hardmaru (@hardmaru) | Robocrunch.Net

hardmaru

hardmaru @hardmaru

research scientist at google brain 🧠 in tokyo🗼

Tweets by hardmaru

Interesting thread about the future of large models, economic implications and the path to commercialization. It’s still a question mark whether startups can build a profitable and sustainable business model around large models, or whether these models simply become a commodity.

Shared by hardmaru at 5/14/2022

Neural Slime Volleyball is now ported to EvoJAX! Feels strange that the game I wrote in JS back in 2015 is now running a gazillion times faster in JAX on a GPU. The yellow agent here is trained in 4 minutes: https://t.co/3hAiu37E5h Original web version: https://t.co/xJwoQa1WYt https://t.co/cCFGWxdMbq

Shared by hardmaru at 5/13/2022

Interesting application combining computer vision and high-powered lasers to eradicate weeds on a farm.

Shared by hardmaru at 5/11/2022

“Are there any software engineers that switched into a machine learning role and found it a lot more stressful due to deadlines combined with the uncertainty of research?” Discussion: https://t.co/OHZdmj1ly2

Shared by hardmaru at 5/10/2022

The full documentation, tutorials, and code for “Evolution Gym: A Large-Scale Benchmark for Evolving Soft Robots” is finally out: https://t.co/CTa0Qhl61X https://t.co/FGbPJdilAN Please give @JagdeepBhatia8, the main author of the project, any feedback.

Shared by hardmaru at 1/30/2022

Interestingly when everyone wears a mask in public, I automatically form mental guesses about what each person might look like, and this mental process is performed seemingly unconsciously and effortlessly. Deep probabilistic models can be trained to do some version of this too:

Shared by hardmaru at 1/25/2022

Article in @petapixel about @GoogleAI's SR3 Image Super-resolution method. Neural network-based photo upscaling will likely be commonplace in most smartphones in the near future. https://t.co/7TKE3jeTv2

Shared by hardmaru at 9/1/2021

Cool “live coding” experiment with @zzznah extending Neural Cellular Automata to produce multiple self-organizing textures. I hope this kind of Twitch-like, “live coding” presentation format can become a thing for presenting some ML (and even RL) experimental results.

Shared by hardmaru at 8/31/2021

Fastformer: Additive Attention Can Be All You Need Fastformer is an efficient Transformer model based on additive attention with linear complexity. Experiments (on 5 NLP datasets) show that it is much more efficient than many existing Transformer models. https://t.co/QGEs2fQ9aS

Shared by hardmaru at 8/25/2021

A fantastic primer on AI Story Generation by @mark_riedl on @gradientpub https://t.co/z366u84LTj

Shared by hardmaru at 8/22/2021

Computer Scientists Discover Limits of Major Research Algorithm: A new result in complexity theory establishes why the gradient descent algorithm cannot solve some kinds of problems quickly. (@QuantaMagazine) https://t.co/yXgoNp5TdZ https://t.co/8RqT8RSKWS

Shared by hardmaru at 8/20/2021

Maze: A new framework for applied reinforcement learning. Some of the features are described in the blog. https://t.co/03aoomTsWt https://t.co/dWLDKBYU7N

Shared by hardmaru at 8/18/2021

Single cortical neurons as deep artificial neural networks Cortical neurons can be approximated (at millisecond resolution) by a temporal convolutional neural network with five to eight layers. @Mikilon @DavidBeniaguev @Segev_Lab https://t.co/NF8n6PfN46 https://t.co/1QI1o0ULtU

Shared by hardmaru at 8/12/2021

How to avoid machine learning pitfalls: A guide for academic researchers https://t.co/9yJc29p7AJ https://t.co/sCjAtdNsY2

Shared by hardmaru at 8/8/2021

Nice article from Unity about multi-agent RL DodgeBall Some strategies that emerged include an agent dedicated to defend their own base, distracting an opponent, waiting at the enemy's destination to recapture flag, and guarding a teammate with the flag. https://t.co/RymtyTsgKQ

Shared by hardmaru at 7/13/2021

Tune hyperparameters to get SOTA on some ML benchmark

Shared by hardmaru at 7/8/2021

Tencent deploys facial recognition to detect minors gaming at night. The initiative, dubbed “Midnight Patrol,” is aimed at deterring minors from using “tricks” to pose as adults between 10pm and 8am. 🤔 Article by @yaling_jiang in @SixthTone: https://t.co/rXiUCjgvfQ

Shared by hardmaru at 7/7/2021

Such an elegant way to represent a locomotion policy.

Shared by hardmaru at 7/4/2021

Talk by @karpathy about how vision-based sensors are outperforming legacy, single-purpose sensors (including radar). He also discuss the efforts behind collecting petabyte-sized datasets, Transformer-fusion architectures, supercomputers, and importance of owning the entire stack.

Shared by hardmaru at 6/22/2021

Engineering Sketch Generation for Computer-Aided Design (Autodesk Research) They look at 2 different generative models: CurveGen (think SVG primatives) and TurtleGen (think Turtle Graphics Logo language from the 1970s), for engineering sketch generation. https://t.co/Rn5gsCoDjC

Shared by hardmaru at 6/19/2021

Sketch-QNet: A Quadruplet ConvNet for Color Sketch-Based Image Retrieval Interesting work to tackle the color sketch-based image retrieval (CSBIR) problem, achieving new SOTA results. Authors are from a startup “Impresee” focused on sketch-based methods. https://t.co/TRaNECS128

Shared by hardmaru at 6/19/2021

There are important open challenges in the intersection of ML and cognitive science, with regards to not only object-part segmentation in both photos and human sketches, but also learning higher-order abstractions of images, and linking these to cognitive development. (2/2)

Shared by hardmaru at 6/19/2021

“Drawing is a cognitive technology.” @judyefan describes a theory of how contextual and perceptual information are combined. Humans navigate this axis (1/2):

Shared by hardmaru at 6/19/2021

Graph neural networks demonstrate promising potential for algorithmic reasoning (some early results for extrapolation), and in principle can be trained via backprop to learn drawing algorithms and even parse-tree generative models (i.e. L-Systems representations of plants). 2/2

Shared by hardmaru at 6/19/2021

Next talk, "Neural Algorithmic Sketching" from @PetarV_93: There is much similarity between free hand sketching and algorithmic reasoning, so sketching can serve as a basis for representation learning / algorithmic representations of concepts. 1/2

Shared by hardmaru at 6/19/2021

The 1st Workshop on Sketch-Oriented Deep Learning Workshop @CVPRConf kicked off! I'll attempt to live tweet the workshop. @junyanz89's talk on "Sketch Your Own Models" discuss combining human sketching to interface with the latent space of photos, and even rewrite entire models.

Shared by hardmaru at 6/19/2021

Event-based backpropagation can compute exact gradients for spiking neural networks They can backprop through a spiking neural network of leaky ‘integrate-and-fire’ neurons for general loss functions. Opens up new possibilities for neuromorphic hardware. https://t.co/6ATShRTO9M https://t.co/bccocv911q

Shared by hardmaru at 6/18/2021

A bit different from other ML roles “Passionate interest in areas such as collective intelligence, cellular automata, L-systems, neuroevolution, artificial life, self-organizing systems, open-ended novelty search, and demonstrate a working knowledge of deep learning frameworks.”

Shared by hardmaru at 6/18/2021

Thinking Like Transformers RNNs have direct parallels in finite state machines, but Transformers have no such familiar parallel. This paper aims to change that. They propose a computational model for the Transformer in the form of a programming language. https://t.co/OuPBSrS1EJ

Shared by hardmaru at 6/16/2021

The Modern Mathematics of Deep Learning A review paper (textbook) on the mathematical analysis of deep learning, with an emphasis on the generalization power of overparametrized neural nets and the role of depth in architectures. https://t.co/praDEEkR0k https://t.co/YgSSM2T3Kk

Shared by hardmaru at 6/14/2021

“We demonstrate that the brain, like language models, constantly predicts upcoming words in natural speech, 100s of milliseconds before they’re perceived. Our findings suggest LMs provide a biologically feasible computational framework for studying the neural basis of language.”

Shared by hardmaru at 6/9/2021

The hottest deep learning architecture right now is the multilayer perceptron.

Shared by hardmaru at 6/5/2021

“Traditional Reinforcement Learning (RL) algorithms either predict rewards with value functions or maximize them using policy search. We study an alternative: Upside-Down Reinforcement Learning, that solves RL problems primarily using supervised learning.” https://t.co/HqL8aNRnNC

Shared by hardmaru at 6/3/2021

Decision Transformer: Reinforcement Learning via Sequence Modeling A nice result in the paper: By training a language model on a training dataset of random walk trajectories, it can figure out optimal trajectories by just conditioning on a large reward. https://t.co/XnkZG4eiIU https://t.co/MeBT1LbCTh

Shared by hardmaru at 6/3/2021

Generative Art Using Neural Visual Grammars and Dual Encoders Cool work that evolves Lindenmeyer systems-based procedural artwork (SVG) from a text description embedding. Left image is a result of “Tiger in the Jungle” blog https://t.co/hVfhSfp0c7 paper https://t.co/kGEqvLcMVK

Shared by hardmaru at 6/2/2021

Cloud TPU Virtual Machines finally released These VMs run on TPU host machines that are directly attached to TPU accelerators, so everything feels like it's run locally. The new 𝗹𝗶𝗯𝘁𝗽𝘂 library supports TensorFlow, PyTorch, JAX, and soon, Julia. 🔥 https://t.co/Yhatj1xaKz

Shared by hardmaru at 6/2/2021

Just finished bidding for 50 NeurIPS papers. I'm at the "candy store" phase where I see many exciting abstracts full of interesting ideas. But once I read the papers, my enthusiasm cools off. The review process is flawed in many ways, but to me, reviewing is still a lot of fun.

Shared by hardmaru at 6/1/2021