Re: Are Emergent and Leabra Stable?
Randall O'Reilly
Unfortunately things are not quite stable yet. The emergent readme summarizes the situation:
https://github.com/emer/emergent
IMPORTANT branch information:
• main is the latest version based on Cogent Lab, used by axon, and is still undergoing significant development.
• leabra is an awkward in-between version that works with the current leabra and Computational Cognitive Neuroscience simulations, which had to be updated for teaching in Fall 2024, so the simulations could run on the web browser.
I had to update leabra and all the sims before several major changes in the overall infrastructure. The plan is to develop everything around Cogent Lab: https://github.com/cogentcore/lab which incorporated all the etensor and etable stuff that used to be in emergent. Axon is updated to use lab, but leabra is not, and it would unfortunately require another fairly major rewrite of any given sim to get it to work with lab.
Overall, I will be working on an update to the CCN content based on axon, rather than leabra, and will circle back to update leabra much later after all the dust settles.
The plan is to make a "neuro wikipedia" with an extensively hyperlinked document repository based on the new content framework in core (soon to be merged into the main branch), with interactive running examples interspersed among the documentation. All of this can be run with much better performance through a web browser, relative to the current leabra CCN sims, because axon uses WebGPU for compute and graphics, whereas Leabra uses CPU for compute, which is very slow in the browser.
So, while you could use Leabra for new projects, I would rather recommend using axon. I am working on trying to stabilize the basic algorithmic level of axon now, and while there will be updates based on changes in lab and core, these will likely be more of the search-and-replace variety.
Thus, I would recommend looking at the ra25 example in axon at this point. There are a lot of other examples in the sims directory and most of them have been updated and are working well..
Sorry for all the disruption. I really hope that in a couple of months, things will be much stabler. I am already much, much happier with the way sims work in the new framework represented in axon, so I think it will all be worth it in the end..
- Randy
> On Jan 20, 2025, at 6:51 PM, Stephen Read <re...@usc.edu> wrote:
>
> I’m thinking of doing some video tutorials this semester as well as doing several modeling projects this semester and I wanted to get a sense of how active any programming is on emergent, leabra and the core elements for the gui?
> Thanks, Steve
> Stephen J. Read
> Mendel B. Silberberg Professor of Social Psychology
> Department of Psychology
> University of Southern California
> Los Angeles, CA 90089-1061
> Website: www.stephenjread.com
Stephen Read
Hmm…So are the libraries for the current class sims and the current leabra ra25 going to be stable at least for the next 4-6 months? My current situation is that I have several ongoing projects that we already spent time getting to work with the current structure for the sims and we want to spend this semester working on further developing them. We don’t have anything planned to start from scratch, at least for a while. We just want to test and develop what we currently have. So can I safely stay with the current class sims framework, at least for a while?
And it sounds like starting something new with axon should wait at least a little bit.
It sounds like everything will be a great improvement when it is done, but for right now I just need a predictable environment for my current projects.
One other question. Is axon going to show any great increase in computation speed compared to classic leabra? When I talk to people about using leabra/axon for building cognitive science models, they are always concerned with how slow it is compared to something like PyTorch, even though they are attracted by the greater neurobiological plausibility of leabra/axon.
Best, Steve
Stephen J. Read
Mendel B. Silberberg Professor of Social Psychology
Department of Psychology
University of Southern California
Los Angeles, CA 90089-1061
Website: www.stephenjread.com
From:
Randall O'Reilly <rcore...@gmail.com>
Date: Monday, January 20, 2025 at 7:57 PM
To: Stephen Read <re...@usc.edu>
Cc: O'Reilly Kai <koreil...@gmail.com>, emergent-users <emergen...@googlegroups.com>
Subject: Re: Are Emergent and Leabra Stable?
Steve (CC'ing emergent-users), Unfortunately things are not quite stable yet. The emergent readme summarizes the situation: https: //urldefense. com/v3/__https: //github. com/emer/emergent__;!!LIr3w8kk_Xxm!otoBlNgA3ks2pe6YrH41Sx8do4ZYqTP500PRDmg_FsWbQJU-zQ2U-WHqhB7NIQKkEbpTU1rHwmlx-SU$
Steve (CC'ing emergent-users),
Unfortunately things are not quite stable yet. The emergent readme summarizes the situation:
IMPORTANT branch information:
• main is the latest version based on Cogent Lab, used by axon, and is still undergoing significant development.
• leabra is an awkward in-between version that works with the current leabra and Computational Cognitive Neuroscience simulations, which had to be updated for teaching in Fall 2024, so the simulations could run on the web browser.
I had to update leabra and all the sims before several major changes in the overall infrastructure. The plan is to develop everything around Cogent Lab: https://urldefense.com/v3/__https://github.com/cogentcore/lab__;!!LIr3w8kk_Xxm!otoBlNgA3ks2pe6YrH41Sx8do4ZYqTP500PRDmg_FsWbQJU-zQ2U-WHqhB7NIQKkEbpTU1rHf6iD9jc$ which incorporated all the etensor and etable stuff that used to be in emergent. Axon is updated to use lab, but leabra is not, and it would unfortunately require another fairly major rewrite of any given sim to get it to work with lab.
Overall, I will be working on an update to the CCN content based on axon, rather than leabra, and will circle back to update leabra much later after all the dust settles.
The plan is to make a "neuro wikipedia" with an extensively hyperlinked document repository based on the new content framework in core (soon to be merged into the main branch), with interactive running examples interspersed among the documentation. All of this can be run with much better performance through a web browser, relative to the current leabra CCN sims, because axon uses WebGPU for compute and graphics, whereas Leabra uses CPU for compute, which is very slow in the browser.
So, while you could use Leabra for new projects, I would rather recommend using axon. I am working on trying to stabilize the basic algorithmic level of axon now, and while there will be updates based on changes in lab and core, these will likely be more of the search-and-replace variety.
Thus, I would recommend looking at the ra25 example in axon at this point. There are a lot of other examples in the sims directory and most of them have been updated and are working well..
Sorry for all the disruption. I really hope that in a couple of months, things will be much stabler. I am already much, much happier with the way sims work in the new framework represented in axon, so I think it will all be worth it in the end..
- Randy
> On Jan 20, 2025, at 6:51 PM, Stephen Read <re...@usc.edu> wrote:
>
> I’m thinking of doing some video tutorials this semester as well as doing several modeling projects this semester and I wanted to get a sense of how active any programming is on emergent, leabra and the core elements for the gui?
> Thanks, Steve
> Stephen J. Read
> Mendel B. Silberberg Professor of Social Psychology
> Department of Psychology
> University of Southern California
> Los Angeles, CA 90089-1061
Randall O'Reilly
>
> Hmm…So are the libraries for the current class sims and the current leabra ra25 going to be stable at least for the next 4-6 months? My current situation is that I have several ongoing projects that we already spent time getting to work with the current structure for the sims and we want to spend this semester working on further developing them. We don’t have anything planned to start from scratch, at least for a while. We just want to test and develop what we currently have. So can I safely stay with the current class sims framework, at least for a while?
> And it sounds like starting something new with axon should wait at least a little bit.
> One other question. Is axon going to show any great increase in computation speed compared to classic leabra? When I talk to people about using leabra/axon for building cognitive science models, they are always concerned with how slow it is compared to something like PyTorch, even though they are attracted by the greater neurobiological plausibility of leabra/axon.
That said, there are significant costs associated with the neurobiological nature of the models: it takes 200 updates (cycles) for every trial in axon (1 cycle = 1 msec, 200 = one theta cycle = 5 hz). That means the models are doing 200x the amount of computation per trial relative to a feedforward backprop net forward pass. In addition, backpropagation happens via actual bidirectional connections, so you have roughly 2x the amount of connectivity. And ML models pack a huge amount of information into the linear activations of ReLU units, so they can generally use many fewer units. My hypothesis is that all of this is essential for consciousness and self-aware thinking, but we'll see.. :)
- Randy
Stephen Read
Thanks for all the responses.
I definitely agree that the bi-directional, highly dynamic nature of axon is going to be critical to capture the nature of human thinking and consciousness. The brain as a dynamical system, and the role of constraint satisfaction processes are all critical. Not as convinced that the close correspondence to the biological mechanisms is going to be critical, but we will see.
Best, Steve
Stephen J. Read
Mendel B. Silberberg Professor of Social Psychology
Department of Psychology
University of Southern California
Los Angeles, CA 90089-1061
Website: www.stephenjread.com
From:
Randall O'Reilly <rcore...@gmail.com>
Date: Tuesday, January 21, 2025 at 10:22 AM
To: Stephen Read <re...@usc.edu>
Cc: O'Reilly Kai <koreil...@gmail.com>, emergent-users <emergen...@googlegroups.com>
Subject: Re: Are Emergent and Leabra Stable?
On Jan 20, 2025, at 9: 08 PM, Stephen Read <read@ usc. edu> wrote: > > Hmm…So are the libraries for the current class sims and the current leabra ra25 going to be stable at least for the next 4-6 months? My current situation is that