Sai Prasanna

Sai Prasanna

@saiprasanna.in

See(k)ing the surreal Causal World Models for Curious Robots @ University of Tübingen/Max Planck Institute for Intelligent Systems 🇩🇪 #reinforcementlearning #robotics #causality #meditation #vegan

2.1K followers689 following287 posts

Sai Prasanna
@saiprasanna.in

If open-endedness has to be fundamentally subjectively measured, what are the factors of the agent makes it so if we fix humans as the final arbiter or evaluator. Does embodiment/action space etc of the agent matter for a human evaluator of open-endedness?

View full thread

Sai Prasanna reposted

Tom Silver
@tomssilver.bsky.social

This week's #PaperILike is "Model Predictive Control and Reinforcement Learning: A Unified Framework Based on Dynamic Programming" (Bertsekas 2024). If you know 1 of {RL, controls} and want to understand the other, this is a good starting point. PDF: arxiv.org/abs/2406.00592

arxiv.orgModel Predictive Control and Reinforcement Learning: A Unified Framework Based on Dynamic ProgrammingIn this paper we describe a new conceptual framework that connects approximate Dynamic Programming (DP), Model Predictive Control (MPC), and Reinforcement Learning (RL). This framework centers around ...
Sai Prasanna
@saiprasanna.in

I realized how I background process tonnes of information, from work/research and emotional stuff. And it works well, leads to good research ideas, wise processing of tough situations! But It's so hard to learn to trust this as conscious thinking for solving problems feels more under my "control"

Sai Prasanna
@saiprasanna.in

One strategy I guess is to have good stream of good (BS filter) and diverse (topics, areas) inputs (books, research papers, what not) And not get bogged by the fact that I am too distracted to go deep into one input stream (book or podcast or article or paper) at a time

Sai Prasanna
@saiprasanna.in

I realized how I background process tonnes of information, from work/research and emotional stuff. And it works well, leads to good research ideas, wise processing of tough situations! But It's so hard to learn to trust this as conscious thinking for solving problems feels more under my "control"

Sai Prasanna
@saiprasanna.in

TIL: "Clever Hans cheat" for next-token prediction. A subtle but interesting issue with next-token prediction. In the purely forward next token prediction objective, teacher forcing can lead to learning dynamics where the models don't even generalize "in-distribution"!! arxiv.org/abs/2403.06963

arxiv.orgThe pitfalls of next-token predictionCan a mere next-token predictor faithfully model human intelligence? We crystallize this emerging concern and correct popular misconceptions surrounding it, and advocate a simple multi-token objective...
View full thread
Sai Prasanna
@saiprasanna.in

To re-think I think the problems could be orthogonal. Clever hans pertains to teacher forcing during training leading to easy solutions for lot of the timesteps skewing it to not learning the hard timestep which is most important for test-time.

Sai Prasanna
@saiprasanna.in

Conditioning gap in latent space world models is due to how uncertainty can go into latent posterior distribution or the learnt prior (dynamics model) and not conditioning on the future would put the uncertainty incorrectly into dynamics model.

Sai Prasanna
@saiprasanna.in

TIL: "Clever Hans cheat" for next-token prediction. A subtle but interesting issue with next-token prediction. In the purely forward next token prediction objective, teacher forcing can lead to learning dynamics where the models don't even generalize "in-distribution"!! arxiv.org/abs/2403.06963

arxiv.orgThe pitfalls of next-token predictionCan a mere next-token predictor faithfully model human intelligence? We crystallize this emerging concern and correct popular misconceptions surrounding it, and advocate a simple multi-token objective...
View full thread
Sai Prasanna
@saiprasanna.in

TIL: "Clever Hans cheat" for next-token prediction. A subtle but interesting issue with next-token prediction. In the purely forward next token prediction objective, teacher forcing can lead to learning dynamics where the models don't even generalize "in-distribution"!! arxiv.org/abs/2403.06963

arxiv.orgThe pitfalls of next-token predictionCan a mere next-token predictor faithfully model human intelligence? We crystallize this emerging concern and correct popular misconceptions surrounding it, and advocate a simple multi-token objective...
View full thread
Sai Prasanna
@saiprasanna.in

It reminds me of another problem, and I'm not sure if it's equivalent or if it's some dual problem. It's called the conditioning gap in latent space inference.

Sai Prasanna
@saiprasanna.in

Conditioning gap: When you train a value encoder that computes an approximate posterior that's conditioned partially (say on past tokens), then the posterior has a worse lower bound than one also conditioned on everything (also future tokens).

Sai Prasanna
@saiprasanna.in

TIL: "Clever Hans cheat" for next-token prediction. A subtle but interesting issue with next-token prediction. In the purely forward next token prediction objective, teacher forcing can lead to learning dynamics where the models don't even generalize "in-distribution"!! arxiv.org/abs/2403.06963

arxiv.orgThe pitfalls of next-token predictionCan a mere next-token predictor faithfully model human intelligence? We crystallize this emerging concern and correct popular misconceptions surrounding it, and advocate a simple multi-token objective...
View full thread
Sai Prasanna
@saiprasanna.in

And there is a new paper that claims to fix this for transformer architecture!!! They call it "belief state transformer". Apparently it fixes lots of practical problems arising due to clever hans cheat! arxiv.org/abs/2410.23506

arxiv.orgThe Belief State TransformerWe introduce the "Belief State Transformer", a next-token predictor that takes both a prefix and suffix as inputs, with a novel objective of predicting both the next token for the prefix and the previ...
Sai Prasanna
@saiprasanna.in

The fix involves modelling forward and backward directions. I haven't grokked it fully, but I learnt about the above problem there. I find this two papers a really nice sequence of a fundamental problem and then a solution!

Sai Prasanna
@saiprasanna.in

TIL: "Clever Hans cheat" for next-token prediction. A subtle but interesting issue with next-token prediction. In the purely forward next token prediction objective, teacher forcing can lead to learning dynamics where the models don't even generalize "in-distribution"!! arxiv.org/abs/2403.06963

arxiv.orgThe pitfalls of next-token predictionCan a mere next-token predictor faithfully model human intelligence? We crystallize this emerging concern and correct popular misconceptions surrounding it, and advocate a simple multi-token objective...
View full thread
Sai Prasanna
@saiprasanna.in

But teacher forcing makes it easy to predict all tokens after the first branching token by paying attention only to previous token and remembering or attending to the edge with this. This strategy doesn't work for the first token where there are the start branches

Sai Prasanna
@saiprasanna.in

Since teacher forcing makes the model learn easy cheat for most easy tokens, the learning dynamics make it hard to find the correct strategy for the first token.

Sai Prasanna
@saiprasanna.in

TIL: "Clever Hans cheat" for next-token prediction. A subtle but interesting issue with next-token prediction. In the purely forward next token prediction objective, teacher forcing can lead to learning dynamics where the models don't even generalize "in-distribution"!! arxiv.org/abs/2403.06963

arxiv.orgThe pitfalls of next-token predictionCan a mere next-token predictor faithfully model human intelligence? We crystallize this emerging concern and correct popular misconceptions surrounding it, and advocate a simple multi-token objective...
View full thread
Sai Prasanna
@saiprasanna.in

Imagine a task where you give a list of edges of a star graph, start and end node, and train a model with a teacher forcing you to predict the list of tokens in the path from the start to the end. (edge 1, edge 2 ...) (start, goal) (start, intermediate1, intermediate 2 .. .goal)

Sai Prasanna
@saiprasanna.in

The easiest coorect solution for the model is to look at the edge with the goal (since it's star graph there is only one edge) and work the way backwards to the start (in it's computation) and output the path one by one forward.

Sai Prasanna
@saiprasanna.in

TIL: "Clever Hans cheat" for next-token prediction. A subtle but interesting issue with next-token prediction. In the purely forward next token prediction objective, teacher forcing can lead to learning dynamics where the models don't even generalize "in-distribution"!! arxiv.org/abs/2403.06963

arxiv.orgThe pitfalls of next-token predictionCan a mere next-token predictor faithfully model human intelligence? We crystallize this emerging concern and correct popular misconceptions surrounding it, and advocate a simple multi-token objective...
Sai Prasanna
@saiprasanna.in

This failure occurs in distribution, not OOD. And it apparently is general for any model learning next-token prediction regardless of recurrence (linear or otherwise) or attention!!!

Sai Prasanna
@saiprasanna.in

If I have a really good photo that could be potentially used in many contexts, what's the best place to make money with it? My friend has a really good eye for photos and we want to try a side venture selling some of her stuff

Sai Prasanna reposted

Ronen Tamari
@ronentk.me

Looks like a cool study. Lots to learn from ants about large scale coordination www.pnas.org/doi/10.1073/... "Our results exemplify how simple minds can easily enjoy scalability while complex brains require extensive communication to cooperate efficiently." h/t @petersuber.bsky.social

pnas.orgComparing cooperative geometric puzzle solving in ants versus humans | PNASBiological ensembles use collective intelligence to tackle challenges together, but suboptimal coordination can undermine the effectiveness of grou...
Sai Prasanna
@saiprasanna.in

Does augmenting ourselves with V/LLMs to cognitive gaps make self actualization even more difficult on average? Stands stark in contrast with (more difficult/slower to show positive outcomr) augmentation strategies like meditation or psychedelics

View full thread
Sai Prasanna
@saiprasanna.in

Mass effect 3 ending "choices" don't feel so bad or unrealistic seen in this light of ultimate end of augmenting ourselves with sophisticated tools and gradually tools that seem more like us in some ways. I picked the merge.

Sai Prasanna
@saiprasanna.in

Does augmenting ourselves with V/LLMs to cognitive gaps make self actualization even more difficult on average? Stands stark in contrast with (more difficult/slower to show positive outcomr) augmentation strategies like meditation or psychedelics

View full thread
Sai Prasanna
@saiprasanna.in

Maybe its not either/or, it depends on the level at which these things can be customised? But the gap between direct neuro response augmentation/shift and indirect ways feel unsurmountable, atleast in the AR glasses type interface. Maybe direct neuro augmentation devices in the future changes it

Sai Prasanna
@saiprasanna.in

Does augmenting ourselves with V/LLMs to cognitive gaps make self actualization even more difficult on average? Stands stark in contrast with (more difficult/slower to show positive outcomr) augmentation strategies like meditation or psychedelics

Sai Prasanna
@saiprasanna.in

For example, for people whom it takes effort and energy to read facial expressions and empathise emotionally instead of cognitively, one can see AR glasses with VLM support easily fixing the baseline ability. But it comes at a cost of further letting direct emotional sensitivity degrade

Sai Prasanna
@saiprasanna.in

Perhaps a really strong benchmark for generalization for RL agents/foundation models trained with intrinsic motivation (game has no explicit rewards) on "normal" geometry/physics.