ogrisel

16 ms·

121.
by ogrisel 10y ago
If you look at the source you see that the \cite{} statements do reference meaningful anchors but that the bibtext file seems to be missing from the LaTeX archive: https://arxiv.org/format/1611.00020v1
122.
by ogrisel 10y ago
Not everything is crappy at IKEA: you can get cheap mass pine wood furniture that looks pretty decent (if you paint it to fit your tastes). The cost of such furniture is often cheaper at IKEA than buying the planks of wood at a consumer sto
123.
by ogrisel 10y ago
> For instance I found a browser bug in the code submitted on a stack overflow answer [1] and I edited it and it got rejected because it wasn't in the spirit of what the original author intended or something like that. It occurred s
124.
by ogrisel 10y ago
Google never stated they use those to train models as far as I know. It seems that they are primarily used to spare energy when deploying trained models at scale.
125.
by ogrisel 10y ago
Apparently the goal of this guide it to setup a kubernetes from scratch (on top of VMs provisioned on an orchestration-agnostic public cloud), not how to use a managed kubernetes cluster as provided by openshift.
126.
by ogrisel 10y ago
I would bet that is limited to low-energy inference which is the most important impediment to deploy deep vision models in mobile devices.
127.
by ogrisel 10y ago
You might want to discover the full Pareto frontier to gain some insights on the structure of the trade-off. I think the best methods to explore the Pareto frontier are based on the concepts of evolutionary computation like NSGA-II and SPE
128.
by ogrisel 10y ago
The classical solver used to train kernel SVMs and implemented in libsvm [1] has a time complexity in between o(n^2) and o(n^3) where n is the number of labeled samples in the training set. In practice it becomes intractable to train a non-
129.
by ogrisel 10y ago
This discovery is about Proxima Centauri which is slightly closer: 4.25 ly.
130.
by ogrisel 10y ago
scipy.optimize can only do single-criterion optimization (scalar-valued objective function instead of vector-valued objective functions).
131.
by ogrisel 10y ago
Poor performance is generally evaluated by considering the validation error. This paper only cares about the training error: it is about the presence or absence of bad local minima in the optimization problem that one has to solve when trai
132.
by ogrisel 10y ago
The statistical model is a Bernoulli distribution where the expected value p is parameterized by the logistic sigmoid function applied to a linear model of the input variables. The choice of the use of the sigmoid function does not stem fro
133.
by ogrisel 10y ago
It's both simple to use and very easy to customize to build ad-hoc architectures with custom nodes. The development is very active and it's well documented. It can also use either tensorflow or theano as runtime.
134.
by ogrisel 11y ago
Yes I though the same. Especially when it walks in natural landscapes (e.g. on the beach and the snowy forest). I find the view of autonomous human artifacts roaming in otherwise unaltered natural landscapes aesthetically very pleasing. Sci
135.
by ogrisel 11y ago
The old backprop from the 80's / 90's is still in use and the primary way to train deep nets. We tend to call it SGD (on a composition of differentiable functions) nowadays but it's the same algorithm.
136.
by ogrisel 11y ago
It's not an arbitrary dataset. To be taken seriously you need to beat SOTA on several standard benchmark datasets where only reaching SOTA on one is already hard. Or alternatively, reach comparable level of validation accuracy with sig
137.
by ogrisel 11y ago
the blog post is very good, my comment was just about adding the missing info to the HN post title. Edit the title of the HN post has changed since I made this comment... probably by being merged with an earlier submission of the same URL
138.
by ogrisel 11y ago
I was not trying to prove a point. It's just that giving a parallel speedup for a mostly CPU-bound task without mentioning the number of cores is not very interesting. 25x on 64 cores is not as interesting as 25x on 32 cores. If you kn
139.
by ogrisel 11y ago
...on a 32 physical cores machine.
140.
by ogrisel 11y ago
This looks very similar to what is offered by https://platform.sh/ . Anybody familiar with both could contrast them?
141.
by ogrisel 11y ago
A human brain also has a limited number of muscles to control to interact with the world.
142.
by ogrisel 11y ago
I don't understand why would the "number of spaces" matters. What matters is can you design a learning algorithm that performs well in interesting spaces such as: - discrete spaces such as atari games and go, - continuous spa
143.
by ogrisel 11y ago
see also by the same team: http://arxiv.org/abs/1602.01783
144.
by ogrisel 11y ago
...and it's possible to train learning agents to sense and interact with a world described by high dimensional continuous vector spaces, for instance using conv nets (for sensing audio / video signals) and actor-critic to learn an
145.
by ogrisel 11y ago
You can do math with continuous and infinite dimensional spaces.
146.
by ogrisel 11y ago
There is https://github.com/pluskid/Mocha.jl but it seems more focused on vision tasks at least for now.
147.
by ogrisel 11y ago
Python and lua both have good frameworks for developing such models (tensorflow, theano possibly with lasagne or keras, torch, caffe...). The visual part need access to efficient GPU-based implementation of convolutions (typically using cuD
148.
by ogrisel 11y ago
Out of curiosity, is this work being submitted for a conference or journal? Also have you tried to run DMN on the Text QA / Reading comprehension datasets from DeepMind? Teaching Machines to Read and Comprehend https://githu
149.
by ogrisel 11y ago
Maybe this could be done with machine learning to learn an approximate function to predict the outcome of the paired simulation from the outcome of the isolated simulations.
150.
by ogrisel 11y ago
It's a theoretical breakthrough that explains an unexpectedly bad behavior of current designs that was not understood so far: a discrepancy between theory and experimental observations. Now that the model is fine enough to explain the

More ›