MIT Technology Review analyzed 16,625 research articles to identify historical trends in AI research.
They scraped the papers from the artificial intelligence section of arXiv
(pronounced “archive”), the canonical online repository for computer science research. Their main takeaway:
Through our analysis, we found three major trends: a shift toward machine learning during the late 1990s and early 2000s, a rise in the popularity of neural networks beginning in the early 2010s, and growth in reinforcement learning in the past few years.
This last trend, the growth of reinforcement learning, has definitely been picking up outside of arXiv papers as well; see DT #6
for a few recent examples. Personally, I’m most excited to see how reinforcement learning will be applied outside of games, where most current research achievements seem to be.
But are there problems with the trend of approaching most ML problems with neural networks? Alan L. Yuille et al. argue that the current singular focus on deep nets in vision research is bad for the field because: researchers are chiefly targeting tasks that have lots labeled data available (which may not be the most important tasks); deep nets often don’t generalize well outside of the dataset they’re trained on (especially in terms of viewpoints); and deep nets are overly sensitive to context in a way a human wouldn’t be. A few of my professors have mentioned similar issues: people are quick to jump straight to complex, recent methods that perform well on benchmark datasets, without trying simpler, older methods first (which may have major computational and explainability benefits).
Uber has released Ludwig, and open-source toolbox for building deep learning models without writing code. The tool is built on top of TensorFlow, Google’s ML library. It’s centered around encoders that map different data types to tensors (like CNNs for images and RNNs for text), combiners that feed these tensors through neural networks, and decoders that map the tensors back to data. Uber has used Ludwig in several internal projects over the past two years, and it’ll be interesting to see what the wider community can do with it.
Ars Technica surveyed the different approaches 10 companies are taking to make lidar work. Lidar is the “laser radar” technology that self-driving cars (and other robots) use to create a 3D map of the world around them.
The basic idea of lidar is simple: a sensor sends out laser beams in various directions and waits for them to bounce back. Because light travels at a known speed, the round-trip time gives a precise estimate of the distance. While the basic idea is simple, the details get complicated fast. Every lidar maker has to make three basic decisions: how to point the laser in different directions, how to measure the round-trip time, and what frequency of light to use.
The surveys describes the tradeoffs between different techniques for each of these, and then dives deep into the choices that individual companies like Velodyne, Luminar, Blackmore, and several others have made. I’m surprised they’re publishing this as an online article–it would probably do well in an academic journal as well.