inFERENCe

I love reading these two for Python/Recommender Systems: 1. Ben Frederickson: https://www.benfrederickson.com/blog/ 2. Erik Bernhardsson: https://erikbern.com/ And for pure ML: 1. Ferenc Huszár: https://www.inference.vc/ 2. John Langford: https://hunch.net (although he hasn't written a technical post in a while)

I don't have any great recommendations and unfortunately my advice may be not what you want to hear. What I tell my students is "You don't need to know math to build good models, but you need to know math to know why your models are wrong." But this is even a contentious statement within the community. (Personally I'm more interested in exploring what we can build and understand rather than focusing on throwing more compute and data at problems. There's a lot of work to be done that does not require significant compute, but it isn't flashy and you'll get little fame. Every famous model you know has some unsung hero(s) who built the foundation before compute was thrown at the problem). I was previously a physicist and we similarly frequently express that you do not know the material unless you can do the math. Physicists are trained in generating analogies as they help communication but this sometimes leads to people convincing themselves that they understand things far more than they actually do. They say the devil is in the details, and boy are there a lot of details. (Of the science communicators, I'm happy those are the ones you mention though!) But do not take this as gatekeeping! These groups are often happy to help with the math and recommend readings. ML is kinda a while west and you can honestly pick a subdomain of math and probably find it useful, but I would start by making sure you have a foundation in multivariate calculus and linear algebra. As to paper…

In the language of do-calculus, counterfactuals have a very specific meaning. Ferenc gives some great examples here: https://www.inference.vc/causal-inference-3-counterfactuals/ The main idea: Let Y be the set {y, not y} and X be the set {x, not x}. y denotes timing out, and x denotes “Kubernetes not starting the pod”. We have some distribution P(Y=y|X=x) (in this case, the probability of “timing out” (y) given “Kubernetes not starting the pod” x. The counter factual distribution is NOT the same as P(Y= not y | X= not x) (the “probability of not timing out given Kubernetes starting the pod”). The counterfactual is P(Y’=not y| X=x, Y=y, X’=not x), or “the probability of not timing out given that it did originally time out when Kubernetes did not start the pod, and given that this time it did start the pod”. This aligns roughly with what you are saying above.

Search for example on "deep learning". Results: + nice: open books (deeplearningbook.org) + not nice: online seminars (udemy, udacity, edx) + not nice: wikipedia copycats (investopedia) + not nice: SEO optimized business analytics sites recognizable by their top 10 this, top 20 that (kdnuggets) + not nice: commercial software (mathworks) + not nice: big companies hitchhiking on buzzwords (sas) What would be nice are websites and people that are often linked from HN. For example http://www.inference.vc rather than sas. Or https://distill.pub/ rather than investopedia. Or https://www.tensorflow.org rather than mathworks. If good sites are showing up in the top results more often, there will also be more incentive to keep up the good work as well.

In vanilla policy gradient, one plays the game to the end and then bumps the probability of all actions taken by the agent up (if AlphaGo won) or down (if it lost). This is very slow because there ~150 moves in an expert game, and we do not know which moves caused decisive victory or loss - i.e. the problem of "long term credit assignment". Also, it is actually preferable to compute the policy gradient with respect to the advantage of every action, so we encourage actions that were better than average and punish actions that were worse than average - otherwise the policy gradient estimator has high variance. I think about MCTS in the following way: suppose you have a perfect "simulator" for some reinforcement learning task you are trying to accomplish (i.e. real-world robot grasping for a cup). Then instead of trying to grasp the cup over and over again, you can just try/"plan" in simulation until you arrive at a motion plan that picks up the cup. MCTS is exactly a "planning" module, and it works so well in Go because the simulator fidelity is perfect. AlphaGo can't model adversary behavior perfectly, but MCTS and the policy network complement each other because the policy reduces the search space of MCTS. As long as the best adversary is not far away from the space that MCTS + policy is able to consider, AlphaGo can match or beat the adversary. Then, we train the value network to amortize the computation of the MTCS operator (via Bellman equality). Finally, self-play is…

First of all, let me say that what you all have put together is truly excellent. It covers the sort of things toy treatments and papers leave out or gloss over if you actually want to get something working in the real world. That said, I strongly disagree with your disagreement. There was a recent paper whose abstract I read that made me think of homotopies between convolutional networks. Unfortunately I lost the paper behind some stream and never got to read it proper. In the context of that, I realized that the search for convnet design will likely soon be highly automatable, obsoleting much of the work that many DLers are doing now. What will be future proof is understanding information theory so that loss functions become less magical. Information theory is needed to understand what aspects of the current approaches to Reinforcement learning are likely dead ends (typically to do with getting a good exploration strategy, also related to creativity). Concentration of measure is vital to understanding so many properties we find in optimization, dimensionality reduction and learning. Understanding learning stability and ideal properties of a learner/convergence means being comfortable with concepts like Jacobians and semi-positive definiteness for a start. Probability theory is needed for the newer variational methods, whether in the context of autoencoders or a library like Edward (whose like I think is the future). Functionals and the variational calculus is becoming…

...because stochastic methods are implicit regularizers, leading to solutions that generalize better. Let's spell it out for those that don't know. https://www.inference.vc/notes-on-the-origin-of-implicit-regularizatio…

https://www.inference.vc/we-may-be-surprised-again/ The author wrote this article one month ago, and he mentioned where he thought wrong about DL

I am surprised the SGD extensions of https://arxiv.org/abs/1704.04289 have not been picked up yet (via https://www.inference.vc/everything-that-works-works-because-its-bayes…

The intro to the paper has a nice comparison to other similar methods (generative and non-generative) and the blog post linked in this article by inFERNCe https://www.inference.vc/unsupervised-learning-by-predicting-noise-an-… has a nice comparison at the end to different unsupervised methods and where this method adds novelty (or doesn't!)

Measuring causality is hard. Without randomized control trials, you pretty much need to assume some causal graph structure (plain data is not enough). See: "ML beyond Curve Fitting: An Intro to Causal Inference and do-Calculus" http://www.inference.vc/untitled/ When it comes to this SO vs HN article based on data one can look at correlations. Even if some delayed correlations, then still - it does not imply causation (e.g. HN crowd my by faster to post links to new technologies than SO crowd to questions).

Those images [0] inputs that were optimised to maximise a certain classification response were cool! Instead of going to this peak of the response function, is there a way to explore the shell where the actual images reside? Would such images look, to our eyes, more like real input than the optimised input? I suspect they won't, but I still would like to see what is between the dogs! [0] http://www.inference.vc/content/images/2017/11/Screen-Shot-2017-11-09-…

Absolutely no. Random search does not work in high dimensions because of the "curse of dimensionality" - the number of directions to search grows exponentially with dimension. Gradient descent avoids the problem because it knows the right direction. Evolutionary strategies are doing (natural) gradient descent using an estimate of the gradient. Here's a good article: http://www.inference.vc/evolutionary-strategies-embarrassingly-paralle…

This reminds me of Ferenc's reminder that 'deep learning is easy' - 'http://www.inference.vc/deep-learning-is-easy/

I think you are on to something. this post was on HN a few weeks ago about how deep learning is no longer interesting because it has become routine. My first thought when I read that was "deep learning just became interesting because now it's stable enough to build a business around". The end of one cycle is just the beginning of another one. Find out what that new cycle is, hop on, and strap in. http://www.inference.vc/deep-learning-is-easy/ https://news.ycombinator.com/item?id=11006067

The process you are describing can be automated a bit. Imagine your feed reader automatically subscribes you to feeds of articles you liked (star/thumbs up/favorite). Let's say you start subscribed to an HN feed of items that got more than 100 upvotes: https://hnrss.org/newest?points=100&count=100 Then you find the OP link and like it in your feed reader: https://www.niemanlab.org/2022/12/this-is-the-year-of-the-rss-reader-r… Your reader automatically subscribes you to the source feed of this article: https://feeds.feedburner.com/NiemanJournalismLab You like more content => you get subscribed to more feeds automatically. Your feed reader can also take care of unsubscribing you from feeds that are posting too often or are not posting interesting to you content. If you don't like the content of a feed for awhile then your connection to that feed becomes weaker and the content from that feed is ranked less prominently. The content from feeds with a high signal-to-noise ratio (ie, how often you like content from that feed) is ranked higher for you. This is how my little project https://linklonk.com works. If you'd like to give it a try: "thumb up" a few items from the above HN feed: https://linklonk.com/feed?url=https:%2F%2Fhnrss.org%2Fnewest%3Fpoints%… and then check your recommendations on the home page. Through this mechanism I am currently subscribed to 201 RSS feeds and I know that I won't miss content from high "signal-to-noise ratio" feeds. Here is a sample…

This claim is unsupported: > The machine, not smart enough to do the actual difficult job of converting these sophisticated image types to each other Obviously it found an easy way to solve the problem it was given: stenography But could it have solved the problem the researchers intended, if they had framed it correctly? There's no evidence either way for this particular algorithm, but in general this is not hard. This is usually called style transfer and I don't see any reason to believe that standard techniques[1] wouldn't be able to solve the street-map-to-aerial-map problem. And it's pretty well established that adding a bit of noise[2] during training helps GANs avoid these kinds of problems. [1]: https://medium.com/tensorflow/neural-style-transfer-creating-art-with-… [2]: https://www.inference.vc/instance-noise-a-trick-for-stabilising-gan-tr…

My mentor was very very adamant about Bayesian network and hierarchical as being deep learning. He sees the latent layer in the hierarchical model as the hidden layer and the Bayesian just have a strict restrictions/assumptions to the network where as the deep learning is more dumb and less assuming. A few of my professor thinks that PGM, probability graphical model is a super set of deep learning/neural network. This is where my thinking come from. IIRC, a paper have shown that gradient descent seems to exhibit MCMCs (blog with paper link inside that led to this conclusion of mine: http://www.inference.vc/everything-that-works-works-because-its-bayesi… But I am not an expert in Neural Network nor know the topic well enough to say such a thing. Other than was deferring to opinions of some one that's better than myself. So I'll keep this in mind and hopefully one day have the time to do more research into this topic. Thank you.

A few remarks: In the "Doing bad digital humanities with color vectors", if you consider colors as 3D vectors, which they do, you'll see that summing enough uniformly sampled vectors always gives medium browns, because that's the color in the middle of the colorspace. Instead, you should model colors in a polar space and sum vectors in that space. This will prevent going inside the sphere and losing color saturation. It's explained quite well here in the Interpolation section: http://www.inference.vc/high-dimensional-gaussian-distributions-are-so… If you want to understand contemporary use of words embeddings in ML, a nice simple model is explained, with full code, here: https://blog.keras.io/using-pre-trained-word-embeddings-in-a-keras-mod… The original model comes from Kim 2014, https://arxiv.org/abs/1408.5882 It's a very neat use of CNNs for language processing, instead of the more popular RNNs/LSTMs. CNNs have the advantage of training much faster.