research on circuits and systems and how combining that with other fields perhaps big data perhaps other things like things from the life sciences might help us all to be more innovative and to have more large-scale social impact thank you all very much thanks st for the very interesting talk any questions from the audience yes can you give a microphone there yeah hi thank you very nice talk please help me understand just a couple of things though for example excuse me on for example the human trafficking you gave an example where the internet was helping in one sense the traffickers and hurting them perhaps in the other and there were perhaps some flavor of that and some of the other topics as well so I'm struggling a little bit to understand the just the the disruptive part of the innovation here and was any of this actually applied in such a way that you could measure and see the discontinuity that you are looking for you're asking about the discontinuity in the disruptive innovation or the discontinuity and the impact I guess the discontinuity the impact or performance that is that's the yeah so I think that what's going on with the internet is really the backdrop for this the Internet's been around for a while a lot of the bad guys have figured out how to use it I think that the disruptive innovation in those examples at least to my way of thinking is the new more powerful analytics the data mining the machine learning and so on you asked about actually measuring the beneficial impact in the case of the human trafficking I'm told that there are increasingly examples of being able to identify the traffickers bring them to trial and get convictions I don't have that data in front of me that would have been a great thing and my guess is that that data will become available in the next couple of years that will I'm showing here is still new enough this is something that's only been talked about publicly for a short time in fact you can imagine there was some discussion and some angst about how much to talk about so as to not give the bad guys any help by disclosing the techniques being used and so I think that the the good results appear to be coming together there and I think the authors of that work will be reporting on it and quantifying the impact shortly in the case of the criminal justice work the work on bail decisions it's not so far along and it's not so clear that there is going to be an appropriate use and a net benefit any other questions may have actually one question in the going in the same direction so some of the techniques especially that you have which are relying on data mining and so on they heavily depend on how much the available data exists so there's a very big debate and actually issues on privacy and all this information and even between countries have been looking at the problem for a while and then you see that in the US versus what you can do in Europe or even what is available in Asia is completely different so so there's a little bit of concerns about how much you can do and there is a way to create some kind of harmonization of the data how can you share data between different countries and so on so do you see any any capabilities to convert based on the opportunities that exist here is it something every matter we can forget about it would be very much per country or something like that yeah it's it's a huge problem isn't in a huge source of concern for all of us so much more data is being collected in so many ways and the data can invade our privacy and be misused as you pointed out there can be significant cultural differences and expectations of privacy a related problem that you didn't mention is proprietary data versus publicly owned data you know as as we enter a world where increasingly the data itself has value the same kind of value that we're used to associating with things but now it's the data that has value people will be more and more people and organizations and companies some of whom don't share our moral concerns will be more and more motivated to collect data and to keep the data to themselves as something that gives them competitive advantage I've been meeting a lot recently with car manufacturers automobile manufacturer mobiles are among the most connected things we own they have myriad sensors they communicate with us in our smartphones they communicate with the cloud thanks to those smartphones that we walk around with our car our automobile now has the capability of knowing what we're doing even when we're not in the car we parked the car we lock it we walk down the street to go eat or go shopping or go to work our car has the potential to track what we're doing with potentially huge consequences of invading our privacy also consequences of doing good things and the car manufacturers that depends so much on their reputation to sell automobiles to us care a whole lot about how they execute this and how we feel about it so I see lots of issues around data that threatens privacy around proprietary versus public data I don't see yet what the path is to resolving that I see concerned people like all of us raising the question and voicing our concerns I see some people who out of self-interest share our concerns like a car manufacturer that cares about their reputation but as the question indicated I'm also concerned about people and organizations who can operate an obscurity or who don't care about their reputation and are just trying to exploit things any other questions yes well there was another one up there sorry it would be Willie will go for you yeah I was wondering if you have anything more to say about homomorphic encryption so there are certain things one could do for example selectively use the homomorphic encryption for selected data not on all data so that would take off some of the burden on applying homomorphic computing because it's computer smell intensive and I was wondering if you have anything more to say it's just from my own learning perspective thank you so for everyone's benefit it was shown a few years ago that in theory this idea of doing computation on encrypted data is possible for all operations but the computational burden of doing it can make it completely impractical as I recall it can approach 10 orders of magnitude 10 to the 10th grader computational effort for some operations making it completely impractical in the last few years research has reduced that to tolerable penalties like a hundred x or a thousand x for certain operations that are identified in advance and so the current state of homomorphic encryption is for certain operations that are identified in advance it's possible to have this ability to calculate on encrypted data be be computationally practical so a lot of the research that's underway right now is to continue to push the cryptography to make this possible for more for wider range of computation I think the excitement will be when that work on cryptography happens also to occur with work on hardware to make sure we can optimize the hardware capability for these computations with models of what the data is actually going to be used for there are lots of opportunities that one can think of to take advantage of this kind of cloud-based aggregation of data where the data itself is encrypted for example you I want to know how your water or electricity usage compares to a population of your neighbors but you might not want the cloud to know where you sit relative to that distribution maybe you let use way more water or mail moving more electricity than anyone else in your community you may be sensitive to that you might not want that to be known but you might appreciate getting the information on how you compare to the distribution or if 100 cars in a row go around a curve on a road and all of them have their traction control systems activated it's probably not that people are driving too quick or drinking too much alcohol it's probably a defect in the road if your automobile could somehow provide valid information on what happened there without identifying you but just identifying that a car activated its traction control system at this location and this time then the data could be aggregated to wear a road repair crew could be sent out to repair the road so I think the potential to use this kind of technology to do good things while preserving privacy is there and I think that's one of the examples whether the progress in homomorphic encryption will be sufficient to have that be one of the things that unlocks the power of an Internet of Things while maintaining security and privacy remains to be seen this double more first questions hopefully he also had one here yeah hi thank you very much for this wonderful presentation so my question is on the machine learning part so for any machine learning and data mining we need to have several attributes for to do the Association rule and all the prediction so let's say if you want to merge the machine learning with the circuit is I so we need to first find those attributes so let's say the smart abayo decisions so we have several kinds of attributes already like those listed attributes what if there are some hidden attributes like whether they smoke whether like rabbits those kind of hidden attributes that is not available in our data set so is there any algorithm that can try to find those attributes that is the most effective way because if we want to use this algorithm out to a circuit design that is like a disruptive like blue ocean we don't know we have no idea which I supposed to pick first so if there are any suggestions how we can do that it's a great question and it gets to the heart of both the incredible capability of these sorts of analytical techniques and also their limitations in what's traditionally called machine learning the attributes are defined upfront and it's an exercise of finding correlations it might be correlations in a system with tens or hundreds of thousands of dimensions so it might be a particularly complex space but it's still the problem of seeking correlations and dependencies in what's typically called deep learning we typically don't start with the list of attributes we bring a more open minded or agnostic approach to the data and we explore the data in a way that allows the data in the analysis to tell us what the correlations and variables are and we can imagine systems that combine those two approaches there have been remarkable successes in both cases I think it's easy for all of us to see intuitively where the more deterministic and human based machine learning and data mining works but to take another example the capability of computer vision has exploded remarkably in the last several years this of course is to solve that socially most important problem of finding cats in pictures on the web and traditionally the approach to interpreting images the goal is a computer a picture and have the computer identify what's in the picture and increasingly in the future tell a story around what's going on in the picture much as a human would and in the early days of computer vision the approach was to look at certain shapes a cat has a round head and a oval or rectangular body but the breakthrough of computer vision in the last few years has come from throwing all of that away and simply having humans tag millions of images and then let the computers simply compare the images and the tags and extract the correlations and the result of that is increasingly accurate computer systems for looking at a photograph and determining what's in it and even increasingly getting some of the additional layers and nuance are the people happy or unhappy what might have happened a moment ago or what's likely to happen a moment in the future the questions suggestion that hardware could be custom tailored to these kinds of problems is of course an incredibly powerful one as you can imagine and things like different neural network architectures can often be realized in circuitry and increasingly as many of you know there are whole areas of research around using hardware as well as software to try to mimic the human brain and to try to see if there are other kinds of structures that can be more powerful perhaps using statistical rather than deterministic techniques perhaps using more parallel ISM perhaps to reduce the need for power consumption while still having the most robust and powerful computation so it's it's a a great question and an area of active research and your the final question otherwise i would like to guess thank you everybody for being here thank you very much Steve for the great presentation and the questions and a musician

Leave a Reply

Your email address will not be published. Required fields are marked *