My old account (star0) is inaccessible to me, so I created a new one. Yuli-Ban asked me to come here and write something, so I have. Yuli had written some of what I am posting here once before, I see -- but my posting now provides a little more context:
So, I've talked about various paths to "true AI" in the past, for example using video-prediction as a route to common sense reasoning -- and that is progressing very rapidly. But I think that this will not go the whole way towards true AI. About a year ago it occurred to me that there is a way to get a lot closer, and that it seems inevitable that this is how the march towards AI will play out. I am not here to tell you that I will build it; my role is simply one of prediction. You may find the prediction strange, as when I predicted a few years ago on this very forum that many big box stores and malls would close before 2025, to be replaced by online shopping. That prediction was met with comments like, "I promise you, there will still be big box stores!" -- misunderstanding what I said ("many" not "all"), but also seemingly disbelieving that online shopping could have such a large impact.
Now hear me out about my latest prediction: I predict that with the arrival of good, inexpensive, mass-market, non-invasive Brain-Computer Interfaces, A.I. will advance incredibly rapidly. And, furthermore, there are multiple different groups working to produce such BCIs in the next 2 years or so, including Mary Lou Jepsen's group at Open Water, and also Facebook's Building 8 group led by Mark Chevillet. Here is one from OBELAB, a KAIST spin-off, that does FMRI-resolution scans from the surface of the Prefrontal Cortex using FNIRS:
(Probably a scanner like that for more brain surface regions would be all you need for the idea below to work!)
How, you might wonder?
Well, I've written a technical discussion of how it could happen, which you can find here:
complete with potential criticisms and rebuttals.
I've written other drafts of this over the past year, and then earlier this year I heard about an approach by some neuroscientists who came up with an even more basic proposal, that already seems to be producing good results -- a New Scientist article about which can be found here:
Where I think this is headed is the following: soon after the arrival of good, non-invasive BCIs with high spatial and temporal resolution -- perhaps even just for the surface of the brain -- research groups will start collecting large amounts of passively-generated data as human subjects perform everyday tasks. Using Machine Learning, it will be possible to fuse this brain-state information with the recorded stimuli, and produce A.I. systems that generate their own synthetic brain states as they are presented with stimuli.
These brain states are not firings of individual neurons, but rather the averaged behavior of many neurons within a population -- they are "population responses". That may sound like you'd be leaving a lot of important details about the brain out, but people have been able to at least partially read off short-term working memory details, geolocation information, emotional state, language use, and many other things, from just these population values. Furthermore, the right Machine Learning algorithms can fill in some of the gaps when using such crude information about the brain, when given the stimuli and behavioral response (all this is detailed in the above reddit link).
Not long after the arrival of those good BCIs, I foresee the following basic application: a few dozen subjects passively watch a large number of short videos as their brain states are recorded. Using the raw pixels and brain data from all those viewings, a "joint model" will be built using Machine Learning, that when given video input, will produce its own time-varying synthetic brain states, that closely match what you would see if a human viewed those same videos. These wouldn't merely be PREDICTED brain states; but would also be USED to guide future brain-state predictions as the media runs. Once the video the machine is "watching" finishes, the final synthetic brain state will give a representation of what the machine "thought of it".
Sound far-fetched? Well, here is a very crude version of what I just described:
The article describes a method to fuse together human FMRI features as people watch a video (I don't believe it does this over multiple time-steps, though), with features derived from the video, using a Deep Boltzmann Machine neural net. During training time, the video features and FMRI features are both available; but during test time, only the video features are available. The neural net used has the nice property that if one of the modalities is missing at test time, you can apply something called "Gibbs Sampling" to fill in the missing modality. Deep Boltzmann Machines would not be the best method to use if you scale up this work, use pixel inputs for multiple frames, and try to predict and USE a synthetic brain response for multiple time steps; but you can at least see there is at least one path forward on building this type of brain-based A.I.
More intriguing would be to build a chatbot of some kind, using brain data. I have a feeling that if you collected enough data (from maybe 100 test subjects), and used some of the best "generative models" out there today, you could build a chatbot that far and away outperforms the best of the best. I refer to such a hypothetical chatbot as a "Zombie AGI", and conjecture that talking to it would be like chatting with a human with moderate dementia: it would be able to hold a conversation for several exchanges (using short-term and working memory), but would get confused when things got too complicated. It would have common sense reasoning ability and good language understanding. And, most interesting of all, if you were to investigate the synthetic brain states it generates, they would closely resemble what you see in a human test subject!
So when can we expect to see something like this? Let's see: assuming good BCIs are available in about 2 to 3 years, and assuming they are cheap enough to where thousands or millions of people can buy them, I expect it might take another 2 to 3 years to collect the data and experiment with the best brain-prediction models. And so, the first versions of such a chatbot might appear in the academic research literature possibly 5 to 6 years from now!
If you want to see A.I. take off like a rocket, faster than what Deep Learning has produced to date, then focus on the arrival of those BCIs!