Chris Van Pelt
Today’s guest speaker shares how the future of AI starts with better design systems. Join Chris Van Pelt and I, as we discuss how to enable AI in software development, why exiting Figure Eight set the stage for Weights & Biases, why the deep learning space needs new data science tooling, and why humans in the loop systems will be the future of tech startups. This is HumAIn.
Welcome to HumAIn. My name is DavidYakobovitch and I will be your host throughout this series. Together we will explore AI through fireside conversations with industry experts. From business executives and AI researchers to leaders who advanced AI for all.
HumAIn is the channel to release new AI products, to learn about industry trends and to bridge the gap between humans and machines in the fourth industrial revolution. If you like this episode, remember to subscribe and leave a review.
Welcome back to the HumAIn podcast. Today, I’m joined by my guest, Chris Van Pelt, who is the founder of Weights & Biases. He’s working on solving exactly that Weights & Biases with his new venture out of Silicon Valley. Chris, thanks for joining with us today.
Chris Van Pelt
You bet. Thanks for having me, David.
Now we’ve had the opportunity to chat off the record before about enabling AI technologies, how industries are changing. And you’ve now had a couple of ventures all in the space. Why is enabling AI so important for you?
Chris Van Pelt
It’s a paradigm shift in, in how we do software development and especially over the last five years or so we’ve seen this explosion in deep learning and I think we’re just getting started. it’s gonna really change the way that software is getting written.
I attend a lot of hackathons and I know you have your own hackathon, if you will, that’s more open-source. Can you tell us more about how you’re enabling AI by opening up to the community?
Chris Van Pelt
We recently started doing something we’re calling “benchmarks”, which folks can think of as like a mini Kaggle competition, oftentimes focused around social good or something that we believe is going to make positive change in the world.
So we just launched a benchmark called Drought Watch. It’s taking satellite imagery of various drought prone regions in the world. And the task it’s a call to folks in the machine learning community to create an algorithm to predict drought conditions before they happen so that we can take appropriate action and ensure that the impact on humanity is minimal.
I think that’s really important, especially when you talk about social good. I’ve had the opportunity to chat with a few other ventures who are trying to tackle that as well. But this is great for the scale because I think food shortage is definitely something we’re going to experience globally.
I’ve read some research earlier this year that each farmer supports more than 50 people with food, which is amazing to think about how we’ve gone away from subsistence farming into a system where we can support many. But, what do you think is going to happen if we can’t reverse these droughts?
Chris Van Pelt
I would wager your bad things. Bad things would happen. I’m actually from the Midwest, the Heartland of America. So, I’m very familiar with what it means to have a bad year in the crops. I think stakes are lower here in a developed nation, like the U S but the stakes are much higher than other places in the world. I would hope that technology helps us to minimize the impact on human life that can often happen with that kind of crazy state that our climate and weather patterns are in.
I like to be optimistic. So I know we, of course, have the weather where we are, but we do see good news out there where there’s research saying China’s planting more trees, GIS and spatial images are being used now to do exactly what you’ve described. Try to predict climate, try to improve environments. And if you were someone who’s a data scientist today and you wanted to get involved in these benchmarks, mini Kaggle competition and droughts. How could I get involved today?
Chris Van Pelt
I can get you a link. It’s just after wb.ai/wnb/strout watch, but we’ll make sure to post that in the description as well.
Awesome. So we’ll have them the show notes for everyone listening to the episode. I love social good initiatives because it shows how together, we can use data science to impact the world and one of the comments you shared just a few minutes ago is how we’re having a paradigm shift in software development. For those technical folks, listening to the episode now, we know in the last 10 years has been this evolution of infrastructure as a service with things like Ansible and Terraform and all these products from Hashi Corp and other providers, which has started evolving into data science as a service industry. And can you tell me more your thoughts on, do you think we’re moving towards data science as a service, or what does that look like today?
Chris Van Pelt
When we look at the space of kind of developer tools for machine learning, we see two different approaches in the marketplace. One is when it’s kind of an end to end plan form where you’re buying a software package that will kind of automate, or essentially give you data science as a service from kind of data, ingestion and transformation to training of models to actually deploying of those models. In Weights & Biases, we’ve actually taken a different approach.
We created a very kind of targeted point solution. That’s meant to be able to be plugged in to any existing open source solutions out there that you may be using for other parts of the pipeline and instead of trying to create an entire platform as a service we’re really focusing on the training and experimentation around creating models.
It’ll be interesting to see where the market goes in terms of teams feeling comfortable, kind of buying into something that’s end to end or wanting to pick and choose from a field quickly evolving pieces of software. I think it’s definitely very early days in this space.
It’s early days, but it’s exciting. Because if you’re a data scientist working in software like Python before the past couple of years, it’s been really difficult to manage experiments and see those results and Weights and Bias you’ve been working on it as your new venture, it sounds like your previous venture Figure Eight inspired you to bring about this new venture’s Weights and Bias.
And how do both services tie together?
For those who don’t know, Figure Eight is a fantastic company for helping you with all your data labeling solutions globally with computer vision, natural language processing, and has had a huge impact in the AI industry. And you’ve evolved from that venture to your new one. How do you think they tie together?
Chris Van Pelt
Everything starts with Figure Eight. In the majority of the enterprise, these models that are being created are supervised machine learning models, which means they need training data. And especially with deep learning models, they’re very hungry for more data.
So, Figure Eight can give companies that data in a highly scalable, efficient and accurate manner. So really before you could even use the Weights & Biases tool to build a model, you’d need a solution, like Figure Eight to actually label the data and then after you do build a model, you would continue to incorporate some form of labeling in your end to end pipeline.
So this is something, Figure Eight calls a “Human in the loop” essentially active learning where as your model is deployed in the wild, you hopefully kind of target examples that maybe the model didn’t do well on to actually go back through a labeling pipeline and get labels on to kind of further improve your model as you retrain it.
I think a Human in the loop is so important today, especially in data science. I’ve had the opportunity to work with a lot of students on capstone projects when we’re training them through bootcamps to get upskilled. And often when the student works on the project, the first time, they’re surprised that they can’t get necessarily 99% accuracy or get fantastic results in a data and especially for those working in images and texts, then I tell them: have you considered labeling some of your data or seeing what sources are out there?
And I think one of the big misnomers out there is that it’s just magic. The data is just magically good and when you see companies like Google saying: “Oh, we have a 99% accuracy that we can predict this disease”.
little does the consumer often know that a company like Google has spent tens of millions of dollars, hundreds of years of compute and processing power on working on data sets and labeling data to get it to a good enough steady state that now they can outperform a human and still have Humans in the loop. Is that a fair assessment perhaps?
Chris Van Pelt
Yeah. Anyone who’s logging into a website and seeing Google’s recapture is seeing an example of that. When you’re asked to look at the various images and select which ones have street lights in them, you’re labeling machine learning data that will be used to likely retrain models or somehow inform some of the data science that people are doing. So it’s a, it’s a very kind of core aspect of any real-world kind of mature machine learning application.
That’s so fun and cool because especially if you’re someone who’s not a developer by trade, you may think: “Oh, the recaptures” that you go on, google.com or other websites seem to be preventing bots and there are these security features to ensure that your data is safe and nothing bad is happening.
But the truth is next time you’re completing when this recaptures and you’re either sliding across the puzzle piece or you’re clicking the parts of the image to have a stop sign considering that you’re actually helping AI get better. So that’s enabling AI and that is in essence, crowdsourcing.
Chris Van Pelt
Yeah, totally. I remember in the early days with Luis von Ahn who was the original creator of recapture, for those that remember it was like you would see two snippets of text, essentially. And that was a very like pioneering and inspirational application of crowdsourcing and definitely led to some of what we created at CrowdFlower (Figure Eight).
That’s awesome. I love everything that’s again, happening in how we’re evolving in the AI industry. It’s so interesting that yesterday the data labeling is becoming even more important because it is one of the starting stages when you’re solving problems.
But then when you get deeper into the models that you’re deploying and solving, you have your new venture Weights & Biases, which is here to help better understand what’s going wrong, or what’s going right when you’re training.
And could you maybe give us a practical example of what’s one of the models that you’ve been seeing success with, or maybe a high-level use case on how the Weights & Biases tools have been helping improve experimentation?
Chris Van Pelt
Sure. As a software developer, who’s been writing software for.
Oh God, nearly 20 years now! I had really taken for granted how good the developer tooling was for writing kind of basic web software or UIs.
I’m talking about tools like GitHub, and GIT our various kinds of IDs. And as I started to build more deep learning models, I quickly saw that just the tooling in the space was pretty lacking. And it is as mentioned earlier, a different paradigm.
So, when we’re writing regular code, we can kind of change a couple of lines and that might change some logic, but then we get to tell us exactly what changed and we can kind of go back and see what happened when we change a couple of lines of our model and then retrain it, the actual, the weights and the biases that are generated through the training process.
Every one of those values is going to change.
So Virgin control really falls apart. And as we talked to two teams and individuals in this space, we were kind of asking what folks were doing. And everyone was doing a very ad hoc. They put in place a very ad hoc experimentation kind of tracking system.
So this usually consisted of like a spreadsheet where they could write down, maybe some notes, every time they retrain the model. And maybe if they were advanced, they would be writing a Jason file or something that the specified all of the input parameters are configuration options so that they can have some record of what was tried.
So that was really the genesis story of Weights & Biases: It was first trying to address this issue of kind of keeping track of what you had done and then hopefully better enabling teams to reproduce any results that had been obtained in the past.
The way we see teams use this today, we’re working with a number of folks in the autonomous vehicle space in Silicon Valley, it’s definitely a really hot sector and they’re building lots of models and they have large teams that are all trying to collaborate with each other. So we’ve seen it have a really positive impact on the team’s productivity and just ability to keep track of what’s happening.
That’s so smart because just like the use case you gave before Chris, GIT´s been around for a quick minute.
It’s been around since the nineties it’s become the leaving version control system. I know this year there was a report from GIT Lab that said that GIT Version Control, now over 95% developers use the GIT version control system they’re not using Mercurial or the other ones anymore. And so this begs the question, how are we going to track models and model development?
And that’s really exciting to see the impact you’ve been having with the different ventures on their projects. And of course, computer vision, especially with autonomous vehicles, is a fast growing industry on one of my other episodes for HumAIn I got to speak with a chief science officer at the fortune 500 company about level zero through level five.
And where we think computer vision is going and how far we are in the autonomous vehicle race. From what you’ve seen with your partner companies, or maybe just the industry as a whole, where do you think we are in that life cycle of getting to autonomous?
Chris Van Pelt
You’re gonna make me predict when we’re going to get to. Having worked with some of the leading teams in the industry, I have been blown away by how good an individual model performance currently is. I’m talking about the model that’s just putting bounding boxes around objects of interest. I remember being really blown away at how accurate the kind of the existing models that I’ve seen are and how they’re able to detect what the little bitty humans, like way off in the distance.
That being said, the full package involves likely an ensemble of models and all sorts of decision-making and of course being robust to different environments. And I’m not going to commit to a date. I think we’re at least a few years out before we see any kind of meaningful usage of the technology, but it’s exciting times for sure.
Yeah it’s super exciting and totally understand that and the date part, but I think we can see that example you just gave on the bounding boxes is so relevant.
One of the platforms I use a lot is Facebook with all my friends and posting images from events that we go to. And I remember earlier in the summer, Facebook had this issue, where images were no longer showing. And they were shown with the text and they would say something like: “image is two Humans with a dog” or image is “house and person on bicycle”.
And it was so fascinating to see behind the scenes, how they were actually classifying these images and perhaps they were using these bounding boxes and other techniques to do that. But it drills it home, how relevant it is today. Because when I go to these events with friends, we’ll say, “Oh, show me a picture of your dog”.
And then they’d go to Facebook and they would pull up the images and I searched for dogs. And sometimes you would see images that you’d be like: “where’s your dog? There’s no Boston terrier in that photo”, but then when you inspect the image very closely out in the corner, or very far in the background, there was a dog. So it’s really amazing to see the breakthroughs we’ve been having, in AI technology.
Chris Van Pelt
Do you have anything about captioning, especially like converting an image into a string of natural languages is fascinating. And it was really computer vision that really started the hype around deep learning a few years back. And it’s been really exciting to see the advances in natural language processing over the last couple of years. And then with a use case, like kind of image captioning, you kind of get to marry both worlds, which is cool for sure.
That’s super cool with image, but then also two texts, like you just mentioned a natural language processing. One of my favorite ventures that I’ve been following over the last few years is Grammarly.
Grammarly helps you with spelling and correcting. It’s like Microsoft Word, but for any software you’re using to get those corrections. And it used to be really simple. It was like turnitin.com or Microsoft Word for the corrections.
But just in the past couple of years, they’ve started doing NLP integrations where if you have a certain phrase and because that phrase has appeared so many times in others research papers, they can make smarter recommendations.
We’re starting to see that with NLP and the industry is evolving really fast. I know there was even research that came out from a joint effort with IBM and Harvard on this glitter project where you could generate articles based on previous data. I even have one of the students I’m working with, who’s working on a rhyming scheme to build their own song generator. It’s amazing. What’s happening with texts today? What are you seeing as some of the advances in texts as well?
Chris Van Pelt
I think one of the biggest examples of a blade kind of thing that made a big splash is opening eyes GPT-2 and their choice to not open up the weights, I think now a few folks have actually gotten the insane amount of compute resources needed to actually generate those weights and give it to the world.
But there’s a lot of fear, especially in the current time around what it means for something to be fake or fake news and these algorithms are shockingly good at generating something that sounds very reasonable. A lot of out of nothing, it’s like truly, I guess, fake. So, along with the transformer architecture, which is kind of a new deep learning architecture are definitely the big topics that we see in the natural language space.
And it sounds like even with what we’re seeing in natural language or computer vision. You just mentioned about the deep fakes and the whole industry we even saw there was one of the presidential candidates recently that pretended to be sick and had a call in from Skype. And then on that, there was actually a doctored video of the candidates, using someone else’s voice so it is quite interesting what’s happening today, especially because all the people in the audience thought it was really the candidate.
So it begs the question on authentication and fingerprinting and spoofing where that’s going. But I think Wes, the fear mentality, but more the optimism mentality that it’s going back to the humans in the loop systems that you mentioned earlier we are going to continue to need humans for these cognitively challenging tasks. To authenticate or to verify or to improve.
Whether you’re generating an article or running an experiment with an autonomous vehicle or checking if a candidate is really live streaming, you have to verify that and although there are some signals today that a human looking at something we can tell it is doctor that may not be possible in a few years to, I think you’re right it’s all about the Weights & Biases and to pivot into that topic: let’s go to the more fundamental level. So we have data and data that can be labeled and it can be used to be run into a system and ultimately generate a model that can be deployed. But data could have issues. And one of those issues is bias. And how would you define bias for our audience today?
Chris Van Pelt
Essentially any time there’s some underlying pattern in your data that it’s not getting after the core of what you’re trying to predict, but instead the systematic of something else in your data collection process. The simple example is: we’re building a model that predicts the likelihood that someone who has been incarcerated will have an episode where they’re brought back into the system.
We can take our existing data and create this model, but there’s likely systemic bias in our own prison system that is racially biased that would make these models then more likely to predict certain races to commit a crime again, or to predict that they would commit a crime again when in reality that’s not a factor that should play into the model’s prediction.
That makes sense. I watched this documentary a few months ago on a flight and they were talking about even in Brownsville, New York and parts of New Jersey exactly about that: more police monitor streets because of certain demographics. I mean, is that something that we should do or shouldn’t do?
And I think what you gave is a really good example of bias that it is almost a loop. It’s a self perpetuating habit that can repeat over time, but ultimately it shows up in all systems and whether these systems are closed or open AI systems. We see the bias showing up and how can we minimize it if I was a researcher today? What are some actionable steps that could do to either minimize bias or discover, or just be aware that it exists?
Chris Van Pelt
The best thing any data scientist can do is try to deeply understand their, their data set. So in the initial training data creation and curation process.
there’s a number of great tools out there that allow you to kind of slice and dice that data, and kind of pull all sorts of statistics over various axes.
And then I would say the same is true for once you’ve created a model to kind of measure how the outputs of that model are performing across say an evaluation data center, some set of data.
We at Weights & Biases think about bias a lot. And we hope that some of the tools that we’ve kind of embedded in our platform will enable data scientists to surface these kinds of issues. So just easier ways to essentially visualize inquiry into how the model performed is, is really one of the core feature sets of the waste and biases tool as well. So I wish there were like a silver bullet that you could just say “hit the no bias button”. But the reality is it’s a lot of work and it just takes a deep understanding of both the underlying data and the model itself.
Yeah, we should make a “no bias button” that’d be super cool. Instead of saying that was easy, no bias, no bias, no bias, but it’s a good reminder. Because when you think of design thinking, you have to constantly think, what are, is my checklist that I can ensure that I’m building the best robust system and I’m covering all end points, because if you don’t, it’s just like what you mentioned earlier, Chris, you have this issue where an incarcerated patient is being reacclimated to society, but then the bias of society is not helping that person reacclimate and could potentially increase the risk that something non desirable happens to that individual. So, design thinking for bias is a huge part there. And perhaps that’s something we’re going to continue to see to emerge.
I know this year, some of the big words coming out, these buzz phrases have been explainable, AI and understandable AI. These are making systems that we understand the bias, we understand how the data is creating a result. What do you think of those general terms? Should the industry be under explainable AI or what’s your thoughts there?
Chris Van Pelt
It’s a big topic today because deep learning especially is really difficult to explain. And we’re transitioning from a world where with boosted trees or a simple regression, it was much easier to explain what the model is doing but as we create these deep learning models with tens of thousands or millions of parameters, it becomes really difficult to explain why any given output was chosen or what their thought process was.
There’s a lot of folks working on better tools and ways to visualize into the network, to be able to answer some of those questions. But there’s a huge number of use cases that we’ll just not use deep learning because of this explainability issue and instead use a more traditional machine learning approach that they can easily walk back and see the various branches of decision-making that the algorithm made.
And some of those traditional techniques could be like a decision tree, as you mentioned, being able to see each decision, but the truth is as we’re becoming more data for society, which assumes becoming AI first reinforcement learning as a theme is emerging that is part of that deep learning and learning and growing over time. What have you seen about the change in reinforcement learning? Is that something that some of the clients are using with Weights & Biases?
Chris Van Pelt
One of our earliest clients is open AI and they’re real pioneers in the reinforcement learning space and have open-sourced a number of tools themselves that are kind of industry standard now are used by lots of folks doing RL.
Reinforcement learning is definitely more on the frontier of ML. But what’s been interesting is in some of our conversations with actual enterprises we’re seeing folks use our RL at least in an experimental context which has been really interesting to see and it’s just a hackathon that on Facebook, the PI torch hackathon and saw one of the groups, a really cool project using reinforcement learning to actually do kind of a classification task or something that maybe you could just kind of plug in to real-world use cases.
I think it’s still early days, but it’s been awesome to see kind of more exposure out of just the pure kind of researching-gameplay world and seeing our RL used in other scenarios that could make sense.
Yeah, one of the scenarios I’ve also seen is that a hackathon recently with RL is about when you’re doing certain hand gestures, that the system can learn over time, what those gestures are, so if it’s trying to do a translation of sign language, it can improve over time. I agree with a lot of these live learning methods.
In RL, it was very exciting and still early days, but definitely part of what we’re going to see as coining data science as a service continues to evolve. What else are you noticing as maybe since you work with a lot of researchers, a lot of developers, any other research trends that are emerging today that people should think about?
Chris Van Pelt
I guess the biggest trend that I’m seeing is kind of more excitement around the unsupervised machine learning use cases so being able to take data that hasn’t been labeled by any human and actually surface or unearth patterns simply by kind of looking at all the data, that to me is super exciting.
I think there’s the likelihood you still continue to have very hybrid approaches as this kind of compute continues to become less constrained and data sets become larger. We’re seeing these models actually kind of pull apart underlying patterns or categories in the data without ever being told what those categories are, which is really cool.
That’s so interesting there’s so many use cases that we can think of. Like if you just have images that have never been trained before with that technique whichever technique that’s going to be. If that’s something with open AI or with Weights & Biases is going to help you classify that data, which I think is fascinating.
And one of the things you mentioned is three themes I’ve been hearing consistently throughout our entire conversation today. Humans are still going to be in the loop, data sets are going to continue to become larger and compute is going to become less constrained. And I think the two parts that we’re still as society trying to solve is how do we do better work with data sets as they become larger.
And that ultimately seems to be around: Compute or this big Oh! notation and how to process data. What do you think is a way that we’re going to be able to better solve compute? Is it just let’s get thousands of GPU’s and TPS to do our processing at more marginal costs or other breakthroughs in algorithms and different types of hardware.I know a lot of the companies are coming out with AI chips now. What do you think might be potentially something that helps solve that problem?
Chris Van Pelt
I think it’s all about the custom hardware. So the reason Nvidia GPUs are winning is the accelerator for machine learning is because they have
the best library in the market to do the operations that the deep learning networks do autonomously when you think about like AMD and their GPU’s versus Nvidia there’s not a ton of difference in how they’re implementing or kind of the amount of compute that either one of them can do what the differences is Nvidia has a much better library to enable developers to actually leverage those operations and then we’re seeing Google with their tensor processing unit that is really well supported by TensorFlow.
They’re definitely gaining market share and there’s a huge number of startups trying to make hardware chips that can do all of this matrix math really quickly and highly parallelized. Those are going to be continued innovation and likely some kind of big step gains as the market matures. Another interesting thing is as researchers create these new model architectures, the hardware makers are always kind of playing catch up.
As the research continues to evolve and we kind of decide as a community, which approaches we want to kind of invest more in than then those approaches get further optimized and it’s easier for us to make models and hopefully it’s this big virtuous cycle.
And as a community one of the big updates we saw in the last few months was TensorFlow and Keras becoming very much tied together with our API. With the emergence of TensorFlow 2.0, I know there’s so many different softwares, including in the Python language and others to do machine learning today. But what have you seen as some of the things that you’ve liked in the new TensorFlow 2.0 release or other AI packages that you’ve been working with?
Chris Van Pelt
Tesla tools should be coming out in the next month or two officially which is exciting. I’ve been teaching classes on machine learning for the last few years and I have always been a huge fan of Keras and that’s the kind of abstraction that we use for all of our course materials so it’s great to see it embedded directly in the library.
I think the main selling point of TensorFlow too, is the eager execution. So making the execution model look a bit more like a PI torch and easing the ability for developers to kind of debug various operations in their compute graph it’s definitely exciting.
We’re continuing to see the market really gravitate towards PI torch especially for research and oftentimes kind of after a model has been really dialed in folks might reimplement it in TensorFlow to actually deploy it but the poachers folks are hard at work at making their deployment. Pipeline is as efficient as possible.
You’re hearing it here first. If you are a developer researcher in the Python space, if you’re a data scientist wanting to get into Python, PI torch and TensorFlow are the two big packages to do a lot of that research and to work with all the state of the art algorithms.
I think one of the consistent themes we’re seeing throughout this entire conversation today is technology is moving so fast and that’s why it’s even more important than ever before to have humans in the loop to ensure that we’re tracking processes that we’re ensuring that deployments are accurate or as best as possible and to ensure that whatever goals you’re aiming to solve, you’re minimizing bias, or at least recognizing where bias occurs in your system.
It’s going to be really exciting to see where the rest of 2019 takes us in research. And even beyond that, anything else from Weights & Bias that our fans can learn about today. In addition to your social hackathon on droughts that you’d like to share?
Chris Van Pelt
Yeah, we actually recorded a number of kinds of classes around deep learning and machine learning, using Keras and TensorFlow. So if you hit our website, wv.com and then you can click on the classes link. Or its tutorial, sorry, you can check them out and give them a try. We’d love for you to use our tool and hopefully we help you unearth any underlying bias or issue with your model and enable you to debug it quickly. And we’d love to hear from you. It’s free to use, so please check us out.
I’ve had the opportunity to play with the tool myself. I’ve looked at the code on Github. I’ve worked with the implementation in Python. It’s super easy, super fun. So hopefully someone can check it out and be excited to see where we continue to move in the human-AI industry. Chris. Thanks so much for being with us today.
Chris Van Pelt
You bet. Thanks for having me.
Hey Humans, thanks for listening to this episode of HumAIn. My name is David Yakobovitch, and if you like HumAIn, remember to click, subscribe on Apple podcasts, Spotify or Luminary. Thanks for tuning in. Join us for our next episode. New releases are every Tuesday.