HNHacker News
TopNewBestAskShowJobs

yonkshi

244 karma · joined August 12, 2016

submissionscomments
yonkshi··on Liquid water 'lake' revealed on Mars
Would you please provide some sources on all your numbers? those are some astronomical numbers and are essential in validating your point.

Edit: I did some of my own research, UHC take 8~11% GDP in other countries. That would put us around 1.8 trillion (2016 GDP). Much less than $3.5 trillion you claimed.

Plus, US is already spending 16% of GDP (~3 trillion) on health care, despite not having universal healthcare. 11% GDP for UHC does not seem like a bad idea to me

source: https://data.oecd.org/healthres/health-spending.htm

yonkshi··on Apple’s Shortcuts will flip the switch on Siri’s potential
The thought of Siri intentionally telling white lies is pretty funny and a bit creepy

- Hey Siri, am I fat?

- No Bruce, you are just big boned

- Thanks Siri.

yonkshi··on Apple Engineers Its Own Downfall with the Macbook Pro Keyboard
How’s the Linux battery management for laptops? This and driver compatibility are usually my biggest concern. I remember 2/3 windows laptops I used were having trouble sleeping/waking up when lids were shut
yonkshi··on Reinforcement Learning from scratch
Ah yeah I agree it's not a terribly difficult concept to learn, but perhaps we have a different definition of graduate level math. Things like information theory are, at best, skimmed over in upper div math classes during undergrad. Having a solid understanding things like KL divergence, information entropy is just an indicator of one's overall math level, and if you are already there, you can consider yourself at graduate level. Though I guess math used in ML is probably kids play comparing to a graduate level mathematician or physicist.
yonkshi··on ZTE of China to Pay $1B Fine in Deal to End U.S. Sanctions
That's assuming if China doesn't do something equally punishing to a US company. China can serious cripple US's race to 5G if they halted Qualcomm ops in China
yonkshi··on Reinforcement Learning from scratch
I agree to an extend. The current state of RL algorithms are more fragmented and have limited application comparing to something like a CNN.

So unless one only wants to learn how to play Atari, they won't be able to "fix" an algorithm if it breaks down in an untested environment. E.g. a non deterministic sparse reward game.

If one day there is a generalized algorithm that can solve large set of RL problems, then I think a high level intuition is probably enough to use RL, but for now, I'd say RL is definitely not for the faint of heart.

yonkshi··on Reinforcement Learning from scratch
For very limited applications you can probably plug and play something like DDQN or A3C, but since we don't have a good algorithm that applies to many RL problems, any innovation will take quite a bit of theoretical background.

A while ago we tried to use DDQN to self-play a multi-agent pacman environment that we built, it was able to converge but the convergence was very suboptimal. We had to modify the sampling process in order to get slightly better convergence.

yonkshi··on Reinforcement Learning from scratch
Yes, bellman equation fundamental idea for all RL algorithms (i.e. updating table of values), but RL is also much more fragile than supervised learning methods, thus to ensure stability, modern algorithms have complex mathematical tools to fix that. I wouldn't say it's the math notations that's scary, rather the concept behind modern algorithms require higher mathematical concepts. For example, Max Entropy algorithm for inverse RL, it's essential to know the concepts of Shannon information entropy to understand why it works.
yonkshi··on Reinforcement Learning from scratch
I think "learn" is a bit misleading here but I do have to say it's a nice and intuitive overview of RL. RL is quite hard and math heavy, I don't know if one can take a short cut in learning RL without solid graduate level math foundation.
yonkshi··on America's Startup Scene Is Looking Anemic
> The S. European economies suffer from a lack of ability to compete on this basis: the ability for people to work together in large teams to develop sophisticated products at scale. Vast numbers of people don't even recognize the kind of professionalism and specialization required to work in these organizations - and I means even simple things like product management, product marketing, a bunch of basic financial ops roles etc...

Would you mind providing a source on this point? Intuitively I thought of it as a cultural issue regarding work ethics and vacation time. This is a bit surprising to me.

yonkshi··on How to Install Tensorflow GPU with CUDA 9.2 for Python on Ubuntu
To add on top of the tutorial, it's recommended to compile with AVX, SSE and FMA instructions enabled if you are using a modern Intel chipset. It has a pretty big boost for calculations that needs to be done on the CPU.

The pip version of TF does not come with AVX and FMA for some reason, so this is one of perks from compiling from source

yonkshi··on Tesla Model 3 Gets CR Recommendation After Braking Update
Seems like an update on the ABS, so less to do with strength, more to do with how ABS are performed. My guess is this probably won't have much effect our everyday driving, but when in emergency, the updated ABS will perform better.
yonkshi··on Tesla Model 3 Gets CR Recommendation After Braking Update
From the article it seems like it was an update tweaking the ABS, which is really only triggered during an emergency braking.
yonkshi··on Seven Pillars of Causal Reasoning with Reflections on Machine Learning [pdf]
Pearl's paper is a solid generalization on logical reasoning and intelligence, but we have to be wary that even us, humans, aren't perfect at generalization and regularization. Without going too meta, we have to contemplate whether logical reasoning causes human level intelligence or it's a mere effect of a complex statistical phenomenon. It may be the case that a human-level AI system will contain all seven pillars of causal reason, but it does not mean that these reasoning were specifically programmed into the AI.

A similar example of this human over-regularization can be found speech technology, linguists have generalized 50~60 phonemes in English speech, though in real life speech, there are magnitudes more phonemes in play, if we force our speech recognition systems to classify speech into 60 phonemes, we miss out on important details. This is why end to end systems such as WaveNet is so much better than older systems, because when we ditch our over-regularized concept, it can learn more freely.

yonkshi··on Seven Pillars of Causal Reasoning with Reflections on Machine Learning [pdf]
I agree. RL is most definitely a form of causal reasoning in my book. It involves associations(value iteration), intervention(action/reward) and counterfactual (policy iteration) that Pearl suggested. I think Pearl did not mention specific technology as it's more of a meta discussion.
yonkshi··on How to write efficient matrix multiplication
We tested it against Tensorflow and scipy sparse matrices, we beat those as well, mostly because those general purpose sparse matrices do not optimize for binary matrix (eliminates the need for multiplication)
yonkshi··on How to write efficient matrix multiplication
I can attest this too, I was recently working with some massive and sparse binary matrix, with only 0.1% elements have values, it was magnitudes faster to index and sum the elements than to do a dot product. Not even a Tesla GPU was close to the performance of a custom matrix operation we wrote.
yonkshi··on Did Google Fake Its Big A.I. Demo?
Google AI hadn't published any major papers on ASR or TTS since WaveNet, even the new WaveNet google demo'd a few months ago weren't even close to this demo's voices. The use of "Uh Huh" "Hmmm" are called back channeling, it's moon shot away given the current state of technology. It requires very low latency and precise timing so it doesn't cut the speaker off.

I believe if this demo were real, it was tested and cherry picked from a very particular environment

yonkshi··on U.S. companies banned from selling components to ZTE
As someone who's not very familiar this area of US government, was this more of an autonomous prosecution kind of like "you broke the law you pay"? Or was this more motivated by the recent trade skirmish with China, being used as a pawn so to speak?
yonkshi··on U.S. companies banned from selling components to ZTE
They could use proxy companies, but ZTE risk getting even further punishment from the US Gov if they were caught again, possibly exiting US entirely.

Doesn't matter how many proxies they use, it takes one tear down to see if they used any US based components.

Although, I wonder, had they used a proxy company when they were doing business with N.Korea and Iran, it probably would've been much harder to prosecute in the US. Perhaps their upper management didn't care enough at the time.

yonkshi··on Planets evenly spaced on log scale
Indulge my wild guess (I'm not a physicist):

If the proto solar system's matter distribution was a Gaussian-like distribution, it would make sense that matter density droped in a logarithmic scale, and thus planets formed droped logarithmically w.r.t radius.

More intuitively.. further away from the sun has less probability for rocks/particles to interact with each other.

yonkshi··on OpenSimpleLidar
In the US 5mW is the legal limit for laser pointers. I think that can be used as a guideline, 3mW IR is much less energetic than a 5mW green/blue laser, perhaps it's not powerful enough to cause huge concern? But I agree no amount of laser is safe for the eye. Police LIDAR guns are 905nm at 50mW, that can technically damage you too.

https://en.wikipedia.org/wiki/LIDAR_traffic_enforcement#How_...

yonkshi··on Spotify opens on NYSE, valuing company at almost $30B
And in Sweden, home of Spotify, the other big news is that NYSE raised Switzerland's flag.

https://www.thelocal.se/20180403/spotify-swiss-or-swedish-wh...

yonkshi··on Likelihood of discontinuous progress around the development of AGI
I agree a true general AI will probably be fairly different on computer than human. Although I want to mention that humans can learn from one example is mostly because we already have a large a priori from our life experience. We spend years learning how to talk, communicate, write and read. Through which we have built very structured symbolic logic system, which is _learnt_.

A example of this would be mathematics. If a person is never taught mathematics, she/he are only limited to basic math operations. It would take the person years of learning and practice in order to comprehend a mathematical literature. Once we have a symbolic logic network built for a certain aspect of our life, we can rapidly retrieve information based on previous logical patterns, thus allowing us to learn from one example.

yonkshi··on Likelihood of discontinuous progress around the development of AGI
Hinton's criticism is very valid, but it's not quite about AlphaGo and its branch of ML. his criticism revolves around supervised learning and back prop that cannot be used to achieve the so called AGI. Because ANN is nothing like our brain's real NN. When Hinton gave the speech back in 2014, NN had a huge explosion of hype, and NN was mostly for supervised learning, which is really only good at classification and regression problems, it cannot make decisions outside of its training.

The famous DeepMind DQN paper (the core of AlphaGo) were published after Hinton's talk. the DQN paper practically opened a new chapter in reinforcement learning field. I am not sure if you are familiar reinforcement learning. RL is learning by trial and error, model-less and largely non-bayesian, similar to how humans learn. Up until AlphaGo, RL field was stuck in a limbo because it was having a very hard time learning non-linear problems (which is the majority of problems in nature)

When I say generalization, I meant generalization of knowledge. Generalization is the second piece of the puzzle because, even as humans, we learn from experienc. After enough examples we began to generalize. Up until DQN came out, we couldn't even effectively learn. It's the equivalent of a human baby with severe memory problem. With deepmind's DQN, we can achieve much more stable learning on non-linear systems, and we can begin to add components such as generalization (such as transfer learning), intuitions (such as intuitive physics), symbolic network.

I am not too sure what you meant by weak AI and general AI. For me, an AI which can learn similar to how humans learn, able to use apply generalized knowledge when facing a brand new problem, independently think and make decisions without human assistance, that's general enough.

Yes, much work needs to be done, but I don't believe this is the wrong direction we are going. Though I'd be glad to be proven wrong and I am very fascinated by this debate, if you would like, we could continue discussing it over email/chat?

yonkshi··on Likelihood of discontinuous progress around the development of AGI
I don't agree with this sentiment, although I agree that AI is not nearly at the level of the hype that pop culture makes it out to be. AlphaZero is still a significant contribution to 'AGI' that shouldn't be buried.

It's true that AlphaZero's knowledge is unable to be generalized for other systems, but its biggest contribution is a _stable_ RL system that can solve problems that no other systems can. This is the first piece of the puzzle of more general AI. Generalization, I would consider as the second piece of the puzzle to more general AI. Generalization may be achieved with potential research in transfer learning, model based RL, symbolic network. But without a stable RL algorithm such as DQN as foundation, generalization has nothing to stand on.

yonkshi··on Likelihood of discontinuous progress around the development of AGI
I think you are only thinking of supervised learning's capabilities, which I assume is the field of your AGI research? I'm working with reinforcement learning and DRL research, and RL was born to address the short coming of supervised learning. DeepMind is arguably the forerunner in RL right now, and AlphaZero is the crystalline of their RL research.

Yes, AlphaGo uses NN and other SL techniques, but the core is very much DQN based RL. No amount of SL can effectively play go and invent new moves. RL can already solve a large number of real world problems with a rather simple algorithm, from self driving cars to video games to NLP. RL can tackle all those problems with pretty much the same core algorithm. The question lies less in IF RL can solve more general AI problems, but rather HOW to solve it. From a high level view, we are having a lot of trouble with its convergence properties mathematically and its extreme sample inefficiency. This is the reason why Boston Dynamics doesn't use much RL, Waymo doesn't use much RL, simply because they can do much better with current techniques without going RL.

AlphaGo is still a major step forward regardless, because it's one of the biggest leap in RL we've taken in the recent years. It suddenly lets RL stably converge on solutions more than we could ever before. AlphaGo's contribution is more than just that it built a specialized Go bot, but rather a much more stable RL algorithm that lets us approximate non linear functions (majority of the real world applications are non linear). If I were to put my money, AI could very well be entering a new era with AlphaGo and their DQN.

yonkshi··on Chinese authorities detain Apple employees suspected of selling customer data
I suppose time will tell :)
yonkshi··on Chinese authorities detain Apple employees suspected of selling customer data
Apple's focus on privacy has always been a splinter in the eye of Chinese authorities.

I strongly believe that this is an excuse Chinese authorities had been looking for that will use to pressure Apple in China at the same time create the illusion to the general public to not trust Apple.

I find it especially suspecious that the Chinese media put so much emphasis the privacy concern of this event and in modern Chinese culture, privacy is much less regarded as compared to western countries.

Anyone remember the propaganda while Google was being driven out of China? Straight up false info about Google were broadcasted on CCTV-1, the prime time national channel. A lot of my friends in China became very patriotic and viewed Google as some sort of evil corporation trying to undermine Chinese culture.

yonkshi··on Python For Finance: Algorithmic Trading
Even if it's a zero sum game (which it's not), most participants are not quant/algo based. As long as you can beat 50% percentile, you can make a profit.

Sure you may not be as profitable as top quant companies, but do you really mind that much?

← PreviousPage 2 of 3Next →