HNHacker News
TopNewBestAskShowJobs

brianchu

2,341 karma · joined July 15, 2012

Looking for the next thing.

Worked on machine learning for medicine at Cardiogram. Studied CS at UC Berkeley. Former undergrad researcher and teaching assistant for Berkeley's ML class. Former intern at Twitter on Ads Analytics Infrastructure.

Email: bc@brianchu.com

github.com/bchu

twitter.com/brrrianchu

submissionscomments
brianchu··on AI’s Language Problem
Deep learning has succeeded tremendously with perception in domains that tolerate lots of noise (audio/visual). Will those successes continue with perception in domains that are not noisy (language) and inference/control, which the article touches on? I think it really is unclear whether those challenges will require fundamental developments or just more years of incremental improvement. If fundamental developments are needed, then the timeline for progress - which everyone in tech seems to be interested in - becomes much more indeterminate.

If you think about audio/visual data, deep nets make sense: if you tweak a few pixel values in an image, or if you shift every pixel value by some amount, the image will still retain basically the same information. In this context, linearity (weighting values and summing them up) make sense. It's not clear whether this makes sense in language. On the other hand, deep methods are state of the art on most NLP tasks, but their improvement over other methods isn't the huge gap as in computer vision. And while we know there are tight similarities between lower-level visual features in deep nets and the initial layers of the visual cortex, the justification for deep learning in NLP is simpler and less specific: what I see is the fact that networks have huge capacity to fit to data and are deep (rely on a hierarchy of features). My guess is we may need a fundamental breakthrough in a newfangled hierarchical learning system that is better suited for language to “solve” NLP.

I think there are similar limitations with control and inference. When it comes to AlphaGo the deep learning component is responsible for estimating the value of the game state; the planning component is done with older methods. This is much more speculative, but when it comes to the work on Atari games, for example, I suspect that most of what is being learned (and solved) is perception of useful features from the raw game images. I wonder whether the features for deducing game state score are actually complex.

I think what I'm trying to say is that when we look at the success of deep learning, we have to separate out what part of that is due to the fact that deep learning is the go-to blackbox classifier, and what part of this is due to the systems we use actually being a good model for the problem. If the model isn't good, does that model merely need to be tweaked from what we currently use, or does the model have to completely change?

brianchu··on Special projects
1) Discussion based on rationality, of those topics, is worthless. Elon Musk says he is almost certain we are living in a simulation, because of probabilities.

2) You're the one calling it idol worship. I'm pointing out that most people would not take this idea seriously if it weren't Elon Musk saying it - and I still believe this is true.

3) The high school comment is pretty relevant. An intro philosophy class will usually mention Descartes, who thought about this stuff 4 centuries ago.

4) I only said it's funny that grown adults take the simulation hypothesis seriously. Your statement was the one that extrapolated that to religion and found offense.

brianchu··on Special projects
Let's not resort to calling people names.

Your argument is ironically the reductive one. I have complete respect for people who choose to believe in an organized religion. But it remains the case that whether or not you can prove a higher deity, that in of itself has no bearing on how you should live your life. You can disprove the existence of god and still choose to believe in religion. You can prove the existence and still choose to not believe in any religious system. So proving the existence is worthless.

brianchu··on Special projects
It can't. We cannot distinguish between flaws and non-flaws. What is a flaw anyways? That is ill-defined - we are conjecturing about what will or will not be present in a reality that we have no indication will resemble our reality.
brianchu··on Special projects
It's funny to see grown adults in tech taking the simulation idea seriously just because Elon Musk said it. I was interested in science fiction in high school and thought about these things, but I concluded they were worthless.

Why? It is ontologically no different from 1) solipsism, a philosophical idea that even most philosophers considered worthless centuries ago, or more loosely 2) belief in a higher deity who created the world.

It's completely unfalsifiable. The answer to the question clarifies nothing about our understanding of the world [1] and furthermore has absolutely no impact whatsoever on how you ought to live your life.

[1] Even if you conclude we live in a simulation, we will never know for sure anything about the nature of the simulation or the nature of the world outside the simulation.

brianchu··on Startup School 2016 applications are open
I've been several times. I think if you're from Northern California (you can drive to the event in < 90 minutes), it's worth it because there's a pretty good crowd of interesting people to talk to. But I don't think it's worth flying in just for the event. It would be worth it if part of a larger Bay Area vacation / business / touring trip.
brianchu··on Turkish government revokes ham radio licenses
It has to be [flagged]/dead to vouch.
brianchu··on Turkish government revokes ham radio licenses
Note: users with enough karma (like myself) can click "vouch" on flagged posts. My guess is if enough users do this, that will unkill it (like it did now).
brianchu··on CS188 Intro to AI – Course Materials
This is a popular undergrad course at Berkeley. To be clear, it's mostly focused on GOFAI (good old fashioned AI) - things like tree/graph search, constraint satisfaction, logic, some basic graphical models, and a bit of RL/ML at the end. It's useful to know about (because these things sometimes show up as components of ML systems), but also fairly disjoint from the core machine learning field that people are undoubtedly interested in.
brianchu··on Linear Programming: How Does It Work?
Theoretically.

In practice, simplex is competitive with and often beats, interior point methods.

brianchu··on Ask HN: Why are some CS and EE programs co-located in the same department?
UC Berkeley only has a single EECS department. The department offers an EECS degree and a CS degree, but that is very misleading because the EECS degree only requires 2 EE classes while the CS degree only requires 1 EE class. Their curricula are essentially the same, the naming difference is mostly due to historical reasons.

There is a tremendous amount of overlap between the fields. Areas of especially large overlap: computer vision, machine learning, robotics, communication (compression and error correction), networking, computer architecture, etc.

brianchu··on End to End Learning for Self-Driving Cars [pdf]
Here is a great, very accessible video where the CTO of MobileEye (which provides some of the components for Tesla's "autopilot") explains his views on the challenges of end-to-end learning for autonomous vehicles, and why it's preferable to decompose the problem instead (still using deep learning for the decomposed modules). https://www.youtube.com/watch?v=GCMXXXmxG-I

I'm inclined to agree, especially because it helps in 1) providing diagnostic information (such as the great driving visualizations shown in the video), and 2) makes it easier to incorporate algorithms and sensors (like with Google's cars) as a redundancy in case the neural network hits a crazy edge case.

brianchu··on The Startup Zeitgeist
Nope. And why would I bother learning when I can just click a few buttons to get Slack? I'm pretty lazy and have no interest spending the weekend learning all the features of IRC.

Also:

> This is all on the free plan obviously, but guess what the next wave of tech startups are going to use

brianchu··on The Startup Zeitgeist
I'm at UC Berkeley right now. Most the college organizations I know use it (including the hacker's club I was a part of). Almost every computer science class uses it for internal staff communication, which is pretty important considering that our intro CS class has 40+ course staff/assistants. Nobody wants to set up and use IRC, which is ugly, time consuming to set up, and not fun to use.

I hear Slack is also gaining a lot of traction for open source community discussion (although IRC is still common).

This is all on the free plan obviously, but guess what the next wave of tech startups are going to use (or are already using). This is not including the big companies that are already adopting it.

There's probably more usage of Slack among CS students than there is usage of bitcoin, VR, drones, or any other technology popular among hackers that people argue (rightly or wrongly) are poised to go mainstream.

brianchu··on Bots won't replace apps, only better apps will replace apps
There is nothing human about chatbots. They are glorified menu-in-text-form delivery mechanisms. NLP technology isn't at the point where they become even slightly human-like when off-the-rails.
brianchu··on Bots won't replace apps, only better apps will replace apps
Great read. I found particularly insightful the idea that conversational UI is like skeumorphism - trying to shoehorn analog forms (conversation) into a digital app, bringing along a bunch of details and actions that don't serve any purpose or are worse than the alternatives.
brianchu··on Tinker with a Neural Network in Your Browser
I'm pretty sure he wasn't talking about the swiss roll specifically. Big gains in neural net performance have been made through better initialization schemes (not dataset specific, just in general, e.g. an initialization scheme might adapt the initial weight distribution depending on the number of hidden units in the next layer), and smaller models are in general more sensitive to initialization.
brianchu··on Ask HN: What makes a Senior Dev
Here's a really good Quora answer talking about how engineers get promoted at Google. I think Google is a good example to look at for promotion (but not necessarily for every issue) because they're known to have a very well-defined process for promotion.

https://www.quora.com/How-do-engineers-get-promoted-at-Googl...

brianchu··on Changing Education – How bootcamps outperform a university education
Okay, I agree in the abstract that "intense immersive study" is a better form of education than any other. This is really obvious. But the breadth and depth of knowledge required in the areas is huge. (e.g. for distributed systems, you'd ideally understand things at all levels of the computing stack - OS, networking, databases, C/C++, data structures, algorithms, some basic discrete math). This could easily require a "bootcamp" to be a year long. Would it really be a bootcamp then?

Also, the world experts in all those areas are at research universities. Now, it's true you could probably offer them a lot of money to teach bootcamp courses, but one of the best routes to developing skills in these areas is to do actual research in these fields. Research takes a long time (> 1 year), so this goes back to my original point.

brianchu··on Changing Education – How bootcamps outperform a university education
I study CS as UC Berkeley. If you want to be a web developer, a mobile developer, a product engineer, or something like that, go to a bootcamp. That, and then getting 2-3 years of work experience is a much better deal than studying CS in college, even Berkeley, and even if you get scholarships.

But I would like to point out that many of these criticisms don't apply to top schools like Berkeley. Berkeley's intro-CS classes are all taught by lecturers - and the lecturers are given tenure (e.g. people evaluated solely on teaching). As a result, I have had a nearly-uniformly-positive experience with the quality of teaching in CS - extending to and especially professors who are also leading researchers.

There are also more areas than buildings apps. If you're interested in distributed systems, robotics, machine learning, finance, etc., then going to a great CS school like Berkeley is a fantastic choice. There is truly cutting-edge, industry-leading work being done in distributed systems, computer architecture, machine learning, deep learning, and databases at Berkeley, and that undergrads are able to contribute to.

brianchu··on Enrollment Is Surging in Machine Learning Classes
Yep. Berkeley only recently started offering a ML class in 2013. The last time it was offered was 1988-1990, the last AI boom (assuming the course numbers meant the same thing back then).
brianchu··on Image Processing 101
Gaussian blur is a low-pass filter.
brianchu··on AlphaGo beats Lee Sedol again in match 2 of 5
Isn't that what I'm saying though? If you're tied/close going into the endgame, AlphaGo will probably win.
brianchu··on AlphaGo beats Lee Sedol again in match 2 of 5
I'm totally uninformed about Go, but by now it seems that unless you're clearly in the lead by the end of the midgame, AlphaGo is going to win, simply because at that point its Monte Carlo Tree Search is going to our-compute you in examining all the tactical variations in the endgame. Lee Sedol really seemed to be under a lot of time pressure at the end.

EDIT: clarified to what I originally meant: "end of midgame"

brianchu··on Q&A: Andrew Ng
For those interested in the "AI threat" discussion, here's an excellent thread on Quora about it with answers from Andrew Ng, Yoshua Bengio, and Pedro Domingos (along with other random opinions): https://www.quora.com/Is-AI-an-existential-threat-to-humanit...

As a side note, Quora has done some really excellent AMAs with a bunch of machine learning researchers, including Ng, Bengio, researchers at Google, and many more.

brianchu··on Ask HN: Is Python dying?
I do a little work in machine learning (as an undergrad). Deep learning specifically is dominated by Python and C++ (Python for the interface, C++/CUDA for the implementation). Theano's interface is in Python. Caffe has C++, Python, and Matlab interfaces (Python and C++ being the most popular in my experience). Tensorflow has C++ and Python interfaces. ipython notebooks started with Python. The odd one out is Torch; its interface is in Lua.

Machine learning also, at least from my academic view, seems to be dominated by C++/Matlab/Java/Python (and things seem to be moving to Python).

But more broadly, you'll have to learn dozens of languages as a programmer anyways. I wouldn't get too caught up in Python vs. other things.

brianchu··on Twitter's rumored algorithmic timeline switch
I don't get the complaints here. Most of the complaints about the algorithmic timeline are implementation details, not objections to the concept itself.

An algorithmic timeline does not necessarily mean live events are out of order. There is no reason the algorithm cannot be sophisticated enough to detect time-sensitive tweets and preserve the ordering of real time events. In fact, a sophisticated enough algorithm would cluster real time event-related tweets in chronological batches, and surface relevant replies / commentary around it.

I'm pretty sure that almost all the people threatening to leave Twitter will still be there when the dust settles.

brianchu··on How convolutional neural networks see the world
That is not an example of how weak our brain is. That is an example of how incredibly intelligent our brain is.

If you saw both tables in the real, physical, natural world, because of the effect of perspective, the table on the left is indeed physically longer than the width of the table on the right. Our brain is correctly inferring this from perspective/trigonometry. The visual system is designed to correctly interpret the natural world, not correctly interpret lengths of line segments on a slide. This illustrates how hard it is to create computer algorithms that robustly infer 3D from 2D (this inference requires taking cues from shading, perspective, etc.)

Almost all optical illusions are examples of the strength, not weakness, of the human brain.

EDIT: I've emphasized the word "natural". The problem with the below examples of "illusions" constructed in the physical world is this: what do you expect the brain to do otherwise? When you see a scene in the world, is it a weakness that our brain discards the (extremely unlikely) possibility the scene is just a massive hologram five inches from your eyes, or a meticulously painted and lit cardboard scene? I don't.

brianchu··on The Unreasonable Reputation of Neural Networks
I don't see much evidence in that paper that neural networks have similar representations to monkey visual cortex activations (the paper studied monkeys). They merely showed that the neural network representations were predictive of visual cortex responses in monkeys (e.g. linear regression was a good fit for using the NN representations to predict visual cortex responses in monkeys).
brianchu··on A Theory of Deep Convolutional Neural Networks for Feature Extraction
CNNs are not rotation invariant.
← PreviousPage 2 of 10Next →