Google Dataset Search Engine

Google promised if a dataset is “[…] hosted, whether it’s a publisher’s site, a digital library, or an author’s personal web page”, you will be able to find them in GDSE. We typed in “cancer” as the key term and we got 20+ results, from sources such as NHIH or Kaggle. Sounds like a service similar to Google Scholar and arxiv. And perhaps that’s why we joked: From now on at AIDL, we should rephrase the answer “go Google” to “go Google Dataset Search”.”

theverge.com

How Do You Start With Machine Learning?

This is a well-written Quora answer on how you should learn machine learning. The part we like is that it mentioned Bishop’s PRML. PRML has attained a status of “CLRS of machine learning”. That is one of those books people said you should read it, but no one really does. But we can assure you, if you even just browse through the book, your skill and understanding of machine learning will get better.

quora.com

Neural Style Transfer Tutorial by Adrian Rosebrock.

We post yet another tutorial by Adrian this time. Why are Adrian’s posts so appealing to us? Well, not only his tutorial covers the basic theory of a deep learning topic. He also include detail commentary on how his own code works. So you can always repeat what he did.

Now one more thing about his neural style transfer tutorial – it works real-time. As far as we know, you can probably optimize some existing tutorial code to achieve real-time information. But Adrian’s is probably the one tutorial we know which already works at real-time.

Anyway, this sounds good to us. And please check it out.

pyimagesearch.com

Dexterous Manipulation with Reinforcement Learning: Efficient, General, and Low-Cost

Here is a post from Berkeley BAIR, on how one can use low-cost hardware and reinforcement learning to learn complex manipulation for robotic hands.

berkeley.edu

Cython for high and low

We found this half-a-year-old gem from Stephen Merity’s blog. In the surface, it doesn’t look like it has anything to do with AI/DL, but then deep learning just happens to be one application you always want to write in hybrid python/C to optimize performance.

So check out Merity’s message, you will learn a lot on how it is done in practice.

smerity.com

Google Dopamine

RL, unlike deep learning, has fewer standard frameworks to initiate beginners. We have the openAI gym in the past, but the valuable site is now unmaintained. And there are a lot of small hacks you need to do to get some sample code working.

So Google’s Dopamine is going to change the landscape, it is an experimental framework which you can easily repeat certain method.

Currently, the framework’s focus is mainly Atari 2600 games, and it already implements several state-of-the-art algorithms.

googleblog.com

Tutorial on Genetic Algorithm

Here is a nice tutorial on genetic algorithm. In our time of deep learning, it’s refreshing to see discussion on GA, which is actually quite competitive when compare to reinforcement learning. Kudo to the author Rishabh Anand!

medium.com

Which GPU(s) to Get for Deep Learning – by Tim Dettmers

Tim Dettmer’s article on choice of GPUs for deep learning is a must-read for us AIDLers. This time he update his popular article “Which GPU(s) to Get for Deep Learning” to include the latest RTX 2080. We like his conclusion about RTX 2080 than 2080 Ti: he believes that 2080 is more cost-effective.

timdettmers.com

Layman Description of Technical Terms in Deep Learning.

Here is a nice layman description of deep learning terminologies.

facebook.com

Understanding an Algorithm

Many of you asked AIDL whether you can understand deep learning after taking several beginner courses. As ML practitioners, our answer is to make sure you can implement classic algorithms in training or inference yourself, compare your work with SOTA, and understand the strengths and weaknesses of various open source package down to the code level.

We think François Chollet explains better:

“A popular quote goes “if you can’t explain it in simple terms, you don’t understand it well enough” (often incorrectly attributed to Einstein or Feynman).

I think a more accurate take is: “if you can’t explain it in arbitrarily precise terms, you don’t understand it well enough””

“In particular, if you understand something clearly, you should be able to describe it in precise algorithmic terms to a computer: you should be able to implement it from scratch (as a simulation, as a framework, etc).””

twitter.com

OpenAI Five Lost to Human Experts in Dota 2

Few weeks ago, we learn that OpenAI created OpenAI Five to compete in one of the most competitive eSport game: Dota 2. We have covered previous Open AI’s Bot for DotA at length in Issue 25 Fact-checking section, we noted that it would be much tougher to move from a 1v1 mode to 5v5 mode because that require learning, complex strategy. But then OpenAI surprised us by showing that they can beat semi-pros in 5v5! So we did feel optimistic for the prospect for this week game at The International.

The difference between “The International” Match and the previous report, (also shared in this issue) is that the bot was dealing with semi-pro rather than genuine professional players. Unfortunately OpenAI failed. But then when you look at data and the phenomena they saw, we might be just hitting a scalability problem. Perhaps training for few more months would make the bot strong enough. See our analysis in the next item.

openai.com

OpenAI Five

We included an an older link from OpenAI so that readers can have a sense of previous OpenAI bot publicized in Jun this year. One major difference, as we mentioned previously is that this time OpenAI bot is dealing with pro players who collect prize money for living. More importantly is that three of the players were playing together competitively.

So what can we say about the strength of current OpenAI Five? First of all, when the players are at solo MMR around 5500, Five doesn’t seem to have problem to beat the team formed by the players in that level. Whereas the pros Five deal with have MMR rating 7000+ (e.g. xiao8’s rating is 7333). So just like machine’s long history of defeating humans in chess, we might be just looking at a scalability problem.

You may ask: is it a scalability problem which we can solve in reasonable amount of time? For example, it seems that machines MMR rating doesn’t just improve linearly over time. For example, we know that from March to June, a three month period, we see the bot trained from scratch can now beat a semi-pro at 5500, but then another three month period of time from June to now, machines don’t quite make up the gap of 1800 points.

Our discussion so far is based on solo MMR, if you look at the players records, team MMR could be a better metric to decide relative strength of a team. Would it be possible to estimate that number? So far, we don’t see such number published yet.

The true picture is also obscured by the fact that there are change of rules of courier. As OpenAI point out, it forces the machine to use a more aggressive strategy, which humans can exploit.

If we were OpenAI researchers, we would probably spend time to analyze the two lost games, and ask how can we learn even faster than now? Given what we see so far, we still place our bet on the machine, and it sounds like we are almost there.

openai.com

Simple Object Tracking With OpenCV by Adrian Rosebrock

Here is another nice tutorial by Adrian Rosebrock. This time on object tracking. Adrian is detail-oriented and knowledgeable in computer vision. He starts from the basic principle of centroid tracking, then presents detail python code for implementation.

pyimagesearch.com

Drive.ai’s Simulation System

This week Drive.ai shares an incredibly cool link on how their on-board passenger display as well as simulation system works. One stunning number they share for any 1 hour of real driving data, it takes 800 hours to annotate it. That might explain the wide-spread use of simulation in SDC training.

medium.com

Connectomics, Deep Learning and Google’s work

This piece came out 2 weeks ago, but it is about a topic we care a lot – connectomics.

You ask: What is connectomes and its study, connectomics? Well. The brain can be thought the connection of neurons. Many believe that if we can trace and map out how these connections really look like, then we can truly understand the mysteries of the brain, such as how we have consciousness (or perceived notion of it), how we can higher level thinking even a lot of us don’t intently train our own deep neural networks in our brain. ( 🙂 ) This thinking follows from very early results from say Hodgkin-Huxley model which describes the action-potential initiation and propagation within the neurons, in some sense, the belief that a physical (electrical model) could explain everything within the brain.

Anyhow, this is why connectomics, the study of these electrical wirings in the brain,is a big deal. But it is such difficult problem! Why? Just think about 1) how you acquire a brain you can study, 2) then get into the neural levels of structure of the brain, 3) then figure out how neurons are actually connected. There are many stories for the first two parts, but Google’s work focus on 3. So for the sake of our discussion, we already get a slice of a brain. What we need to do is just to decide which part of is neurons from an image. This brings us to a familiar problem in deep learning – image segmentation.

If you delve into the paper version of the work, you might notice that the idea of using Convnet in connectomic analysis is actually not new. So what did Google do? What the researcher found is this – usually when you apply Convnet the image segmentation in connectomics. First you use a Convnet to decide if a certain voxel (i.e. pixel) in a 3-d image is a boundary, then you use a separate algorithm to cluster all non-boundary voxel to segments. So they figure, could you make an all-in-one network for these two steps?

And they did, the result is what they called flood-filling network (FFN). Their results show that this new method works better than the past methods. Now one interesting thing though – they also found this all-in-one method could much longer run time than previous two stages method. According to the authors, that is because,

multiple and partially overlapping inference computations are required to segment both
a single object as well as implement the sequential nature of multiple object segmentation.

Interesting to us and our readers. May be this open for some optimizations? Or did Googlers already do some optimization but want to publish a paper later on. We will see.

All-in-all, it’s an interesting paper, so do check it out.

googleblog.com

Google’s Cirq

We don’t pretend we understand quantum computing and quantum machine learning. But if you want to learn more about NISQ check out this Wikipedia link? And this editorial by Nathan Shammah’s will give you a good summary on how open source software interact with quantum computing.

googleblog.com

Google at ICML 2018

Here’s a list of papers by Google at ICML 2018.

googleblog.com

Accelerated Training and Inference with Google Object Detection API

The popular Google object detection API is now getting an extension. Check out the new architectures, accelerated training with the new release.

googleblog.com

AI Predictions in World Cup 2018

Here is an interesting article on several fails of World Cup 2018 prediction. As you know by now, France won the championship this time. And it seems like most ML models predicted otherwise.

Kaveh Bakhtiyari, a PhD candidate, wrote an article on analyzing why there are models are making mistakes in prediction. He suggested all three models he analyzed couldn’t quite account for external factors such as unfair referees or political situation of a country.

medium.com

AI uses in World Cup 2018

Here’s a well-written piece on how AI is actually being used in World Cup 2018. As you can guess, they are not all about prediction. In fact, many AI applications are about automated reporting.

medium.com

DeepMind papers at ICML 2018

Several interesting papers in our mind:

• Machine Theory of Mind
• Parallel WaveNet: Fast High-Fidelity Speech Synthesis
• Efficient Neural Audio Synthesis
• Measuring abstract reasoning in neural networks – see the next item.

deepmind.com

DeepMind’s Measuring Abstract Reasoning

This recent work from DeepMind probably catch many of our eyes. What are DeepMind doing? And how are they measuring reasoning from AI?

First of all, what is the abstract reasoning the blog is talking about. In fact, we are really talking about a particular type of intelligence test called “Raven’s Progressive Matrices” (RPMs). See the Wikipedia Page for some examples. Basically you are given panels of figures each have geometric shapes or patterns and your goal is to reason what should the next panel look like. It is mostly a visual reasoning test for humans. They were first presented with sequences of shapes, colors and patterns in several panels and then asked to derive the las.

With this setup, it’s not hard to imagine that you can come up with a task which machines can be used for guessing answers for RPMs. As you might read from the blog, the best architecture is what the authors called wild relationship network (WRAN). What is WRAN? Actually it’s rather simple. Say you are given K panels. You just run your favorite network on the task, then the answer would be summed and feed into another MLP.

At this point you might ask, “Hmm. So how much of this really measure abstract reasoning that I am thinking about?” The truth is, not much, even the authors admit that visual reasoning is only part of general-purpose abstract reasoning. Still it’s interesting to see if machines can be used to solve a popular IQ test of humans.

If you view it this way, experimental results of the paper are quite fascinating, e.g. the DeepMind researchers have tried to add distractor into panels to see if performance drop. For more details, you might want to take a look of the paper version of the work. You may also find the original data set at here

deepmind.com

An interview with François Chollet by Adrian Rosebrock

François Chollet is the lead developer of Keras, a popular package that interfaces with Tensorflow, Theano and Pytorch. A little about the interviewer too: Adrian Rosebrock has been making a name for his many hand-ons tutorial of using OpenCV, Nvidia DIGITS for deep learning applications. His questions are sharp and touch on many issues in AI today.

pyimagesearch.com

Capture The Flag

After AlphaGo, we are all looking for the next breakthrough in AI. So this piece on how DeepMind cracks Quake III Capture the Flag is interesting.

To understand this piece, you should understand that the agent is learning only from pixels and scores, they weren’t told about the rules. This is quite unlike OpenAI’s work on DotAI which quite a few of assumptions were made.

More interestingly, after training with human players and self-play, the agent shows cooperative behavior. That makes sense: if you put an RL-agent in an environment which cooperation is rewarded, then obviously the agent will cooperate more.

As Jack Clark mentioned in his Import AI newsletter, this result doesn’t seem to get as much press as other DeepMind’s result. It’s a pity. Large scale RL-training can now competitive and cooperative players, we think it’s a big deal.

You can also find the paper version at here.

deepmind.com

Ways to Think about Machine Learning

This is written by Ben Evans who works at a16z, trending around two weeks ago. Evans suggest several helpful ways to think of ML. For example, he believes thinking of ML as AI is really unhelpful. It’s much more useful to think of ML as automation. In a way, that follows the view from industrial revolution. i.e. machines are slowly replacing humans. Evans works in a16, so his view certainly comes from practice.

Another thought, which is intertwined with his view that “ML as automation”, is the idea that ML is a foundational technology similar to relational database in 80s. This is an interesting view. In fact, say if you just think about technology such as image classification, it is really the foundation of other relevant technologies such detection, segmentation. And of course, most medical image classification is still, in a nutshell, image classification based on Convnet. And who knows, what will people dream up in the next 10 years?

We think Evans’ article is a useful view to look at AI, his view is more helpful than say thinking AI is a super-intelligence, or AI will steal our jobs. Those are wishful thinkings mixed up with fears and silly imaginations.

ben-evans.com

FB DensePose

Facebook is open sourcing DensePose, a database which map 2D RGB image to 3D surface models, which you can imagine, is a useful resource if you want to detect human action or posture from image.

fb.com

Predicting World Cup Using ML

If you are US-based, probably you don’t care about soccer as much as the rest of the world do. In some parts of the planet though, soccer is like a religion. So it’s no surprise people are making serious effort on predicting the winner of FIFA 2018.

The paper which is making rounds is “Prediction of the FIFA World Cup 2018 – A random forest approach with an emphasis on estimated team ability parameters” by Groll et al, which use random forest for modeling. Looking deeper into the paper, the author used 16 factors including FIFA ranking and economic metrics of the country as the input variables.

Who is most likely to win? According to the authors: Spain which has 17.8% chance to win. What about German and Brazil? 17.1% and 12.8% respectively. Well well, we will see if these predictions are correct, and whether Spain would be in top-3.

technologyreview.com

Kaggle Data Note: Predict the World Cup 2018 Winner

From Kaggle, isn’t this the most important dataset now?

kaggle.com

Improving Language Understanding through Unsupervised Learning.

This work, by OpenAI, addresses the low resource problem in NLP, specifically in text classification. The authors asked if we get better performance by fine-tuning a pre-trained unsupervised model. And they found, in several tasks, the fine-tuned models give better results than SOTA. This works is based on previous OpenAI’s work on unsupervised sentiment neuron.

OpenAI’s approach is not without issues: e.g. while pre-trained model is effective, we need to assume we have a pre-trained model in the first place. But then training one would require large amount of computation. The authors also found that there are still issues of generalizability.

openai.com

The AI Bubbles and AI Researchers

When there is such irrational exuberance in AI, how should you see yourself as a researcher or perhaps even a developer?

Generally we believe that the so called “AI hype” itself is more fluffy talk. Though recently there are indeed more indications that companies are irrationally exuberant on the prospect of AI. e.g. There are so many self-styled experts. And you see AI influencers are treated as celebrities. I think these are unhealthy phenomena as well.

Many articles only try to pour cold water on AI, and some of their analyses are just plain silly. The authors of the OP though as researchers, are just setting themselves in the right footing, hoping that they can continue to research or do whatever they like to do. So I don’t see anything wrong.

hunch.net

How to build a custom face recognition data set?

Yet another great article by Adrian Rosebrock. This time on how to build a custom face recognition data set.

pyimagesearch.com

AI at Google : The 7 Principles

“AI at Google, the principe” was published 2 days ago. It reads like a sudden change of mind. Given the mantra of Google : don’t be evil. (It was dropped lately.), you would hope that Google published these principles before working on a defense project. Yet it only comes two days ago when Google already establish a sales department for selling AI research to agencies. And of course, pulling out from project Maven doesn’t mean that Google would stop acting like a defense contractor in the future.

What we know: the conflict between Google the web search company and Google the defense contractor will continue. So stay tune?

blog.google

Keras 2.2.0 is here

Keras 2.2.0 was just released. Perhaps the most important change is allow user to write a new model through Model subclassing.

github.com

Real Time t-SNE

Google presents yet another interesting results which can speed up t-SNE, the very versatile visualization algorithm. So what did Google researchers do?

In essence, there are two parts. First through a sampling method, they figure out a way to make the nlog(n) calculation in t-SNE to be linear. Then they also figure out how to draw the point faster by using WebGL texture function. The two combined was why the approach is much faster and can be real-time.

The source code can be found at here. You may also read the paper at here.

googleblog.com

A Good Advice of Debugging Python Script

Denny Britz on Twitter wrote about the Zen of Debugging Python Programs, while by itself is not just about AI. I think it’s very suitable for our group.

“How to debug Python, for beginners:

1. Carefully read the exception message. Repeat it to yourself 3 times, slowly.
2. Step away from the computer. Ideally, lie down somewhere.
3. Close your eyes, take 10 deep breaths.
4. You will have found your mistake.”

twitter.com

Using AI to turn People into Anime Art

Based on the paper published last year, “Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks” ,from Fudan University.

kotaku.com