1257
Twitter
If traffic is so bad on game days—why isn't the city running Dodger Express buses from these Red Line stations? Whe…
from twitter_favs
11 hours ago
Twitter
Believing that it’s air conditioner water is the faith you’re baptized into.
from twitter_favs
19 hours ago
A conversation about how public transport really works | FT Alphaville
“So, I think, we’ve seen, as with this attempt to somehow harness physical space through the power of apps, that there’s just a basic philosophical problem there, which is that transport is fundamentally a physical, spatial problem. It is not fundamentally a communications problem or to the extent that it was a communications problem, we’ve gone most of the way, I think, in taking that friction out of the system. And what Uber is discovering, I think, what a lot of these tech firms are discovering is that taking that friction out of the system did not transform the fundamental reality of space and the math of labour and so on, which have really been the facts that have determined what’s possible in passenger transport and will continue to determine those things.”

“No, of course, the driverless car people will say, no, cars will fit closer together and they’ll be smaller and so we’ll fit more of them over the bridge but that’s a linear solution to an exponential problem. The other dimension of this problem that you must keep in mind is the problem of what we, in the business, call induced demand. And induced demand is the very simply idea that when you make something easier, people are more likely to do it and this is why, for example, when you widen a motorway, the traffic gets worse or it fills up to the same level of congestion that you had before. It’s because when you actually create new capacity, people use the capacity and you end up back in the same point.”

“If there’s a single concept that transport professionals almost all understand and almost nobody else understands, it’s this notion that the relationship between demand and usage is actually circular, the relationship between demand and capacity is circular. That is to say if/when we create more capacity, we trigger more demand. There’s a huge issue then because one of the things that Uber has done is very effectively induce demand for a whole bunch of new car trips in the city that weren’t happening before and this has had the effect, of course, of increasing congestion. The other thing they’re done, of course, is draw people off of public transport, which is a great way to increase congestion. And so this is why it’s tricky and this is why in your example of a bridge, if you widened the bridge but lots of people want to cross it, you’ll end up with a wider bridge that’s exactly as congested as it is now.”

“People who are in fortunate situations, who are much wealthier than average need to be very, very careful about using their personal tastes as an indication of what would be good city planning. The problem with that assumption is not that there’s anything wrong with being elite; the problem with that assumption is that if you’re an elite, that means you’re a minority and cities have to be designed so that they work for everyone. The unique feature of a city is that it doesn’t work for anyone unless it works for everyone. Everyone has to be able to get where they’re going and everyone else’s transportation choices affect your experience; that’s what congestion is.”

“If you build a tower of any kind where hundreds of people live at, basically, the same place and you expect them all to get into cars at eight in the morning with two or three empty seats, they are not going to be able to go anywhere because there is simply not room at such a high density for everybody to get into cars. This was the reigning fantasy of the 20th century. A very influential architect named Corbusier, in the 1940s, drew a famous image, in which we would all live in giant towers and there would be miraculously uncongested freeways running between them. He simply did not run the numbers on how many cars that is if everybody in that tower has a car. It doesn’t work. They don’t fit. It doesn’t matter whether they’re driverless or autonomous. It doesn’t matter what kind of car they are.”

“And the answer is, we did fine because most big data is big detail and the Tube isn’t going to change anything it does; the New York City Subway is not going to change anything it does, just because we now have data, knowing that at 07:34 p.m., six people from over here go over there to that park when the weather’s nice. That’s the sort of data we’re getting and it’s not relevant to the design of high-capacity transit systems. It is in the nature of highly efficient public transport that you do not vary your routing according to that sort of detail.”

“’m not surprised that a lot of very fortunate people who get around by car are trying to solve the problem of how I can continue getting around by car.”

“It’s difficult because their starting point is always, how can software help? And how can our product help? That’s what every merchant is going to ask. And the answer is, we can come up with ways for software to help around the edges; a great deal of that has been done. Real-time information about transit arrivals and departures was completely transformative to public transport; makes it so much easier to use. A lot of good technologies have come into public transport operations; I think, there’s room for more improvement in that but there isn’t a revolution out there because the problem remains spatial”
urbanism  transport 
2 days ago
How to unit test machine learning code. – Chase Roberts – Medium
Should have been called "some unit tests that might make sense for tensorflow models"
tensorflow  neuralnetworks  machinelearning  testing 
2 days ago
Can Silicon Valley workers rein in Big Tech from within? | Ben Tarnoff
“In my own conversations with its participants, they’ve explained that they don’t want to just keep pressuring their CEOs into dropping certain contacts. Rather, they want a seat at the table where decisions are made. They want to help determine how the technologies are built – and if they’re even built in the first place. As a letter written by Amazon workers explains: “We demand a choice in what we build, and a say in how it is used”

“People outside the industry might wonder why they should trust tech workers with more power. After all, tech workers already earn relatively high salaries, particularly at the big firms. If they gained more leverage, wouldn’t they just use it to demand even more money? On the other hand, if they did try to shape technology for social benefit, isn’t that a recipe for the worst kind of paternalism?”
unions  tech  politics 
4 days ago
Good cheap cars for under $5k
1489523754233.jpg (JPEG Image, 3842 × 2920 pixels) - Scaled (27%)
car  usa 
12 days ago
Locked doors, headaches, and intellectual need | Affording Play
“Consider Dan Meyer’s question for math educators: if math is the aspirin, then how do you create the headache?

Think of yourself as someone who sells aspirin. And realize that the best customer for your aspirin is someone who is in pain. Not a lot of pain. Not a migraine. Just a little.

One of the worst things you can do is force people who don’t feel pain to take your aspirin. They may oblige you if you have some particular kind of authority in their lives but that aspirin will feel pointless. It’ll undermine their respect for medicine in general.”

“Monads are a solution to a specific problem: the problem of repetitive code. If you write enough code in a functional programming language, you start to notice that you’re writing a lot of suspiciously similar code to solve a bunch of superficially different problems. Wouldn’t it be nice if you could just write this code once and then reuse it, instead of rewriting it slightly differently every time? I’m omitting a lot of detail here, but this is effectively what monads allow you to do.”

“By definition, the chief difference between experienced and inexperienced functional programmers is that experienced functional programmers have written tons of code in functional languages. They’ve all encountered repetition and sought solutions to it. In other words, they’ve felt the headache for which monads are the aspirin.

Beginners, on the other hand, haven’t written nearly as much functional code. They might not have noticed any recurring patterns yet; if they have, the repetition doesn’t yet bother them. The headache just isn’t there.

This is why it’s so hard to explain monads to beginners, especially with canned tutorials that try to explain what monads are without spending too much time on what they’re for. Monads are the solution to a problem that beginners haven’t yet experienced for themselves; and, as a result, they feel pointless, like something out of high-school math.

Remember: One of the worst things you can do is force people who don’t feel pain to take your aspirin. Likewise, I suspect that trying to “teach monads” to novice functional programmers who don’t yet understand the need for monads is likely to do more harm than good, creating further unnecessary confusion and perpetuating the myth that monads are intrinsically hard to understand.”
functional  maths  education 
15 days ago
World Cup 2018: The Yob-Swagger of Inger-Land
“I am always shocked by the way it takes hold, this faith in England, the desire for England to go all the way, as they say, in spite of the long record of thwarted hopes, hopes that are so necessary a prelude to the lingering after-taste—the permanent after-taste, if such a thing is possible—of ashes in the mouth that, as when savoring a complex wine, the fire of hope itself already burns with more than a hint of the ashes to come. Even bearing all this in mind, and even if the Russians are better-prepared, fitter, and, not for the first time in their history, fighting on home soil, I still have faith in our hooligans to show their mettle and do us proud.”

“I love Jamie Vardy with his yob face and yob haircut. Good old England, good old Yob-land. We want England to be non-racist, non-homophobic, non-misogynistic, and all that, but, God knows, however much we hate yobs, we don’t ever want England to be yob-free. Yes, yes, good old Bobby Moore, wiping his hand like a young King Lear before collecting the Jules Rimet trophy from the young Queen at Wembley in 1966, but in some ways, Liam Gallagher, former Oasis frontman and unrepentant super-yob, still full of yob-swagger—would be the ideal England captain.”

“Football is all about cheating, it’s nothing but cheating, cheating and moaning about being cheated. They might as well rebrand it cheatball and have the trophy recast in the image of Thierry “le tricheur,” Henry’s handling the ball against Ireland in 2009. I look back nostalgically to the days when diving—or simulating being fouled—was supposedly the preserve only of highly skilled foreigners, but England now has home-grown, world-class divers like Dele Alli of Tottenham.”

“The thing is, you see, I love England even if it is, in some respects, a bit of a shithole and, in others, a complete shithole. No one will ever put it better than D.H. Lawrence of Nottingham Forest FC who considered himself “English in the teeth of all the world, even in the teeth of England.””
culture  england  football 
16 days ago
Thermostats, Locks and Lights: Digital Tools of Domestic Abuse
“Their stories are part of a new pattern of behavior in domestic abuse cases tied to the rise of smart home technology. Internet-connected locks, speakers, thermostats, lights and cameras that have been marketed as the newest conveniences are now also being used as a means for harassment, monitoring, revenge and control.”
technology  harassment  crime 
16 days ago
Essays: Artificial Intelligence and the Attack/Defense Balance - Schneier on Security
Both attack and defense will benefit from AI technologies, but I believe that AI has the capability to tip the scales more toward defense. There will be better offensive and defensive AI techniques. But here's the thing: defense is currently in a worse position than offense precisely because of the human components. Present-day attacks pit the relative advantages of computers and humans against the relative weaknesses of computers and humans. Computers moving into what are traditionally human areas will rebalance that equation.
machinelearning  security 
16 days ago
Why I left small-city Wisconsin: the under-studied reason why Millennial couples are clustering in supercities – Greater Greater Washington
With more households looking to advance two careers, the pressure on job-rich, transit-rich areas has increased. With the power of two incomes, we are able to outbid single-income households. Single-career households are also more flexible in where they locate and are less dependent on transit.

Viewed from this perspective, the “new urban crisis” can be understood in part as the result of changes in gender roles and professional ambitions, as more couples equally prioritize two distinct careers.
urbanism 
16 days ago
Choose Boring Technology
“But what I’m aiming for there is not technology that’s “boring” the way CSPAN is boring. I mean that it’s boring in the sense that it’s well understood. It’s bad, but you know why it’s bad. You can list all of the main ways it will let you down.”

“If you’re giving individual teams (or gods help you, individuals) free reign to make local decisions about infrastructure, you’re hurting yourself globally.”
tech  production 
17 days ago
Coarse-ID Control – arg min blog
This is the thirteenth part of “An Outsider’s Tour of Reinforcement Learning.”

The general framework of Coarse-ID Control consists of the following three steps:

Use supervised learning to learn a coarse model of the dynamical system to be controlled. I’ll refer to the system estimate as the nominal system.
Using either prior knowledge or statistical tools like the bootstrap, build probabilistic guarantees about the distance between the nominal system and the true, unknown dynamics.
Solve a robust optimization problem that optimizes control of the nominal system while penalizing signals with respect to the estimated uncertainty, ensuring stable, robust execution.
reinforcementlearning 
17 days ago
Lost Horizons – arg min blog
As we discussed in the previous posts, 95% of controllers are PID control. Of the remaining 5%, 95% of those are probably based on receding horizon control (RHC). RHC, also known as model predictive control (MPC), is an incredibly powerful approach to controls that marries simulation and feedback.
reinforcementlearning 
17 days ago
Catching Signals That Sound in the Dark – arg min blog
This is the eleventh part of “An Outsider’s Tour of Reinforcement Learning.”
reinforcementlearning 
17 days ago
The Best Things in Life Are Model Free – arg min blog
This is the tenth part of “An Outsider’s Tour of Reinforcement Learning.”

PID stands for “proportional integral derivative” control. The idea behind PID control is pretty simple: suppose you have some dynamical system with a single input that produces a single output. In controls, we call the system we’d like to control the plant, a term that comes from chemical process engineering. Let’s say you’d like the output of your plant to read some constant value yt=v. For instance, you’d like to keep the water temperature in your espresso machine at precisely 203 degrees Fahrenheit, but you don’t have a precise differential equation modeling your entire kitchen. PID control works by creating a control signal based on the error et=v−yt. As the name implies, the control signal is a combination of error, its derivative, and its integral:

I’ve heard differing accounts, but somewhere in the neighborhood of 95 percent of all control systems are PID. And some suggest that the number of people using the “D” term is negligible. Something like 95 percent of the myriad collection of control processes that keep our modern society running are configured by setting two parameters. This includes those third wave espresso machines that fuel so much great research.
reinforcementlearning 
17 days ago
Clues for Which I Search and Choose – arg min blog
This is the ninth part of “An Outsider’s Tour of Reinforcement Learning.”

We have seen that random search works well on simple linear problems and appears better than some RL methods like policy gradient. Does random search break down as we move to harder problems? Spoiler Alert: No. But keep reading!

Let’s apply random search to problems that are of interest to the RL community. The deep RL community has been spending a lot of time and energy on a suite of benchmarks, maintained by OpenAI and based on the MuJoCo simulator. Here, the optimal control problem is to get the simulation of a legged robot to walk as far and quickly as possible in one direction. Some of the tasks are very simple, but some are quite difficult like the complicated humanoid models with 22 degrees of freedom. The dynamics of legged robots are well-specified by Hamiltonian Equations, but planning locomotion from these models is challenging because it is not clear how to best design the objective function and because the model is piecewise linear. The model changes whenever part of the robot comes into contact with a solid object, and hence a normal force is introduced that was not previously acting upon the robot. Hence, getting robots to work without having to deal with complicated nonconvex nonlinear models seems like a solid and interesting challenge for the RL paradigm.

Recently, Salimans and his collaborators at Open AI showed that random search worked quite well on these benchmarks. In particular, they fit neural network controllers using random search with a few algorithmic enhancements (They call their version of random search “Evolution Strategies,” but I’m sticking with my naming convention).
reinforcementlearning 
17 days ago
Updates on Policy Gradients – arg min blog
This is the eighth part of “An Outsider’s Tour of Reinforcement Learning.”
reinforcementlearning 
17 days ago
A Model, You Know What I Mean? – arg min blog
This is the seventh part of “An Outsider’s Tour of Reinforcement Learning.”

The role of models in reinforcement learning remains hotly debated. Model-free methods, like policy gradient, aim to solve optimal control problems only by probing the system and improving strategies based on past awards and states. Many researchers argue for systems that can innately learn without the complication of the complex details of required to simulate a physical system. They argue that it is often easier to find a policy for a task than it is to fit a general purpose model of the system dynamics.

On the other hand, in continuous control problems we always have models. The idea that we are going to build a self-driving car from trial and error is ludicrous. Fitting models, while laborious, is not out of the realm of possibilities for most systems of interest. Moreover, often times a coarse model suffices in order to plan a nearly optimal control strategy. How much can a model improve performance even when the parameters are unknown or the model doesn’t fully capture all of the system’s behavior?

In this post, I’m going to look at one of the simplest uses of a model in reinforcement learning. The strategy will be to estimate a predictive model for the dynamical process and then to use it in a dynamic programming solution to the prescribed control problem. Building a control system as if this estimated model were true is called nominal control, and the estimated model is called the nominal model. Nominal control will serve as a useful baseline algorithm for the rest of this series.
reinforcementlearning 
17 days ago
The Policy of Truth – arg min blog
This is the sixth part of “An Outsider’s Tour of Reinforcement Learning.”

Our first generic candidate for solving reinforcement learning is Policy Gradient. I find it shocking that Policy Gradient wasn’t ruled out as a bad idea in 1993. Policy gradient is seductive as it apparently lets one fine tune a program to solve any problem without any domain knowledge. Of course, anything that makes such a claim must be too general for its own good. Indeed, if you dive into it, policy gradient is nothing more than random search dressed up in mathematical symbols and lingo.

I apologize in advance that this is one of the more notationally heavy posts. Policy Gradient makes excessive use of notation to fool us into thinking there is something deep going on. My guess is that part of the reason Policy Gradient remained a research topic was because people didn’t implement it and the mathematics looked so appealing on its own. This makes it easy to lose sight of what would happen if the method actually got coded up. See if you can find the places where leaps of faith occur.
Lots of papers have been applying policy gradient to all sorts of different settings, and claiming crazy results, but I hope that it is now clear that they are just dressing up random search in a clever outfit. When you end up with a bunch of papers showing that genetic algorithms are competitive with your methods, this does not mean that we’ve made an advance in genetic algorithms. It is far more likely that this means that your method is a lousy implementation of random search.
reinforcementlearning 
17 days ago
A Game of Chance to You to Him Is One of Real Skill – arg min blog
This is the fifth part of “An Outsider’s Tour of Reinforcement Learning.”

Iterative learning control and RL merely differ insofar as what information they provide to the control design engineer. In RL, the problems are constructed to hide as much information about the dynamical system as possible. Even though RL practice uses physics simulators that are generated from well-specified differential equations, we have to tie our hands behind our back pretending like we don’t know basic mechanics and that we don’t understand the desired goals of our control system. As a result, RL schemes require millions of training examples to achieve reasonable performance. ILC on the other hand typically never requires more than a few dozen iterations to exceed human performance. But ILC typically requires reasonable models about the underlying system dynamics, and often assumes fairly well specified dynamics. Is there a middle ground here where we can specify a coarse model but still learn on actual physical systems in a short amount of time?
reinforcementlearning 
17 days ago
Lessons from Optics, The Other Deep Learning – arg min blog
It would be easier to design deep nets if we could talk about the action of each of its layers the way we talk about the action of an optical element on the light that passes through it.

We talk about convolutional layers as running matched filters against their inputs, and the subsequent nonlinearities as pooling. This is a relatively low-level description, akin to describing the action of a lens in terms of Maxwell’s equations.

Maybe there are higher level abstractions to rely on, in terms of a quantity that is modified as it passes through the layers of a net, akin to the action of lens in terms of how it bends rays.

And it would be nice if this abstraction were quantitative so you could plug numbers into a formula to run back-of-the-envelope analyses to help you design your network.

It would be easier to design deep nets if we could talk about the action of each of its layers the way we talk about the action of an optical element on the light that passes through it.

We talk about convolutional layers as running matched filters against their inputs, and the subsequent nonlinearities as pooling. This is a relatively low-level description, akin to describing the action of a lens in terms of Maxwell’s equations.

Maybe there are higher level abstractions to rely on, in terms of a quantity that is modified as it passes through the layers of a net, akin to the action of lens in terms of how it bends rays.

And it would be nice if this abstraction were quantitative so you could plug numbers into a formula to run back-of-the-envelope analyses to help you design your network.

There’s a mass influx of newcomers to our field and we’re equipping them with little more than folklore and pre-trained deep nets, then asking them to innovate. We can barely agree on the phenomena that we should be explaining away. I think we’re far from teaching this stuff in high schools.
deeplearning  optics  physics  education 
17 days ago
The Linear Quadratic Regulator – arg min blog
This is the fourth part of “An Outsider’s Tour of Reinforcement Learning.”
reinforcementlearning 
17 days ago
The Linearization Principle – arg min blog
This is the third part of “An Outsider’s Tour of Reinforcement Learning.”
17 days ago
Total Control – arg min blog
This is the second part of “An Outsider’s Tour of Reinforcement Learning.”
reinforcementlearning 
17 days ago
Introduction to Learning to Trade with Reinforcement Learning – WildML
Before looking at the problem from a Reinforcement Learning perspective, let’s understand how we would go about creating a profitable trading strategy using a supervised learning approach. Then we will see what’s problematic about this, and why we may want to use Reinforcement Learning techniques.

Luckily, there are solutions to many of the above problems. The bad news is, the solutions are not very effective.

Supervised Model Training: If necessary, you may train one or more supervised learning models to predict quantities of interest that are necessary for the strategy to work. For example, price prediction, quantity prediction, etc.

Policy Development: You then come up with a rule-based policy that determines what actions to take based on the current state of the market and the outputs of supervised models. Note that this policy may also have parameters, such as decision thresholds, that need to be optimized. This optimization is done later.

There are several possible reward functions we can pick from. An obvious one would the Realized PnL (Profit and Loss). The agent receives a reward whenever it closes a position, e.g. when it sells an asset it has previously bought, or buys an asset it has previously borrowed. The net profit from that trade can be positive or negative. That’s the reward signal. As the agent maximizes the total cumulative reward, it learns to trade profitably. This reward function is technically correct and leads to the optimal policy in the limit. However, rewards are sparse because buy and sell actions are relatively rare compared to doing nothing. Hence, it requires the agent to learn without receiving frequent feedback. An alternative with more frequent feedback would be the Unrealized PnL, which the net profit the agent would get if it were to close all of its positions immediately. Both of these reward functions naively optimize for profit. In reality, a trader may want to minimize risk. A strategy with a slightly lower return but significantly lower volatility is preferably over a highly volatile but only slightly more profitable strategy. Using the Sharpe Ratio is one simple way to take risk into account, but there are many others.

Instead of needing to hand-code a rule-based policy, Reinforcement Learning directly learns a policy. There’s no need for us to specify rules and thresholds such as “buy when you are more than 75% sure that the market will move up”. That’s baked in the RL policy, which optimizes for the metric we care about. We’re removing a full step from the strategy development process! And because the policy can be parameterized by a complex model, such as a Deep Neural network, we can learn policies that are more complex and powerful than any rules a human trader could possibly come up with. And as we’ve seen above, the policies implicitly take into account metrics such as risk, if that’s something we’re optimizing for.
quant  reinforcementlearning  machinelearning  finance 
17 days ago
Program Synthesis in 2017-18 – Alex Polozov
The list is skewed toward ML/AI rather than PL. This isn’t because of any personal biases of mine, but simply because the volume of synthesis-related papers in NIPS/ICLR/ACL/ICML over the last couple of years started exceeding PLDI/POPL/OOPSLA. Many prominent researchers now publish in both communities.

Many program synthesis applications are built on a top-down enumerative search as a means to construct the desired program. The search can be purely enumerative (with some optimizations or heuristics), deductively guided, or constraint-guided. A common theme of many research papers of 2018 has been to augment this search with a learned guiding function. Its role is to predict, for each branch at each step of the search process, how likely it is to yield a desired program, after which the search can focus only on the likely branches. The guiding function is usually a neural network or some other probabilistic model. The overall setup can be framed slightly differently depending on the interaction of components in the search process, but the key idea of combining a top-down search procedure with a trained model guiding it remains the same.

DeepCoder pioneered such combination of symbolic and statistical techniques in early 2017. That system learns a model that, given a task specification (input-output examples in the original work), predicts a distribution over the operators that may appear in the desired program. This distribution can be then used as a guiding function in any enumerative search

Graph Neural Networks, or GNNs, arose recently as a particularly useful architecture for tasks that involve reasoning over source code, including program analysis and program synthesis. The motivation here is the need to treat a program as an executable object. Most of the ML applications to code before 2017 treated programs either as a sequence of tokens or, at best, a syntax tree. This captures natural language-like properties of code, i.e. the patterns that arise in programs thanks to their intent to communicate meaning to other programmers. However, code also has a second meaning – it is a formal executable structure, which communicates a particular behavior to the computer. In other words, NLP-inspired methods, when applied to code, have to infer the semantics of for, if, API chaining patterns, and so on. One would hope that this is learnable by a deep neural network from the ground up, but in practice, it’s been exceedingly difficult.
deeplearning  programsynthesis  machinelearning  cs 
17 days ago
Program Synthesis Explained — James Bornholt — University of Washington
We start with some specification of the desired program. A synthesiser produces a candidate program that might satisfy the specification, and then a verifier decides whether that candidate really does satisfy the specification. If so, we’re done! If not, the verifier closes the loop by providing some sort of feedback to the synthesiser, which it will use to guide its search for new candidate programs.

If there is a solution, we pass the candidate program to the verify step. Remember that the candidate program satisfies all the test cases in the existing collection. The verify step is going to use an SMT solver to answer the question: Does there exist another program P’, different from the candidate program P, that also satisfies all the test cases in the existing collection, but on some input z disagrees with P?

Oracle-guided synthesis assumes you have an oracle, an existing implementation of the program you’re trying synthesise. We don’t actually need to inspect the implementation; we treat it as a black box that provides outputs when we supply inputs. But even this seems a little absurd: why synthesise a program we already have?

The authors use two domains to illustrate why their technique is useful. The first is the traditional suite of bit-vector benchmarks from Hacker’s Delight. Many synthesis papers use these benchmarks because they are examples of small, unintuitive programs. Instead of trying to come up with the most efficient bit-twiddling hack, programmers can write a simple, inefficient implementation of a bit-vector manipulation, and the synthesiser uses this implementation to produce an optimal program. The second domain is program de-obfuscation – taking an obfuscated program as the oracle, and synthesising a new, simpler program that matches its behaviour.
Stochastic superoptimisation uses Markov-chain Monte Carlo (MCMC) sampling to search the space of programs in a more guided way. Essentially, stochastic superoptimisation defines a cost function that measures how “good” a candidate program is, and uses MCMC (in particular, the Metropolis algorithm) to sample programs that are highly weighted by that function. This bias means MCMC search is more likely to visit programs that are nearer to optimal. Although this is still a random search of a very large space, the bias means that empirically, stochastic superoptimisation often quickly discovers correct programs that are almost optimal.
cs  programming  programsynthesis 
17 days ago
Notes on notation and thought
I'm collecting quotes on interesting notations—both powerful ones and bad ones—and how they influence thought.
cs  maths  physics  dance 
18 days ago
Renaissance's Skill and Facebook's Buyback - Bloomberg
"Signals that made a lot of sense that were very strong" are what Renaissance is not looking for. Investing wisdom that you can sum up in a sentence, or a paragraph, or a book, or a human lifetime of learning and understanding: That's not for them. What they want are signals that don't make much sense, that aren't very strong, that you can't grasp intuitively and then go out and implement. They want stuff that looks like noise, but that in the hands of a powerful enough computer with cheap enough implementation costs, can make a little bit of money. Over and over again.

And so Renaissance doesn't really hire people because they understand investing. "A résumé with Wall Street experience or even a finance background was a firm pass" early on at Renaissance, and it's full mostly of speech-recognition specialists and astrophysicists and string theorists, scientists who "excel at screening 'noisy' data." It's a pattern-recognition firm, a data-processing firm, a firm for finding correlations and signals and using those signals to make frequent, tentative, marginally profitable decisions. Which add up to 40 percent annual returns, year after year.

It's a little embarrassing, no? That investing is best understood by people who don't understand investing? That it's a trivial application of broader data-science principles, best addressed by people who were trained on harder and more interesting applications?
datascience  machinelearning  quant  finance 
18 days ago
Remote Only
Hiring and working from all over the world instead of from a central location.
Flexible working hours over set working hours.
Writing down and recording knowledge over oral explanations.
Written down processes over on-the-job training.
Public sharing of information over need-to-know access.
Opening every document to change by anyone over top down control of documents.
Asynchronous communication over synchronous communication.
The results of work over the hours put in.
Formal communication channels
professional  remote 
18 days ago
Foot-candles: the different paths to tech – Alice Goldfuss
The deeper you dive into programming, the more you will run into topics covered by CS degrees. This may make you feel extremely behind and out of your depth. When this happens, keep the following in mind:

Your lack of knowledge in these topics doesn’t negate the work you’ve already done.
You know things CS grads don’t.
It’s likely your understanding of the topic is fresher and more complete than a CS grad who hasn’t touched it in years.
Everyone learns things in different orders and at different times, including CS grads.
Some things you will never need to know.
professional  cs 
18 days ago
Been Down So Long It Looks Like Debt to Me | M.H. Miller
College, which cost roughly $50,000 a year, was the only time that money did not seem to matter. “We’ll find a way to pay for it,” my parents said repeatedly, and if we couldn’t pay for it immediately, there was always a bank somewhere willing to give us a loan. This was true even after my parents had both lost their jobs amidst a global financial meltdown. Like many well-meaning but misguided baby boomers, neither of my parents received an elite education but they nevertheless believed that an expensive school was not a materialistic waste of money; it was the key to a better life than the one they had. They continued to put faith in this falsehood even after a previously unimaginable financial loss, and so we continued spending money that we didn’t have—money that banks kept giving to us.

I’ve spent a great deal of time in the last decade shifting the blame for my debt. Whose fault was it? My devoted parents, for encouraging me to attend a school they couldn’t afford? The banks, which should have never lent money to people who clearly couldn’t pay it back to begin with, continuously exploiting the hope of families like mine, and quick to exploit us further once that hope disappeared? Or was it my fault for not having the foresight to realize it was a mistake to spend roughly $200,000 on a school where, in order to get my degree, I kept a journal about reading Virginia Woolf?
debt  investments  education  usa  finance  politics 
18 days ago
As We May Think - The Atlantic
As Director of the Office of Scientific Research and Development, Dr. Vannevar Bush has coordinated the activities of some six thousand leading American scientists in the application of science to warfare. In this significant article he holds up an incentive for scientists when the fighting has ceased. He urges that men of science should then turn to the massive task of making more accessible our bewildering store of knowledge. For years inventions have extended man's physical powers rather than the powers of his mind. Trip hammers that multiply the fists, microscopes that sharpen the eye, and engines of destruction and detection are new results, but not the end results, of modern science. Now, says Dr. Bush, instruments are at hand which, if properly developed, will give man access to and command over the inherited knowledge of the ages. The perfection of these pacific instruments should be the first objective of our scientists as they emerge from their war work. Like Emerson's famous address of 1837 on "The American Scholar," this paper by Dr. Bush calls for a new relationship between thinking man and the sum of our knowledge. — THE EDITOR
science  history  war  InformationManagement  usa 
18 days ago
How Quora’s Head of Data Science Conducts Candidate Interviews
Eric does want people to be interested in the data. Quora has more than 200 million monthly active users, who interact through questions ranging from, “Why do bees die after they sting?” to “What are the best tips for raising venture capital funding?”

That data is interesting! But when the first thing that a candidate from a research background says in an interview is that they’re interested in Quora’s dataset, “that can be a red flag. It sounds like you’re still thinking about writing papers.”
datascience  interview  professional 
18 days ago
Getting There: Prominent urban thinker gives Spokane transportation a mixed grade | The Spokesman-Review
“From a communication perspective, it’s one of the most counter-intuitive conversations you can have with your constituents, that more parking isn’t necessarily good, that wider roads just induce more driving,” he told the 20 or so city planners, mayoral advisers and downtown decision-makers gathered on the first floor of City Hall.
urbanism 
19 days ago
« earlier      
1970s 20c academia adversarial advertising ai algorithms amazon analytics anecdata anomaly antarctica api apple architecture art arxiv astro async audio aws backup bash bayes bias bigdata bitcoin book books brexit business c calculus california capitalism car causality cia climate clojure cloudfront communication concurrency conference crime cryptocurrency cryptography cs csv culture dask data database dataengineering datascience datastructures deeplearning design devops differentialprivacy diversity docker economics education engineering english espionage ethics etsy eu europe facebook family fatml federatedlearning feminism fiction film finance fpga functional fzf gaussianprocesses gawker git github golang google h1b hardware haskell health hiring history homomorphic housing http https immigration infrastructure internet interpretability interview investments java jobs journalism js jupyter keras kubernetes labour lambda language law lda legal library linearalgebra linux literature losangeles machinelearning macos make management map mapreduce maps marketing math maths me media module money movies music neuralnetworks newyork nlp notebook numpy nyc oop optimization os package pandas patterns phone physics politics polling presentation privacy probabilisticprogramming probability product professional programming psephology publishing pycon2017 pymc3 pytest python pytorch pywren quant r race racism recipe recommendation reinforcementlearning religion remote research retirement review rnn russia rust s3 sanfrancisco science sciencewriting scientism scifi scikitlearn search security sentiment serverless sexism siliconvalley slack socialism socialmedia spark sql ssh ssl stan startup statistics streaming summarization surveillance talk tax tech technology tensorflow testing texas text timemachine timeseries tmux transport travel trump tutorial tv twitter uber uk unix urbanism usa versioncontrol video vim visualization vpn war web webdev word2vec writing

Copy this bookmark:



description:


tags: