Thanks, your comment made my day :)
2,242 karma · joined February 8, 2015
I'm a ML researcher working on Deep Learning for robotics. I'm skeptical of the symbolic approach by which 1) ontologists manually enter symbolic assertions and 2) the system deduces further things from its existing ontology. My skepticism comes from a position of Slavic pessimism: we don't actually know how to formally define any object, much less ontological relationships between objects. If we let a machine use our garbage ontologies as axioms with which to prove further ontological relationships, the resulting ontology may be completely disjoint from the reality we live in. There must be a forcing function with which reality tells the system that its ontology is incorrect, and a mechanism for unwinding wrong ontologies.
I'm reminded of a quote from the Alien, Covenant movie.
Walter : When one note is off, it eventually destroys the whole symphony, David.
There's an experience worse than blindness—it's the certainty that your vision is perfect and the horror that there's no world around you to see.
Q2: What physical & behavioral similarities do you observe between you and your half-siblings? Do you have any insight on the nature vs. nurture debate?
I am not really aware of any RL (deep or not) being deployed in commerce besides bandit problems used in recommender systems and the strong similarities between shortest-paths / graph planning. Any thoughts?
(cup, table, ball) = (cup, table) + (ball)
Unlike word2vec (where this property magically arises from training a language model), we enforce this property in our training objective in order to learn good visual models.
Unlike typical use of graph and relational networks, we explicitly wanted to throw away spatial relations between objects, to force the representation to keep track of what objects are still remaining even if the robot re-arranges the scene during the grasp. I think using a graph network makes sense if one wishes to model pairwise relations between objects though.
And then I followed up with an early TensorFlow visualizer https://github.com/ericjang/tdb
It turns out that while such tools seem useful at first glance, they turn out to not be that helpful to power users. For models bigger than LeNet, things get really ugly to visualize. And once you understand a high-level module and can take its training for granted, there isn't a need anymore to really look at it anymore. It can also be kind of annoying to tumble around in 3D when you just want to look at some activation maps. What does the 3D aspect of the visualization buy you here?
Tools like TensorBoard + Jupyter notebooks for inspecting weights and ad-hoc visualizations (e.g. VizDom) seem to strike the right balance.
If you want to continue pushing in this direction, I highly encourage embarking on an actual Deep Learning research project using your tool. In ML it's so important to dogfood your own software!
Another way to formulate this question: "given training data that only tells you about digits, how do you know whether something is a digit or not?" Given that the training data never actually defines what isn't a digit, how can we ensure that the model actually sees a digit at test time? If we cannot ensure this (e.g. an adversary or the real world supplies inputs), how can we "filter out" bad inputs?
A quick hack solution that works well in practice is to examine the "predictive distribution" across digit classes. Researchers have empirically found that entropy tends to be higher (i.e. more smooth) when the model sees an OoD input. However, the OoD problem is not fully solved.
Here's a nice survey paper on the topic: https://arxiv.org/abs/1809.04729
Note that methods that tie OoD to the task at hand (classification) are not actually solving OoD, they are solving "predictive uncertainty" of the task.
I feel like in the 21st century, a lot of folks have come to put a lot of faith in the "stocks in the long run" mantra.
But any stationary effect in the markets can be arbitraged away. If 9.7% long-term returns were guaranteed, wouldn't everyone just borrow 30-year loans on margin at 5% interest rate and invest it in S&P?
Edit: agree with sibling comment. If one loses 27% in the first year and makes 9.7% annually after that, they would be operating at net loss. So even if long-term gains were assured, volatility can still make investing on borrowed money unprofitable.
Founders of highly successful startups (or at least those on a promising growth trajectory) are in practice sensitive to which VCs participate in each round (and I suspect in the next few years they will start to pay attention to LPs as well). The "buy-in" price that you mention is not efficient because some people cannot participate no matter how much they are willing to pay.
Similarly, imagine if Uber wanted to hire a top Computer Vision researcher, and they could not. This hiring situation might not even get noticed by a LP, but an LP would surely notice if a VC was not able to invest their money in the next Uber / Facebook.
FWIW - I do think your assessment of oil is correct (since oil and money are both fungible). But for venture capital, one side of the investment (startup talent, VC skill and network, etc.) are not fungible.
(elaborated a bit more in a twitter rant: https://twitter.com/ericjang11/status/1054115724996694016)
As another commenter mentions, I do think it is productive to discuss the line with which something is no longer "rigorous" enough to be considered science. The caveat is that these discussions should be made by the lab doing the work, not by the Internet peanut gallery.
Media Lab's "research" resembles constructive science rather than deductive science (how does the world work?). We normally forgive science papers for doing engineering when they present something that is immediately useful to humanity (e.g. TensorFlow whitepaper). Artists (fashion designers, creatives, etc.) often make cool demos but fall short "revolutionary" products because the fabrication technology is not ready yet for their ideas. It still has a place, though, in science - if only to suggest unorthodox ways of approaching scientific inquiry.
At the very least, these works realize thought-provoking concepts a step farther than most of us are willing to commit to -- chatting about it on the Internet.
FWIW, my day job is doing science, if that lends my comment any credibility.
[1] http://matter.media.mit.edu/publications/article/making-data...
[2] https://twitter.com/ericjang11/status/1003471515323494405
Previous works in this area (and there are many) have tackled the exact problem of natural image generation. The specifics of the methods presented in the paper are of considerable interest to the ML/AI research community, but what I'd like to highlight here the results.
One might suppose that a neural network could never make images consistent with reality because there are so many hidden latent variables & strong causal factors that lead to the generation of coherent real-world imgaes. For instance, laws of light transport and camera perspective, smooth texture variation on objects, bilateral symmetry in organisms, what objects are "possible" and which objects are not.
Often to generate images that are at least plausible under our reality we need to make use of explicit physically-based rendering techniques like simulating the transport of light and accurately modeling a lot of high-fidelity geometry. We have to hard-code so many physical equations and rules into these systems, and even then we still have a hard time rendering everyday things (candlelight, soap bubbles, food).
Prior to this paper, I certainly had my doubts that neural networks could ever capture enough implicit "knowledge" about the world to synthesize an image (not in the training set, mind you) that could convince a human it was real. Machine Learning is the study of generalization and we know of no useful guarantees on generalization for finite-size training sets and overparameterized models.
To my knowledge, this is the first paper to generate high-fidelity natural images with no apparent visual artifacts (blurriness, weird textures that could not exist in the real world). The laws of light transport (NP-hard to compute) appear to be convincingly preserved, and I am blown away.
Say hello to Kwabena for me if you get the chance ;)