Tag: cs

MuZero

MuZero learns a model that, when applied iteratively, predicts the quantities most directly relevant to planning: the reward, the action-selection policy, and the value function. When evaluated on 57 different Atari games – the canonical video game environment for testing AI techniques, in which model-based planning approaches have historically struggled – our new algorithm achieved a new state of the art. When evaluated on Go, chess and shogi, without any knowledge of the game rules, MuZero matched the superhuman performance of the AlphaZero algorithm that was supplied with the game rules.

Declarative web apps

I still have nearly all the questions I started this piece with. At the same time, by engaging with the paper and the topic I’m getting clearer in my own mind as to what my requirements for a future application composition system would be. To quote Grady Booch one more time: “the whole history of computer science is one of ever rising levels of abstraction.”

Against File systems

File systems unfit as distributed storage backends

Breaking the assumption that a distributed storage backend should clearly be layered on top of a local file system allowed Ceph to introduce a new storage backend called BlueStore with much better performance and predictability, and the ability to support the changing storage hardware landscape.

Interpretable models

Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead

2023-03-01: Math approaches can help with interpretation

Let’s take the set of all cat images and the set of all images that aren’t cats. We’re going to view them as topological shapes, or manifolds. One is the manifold of cats and the other is the manifold of non-cats. These are going to be intertwined in some complicated way. Why? Because there are certain things that look very much like cats that are not a cat. Mountain lions sometimes get mistaken for cats. Replicas. The big thing is, 2 manifolds are intertwined in some very complex manner.
I measure the shape of the manifold as it passes through the layers of a neural network. Ultimately, I can show that it reduces to the simplest possible form. You can view a neural network as a device for simplifying the topology of the manifolds under study.

Unix at 50

Maybe its pervasiveness has long obscured its origins. But Unix, the operating system that in 1 derivative or another powers nearly all smartphones sold worldwide, was born 50 years ago from the failure of an ambitious project that involved titans like Bell Labs, GE, and MIT. Largely the brainchild of a few programmers at Bell Labs, the unlikely story of Unix begins with a meeting on the top floor of an otherwise unremarkable annex at the sprawling Bell Labs complex in Murray Hill, New Jersey.

Self-locating uncertainty

Self-locating uncertainty is a different kind of epistemic uncertainty from that featured in pilot-wave models. You can know everything there is to know about the universe, and there’s still something you’re uncertain about, namely where you personally are within it. Your uncertainty obeys the rules of ordinary probability, but it requires a bit of work to convince yourself that there’s a reasonable way to assign numbers to your belief. In one sense, all of these notions of probability can be thought of as versions of self-locating uncertainty. All we have to do is consider the set of all possible worlds — all the different versions of reality one could possibly conceive. Some such worlds obey the rules of dynamical-collapse theories, and each of these is distinguished by the actual sequence of outcomes for all the quantum measurements ever performed. Other worlds are described by pilot-wave theories, and in each one the hidden variables have different values. Still others are many-worlds realities, where agents are uncertain about which branch of the wave function they are on. We might think of the role of probability as expressing our personal credences about which of these possible worlds is the actual one.

2 Generals problem

1 night in September 2018 the food delivery service Deliveroo went haywire. It sent some customers the same food order several times, and other customers got nothing. In this video, Tom Scott explains that this was a classic example of the “2 Generals” problem, and how Deliveroo’s “Night of the Multiple Orders” could have been mitigated — by using something called an “idempotency token,” which allows a transaction to only happen once.

Tensor Considered Harmful

Despite its ubiquity in deep learning, Tensor is broken. It forces bad habits such as exposing private dimensions, broadcasting based on absolute position, and keeping type information in documentation. This post presents a proof-of-concept of an alternative approach, named tensors, with named dimensions. This change eliminates the need for indexing, dim arguments, einsum- style unpacking, and documentation-based coding. The prototype PyTorch library accompanying this blog post is available as namedtensor.