Taco Cohen (@TacoCohen) | Robocrunch.Net

TacoCohen

Taco Cohen @TacoCohen

Deep learner at Qualcomm AI Research (via Scyfer acquisition). Spent time at UvA, Deepmind, OpenAI.

Tweets by Taco Cohen

Sure. But via what channel would information from our evolutionary past affect the brain of a newborn, aside from DNA? Information transmission via chemical signaling or neural connection to the mother during gestation? I dont see a plausible high bandwidth channel

Shared by Taco Cohen at 5/14/2022

Current causality frameworks & algos generally aren't. They typically require expert input about which variables to use, which causal relations might be present, etc. That's fine for applications in science, but not for autonomously learning AI agents. 15/n

Shared by Taco Cohen at 5/14/2022

Furthermore, many algorithms for e.g. causal discovery / graph learning rely on conditional independence testing and have terrible scaling behaviour (computationally and statistically) 16/n

Shared by Taco Cohen at 5/14/2022

I'm currently exploring the hypothesis that something that could be called causality might help us address some of these limitations 12/n

Shared by Taco Cohen at 5/14/2022

More generally, it seems like the current DL approach is good at absorbing tons of information from data, but not that great at extracting "the essence". 10/n

Shared by Taco Cohen at 5/14/2022

By essence I mean something that takes few bits to describe, but explains a lot, and generalizes very far even if it cannot always be used to predict every detail. Grammars, laws of physics, causal mechanisms, symmetries, etc. 11/n

Shared by Taco Cohen at 5/14/2022

One could argue that we can find that 350mb of prior+algo by meta learning or something, and again it might just work. But so far I have not seen any overwhelming successes from e.g. neural architecture search or learned optimizers. 8/n

Shared by Taco Cohen at 5/14/2022

Everyone still uses handcrafted transformer/CNN architectures and Adam, so these learned architectures and optimizers can't be that much better, if they are indeed better at all. 9/n

Shared by Taco Cohen at 5/14/2022

The information that evolution has equipped us with can be encoded in ~350mb of DNA, whereas pre-trained nets are far far bigger these days. Weights can be compressed, but not by many orders of magnitude. 5/n

Shared by Taco Cohen at 5/14/2022

So it seems that what evolution has equipped us with is some very general prior knowledge and learning algorithms, that work robustly in a huge range of environments 6/n

Shared by Taco Cohen at 5/14/2022

That thing is getting to a high level of competence in a new or modified domain *quickly*, using relatively little (or sometimes no) labelled data/human-produced text/real-world interactive experience 2/n

Shared by Taco Cohen at 5/14/2022

People will point to transfer capabilities as evidence that this ability can come from scaling, and indeed it might just work. So far at least it has worked better than most people believed just a few years ago. Let's try and see. 3/n

Shared by Taco Cohen at 5/14/2022

Usually this happens at the emeritus prof level, but I’m trying to speed up my career. In all seriousness though, if we want to learn causal representations/variables then we better know what they are. We need a crisp mathematical definition before we can start the engineering

Shared by Taco Cohen at 5/11/2022

Neuroscience is hard because brain state is hard to measure and intervene upon. Figuring out principles of intelligence by building AI seems easier.

Shared by Taco Cohen at 4/30/2022

8 years of progress in generative modelling. What a time to be alive

Shared by Taco Cohen at 4/30/2022

After LLMs, the next big thing will be LCPs: Large Control Policies. Very general pretrained goal-conditioned policies for embodied agents. If you provide it with a goal vector / example / text, it can do a large number of tasks in a large number of environments. Then we retire🤖

Shared by Taco Cohen at 4/14/2022

Nice example of theory informing practice: Tune hyperparams on a small model using muParameterization, transfer them to a large model without further tuning. Big deal if it works as advertised.

Shared by Taco Cohen at 3/8/2022

Interested in generative modelling and image/video/audio compression? Qualcomm AI Research is hiring researchers in this exciting area in Amsterdam and San Diego! https://t.co/V1rNzwMhHZ

Shared by Taco Cohen at 8/17/2021

Like our NeurIPS paper (https://t.co/zLpKPpCOuw) we use the language of fiber bundles, which turns out to be a perfect fit for CNN feature spaces on homogeneous spaces as well as general manifolds and other spaces.

Shared by Taco Cohen at 5/28/2021