There is a proposed experiment to test for fan effects based on typicality effects.
Experiment tests Fisher and Langley theory Experiment tests for ICL based fan effects and IWL fan effects Unifies a cognitive effect across learning types in LLMs
We discussed potential ideas for how to measure fan effects in language models. I proposed that we vary the number of category members given in the context to a language model and interrogate the presence of the elements.
At some number of context members, the probability of absence given actual presence will tend toward the probability of absence given actual absence. That is…
\[\begin{aligned} p(absent|present) \approx p(absent|absent) \end{aligned}\]When this occurs, if the probability of correctly evaluating the presence of atypical category members is higher than for typical category members, this would be a language model specific variant of the fan effect.
So, GPT-2 can be used with a population and a number of category items that are understood in terms of their typicality. The context can be filled to varying degrees (experimental variable 1). When the above probabilitic scenario occurs for the cohort of present typical items, the probability of presence of the atypical within context cohor should be measured. If the accuracy is higher for atypical than for typical when the context has reached the described probabilistic scenario, then GPT-2 is more readily able to retrieve atypical items within the context than typical items. Thus, a form of fan effects are present.
Thinking about the theoretical idea of computational criticism and synthesizing the desiderata from the fisher and shin paper with the framework from stiny and gips has resulted in a new understanding of the important facets of computational criticism. Specifically, it has become clear that many of the desiderata are isomorphic restatements of larger problems in AI and cognitive science. As an example, fisher and shin say that a critic should possess an understanding of the medium specific formal characteristics. This should be interpreted to mean that the critical process which yields translation from artwork to latent description should be sensitive to a subset of the medium specific formal characteristics. The more general problem in AI is known as feature selection. Given a set of features, which of these features is important. In cognitive science the question is how does attention cause the features to be extracted from the background? A similar problem in computational criticism is the need to exclude information which is irrelevant. This is important from the standpoint of expertise in cognitive science as those who are experts tend to not be distracted by irrelevant information. This is interesting because it seems that computational criticism inherits from both AI and cognitive science in important ways. This should be the case as criticism is a behavior that arises from intelligent behavior.
It’s interesting to consider that creators and critics exist in a symbiotic relationship with one begging the existence of the other. If a musician releases an album, opinions will be developed regarding their music. On the other hand, if a specific taste seems to emerge from a society, a creator will quickly fill the need for further creation. Often the creator will act as the critic as well. Such is the case for most human artists, thus the cliche of being one’s own biggest critic. This leads to a bit of a chicken or egg question. In the creative ecosystem who is responsible for the impetus that leads to new creations? Creator or critic? The obvious answer may be to assume that the creator was to blame. However, consider that if Bach had possessed a different critical bias then he would have likely created something else to suit his taste. Therefore, it seems that a creator possesses a critical bias based on there internal reward circuitry which then guides their search for their next creation. So, the critic comes before the creation. This leads to questions about the possibility of using the creator critic pair to simulate emergent creativity. By allowing the critic to evolve, the creator will begin to evolve to meet the expectations of the critic thus resulting in emergent, novel creations.
From above, it seems that a creator is actually made of a creative component and critical component with the critical arising from the reward giving limbic system. However, the creator does not produce a written criticism. This does not make the critical process in the creator any less real. This suggests that the written criticism is not an intrinsic process to general criticism. Below is an alternative model of creativity and criticism which has the limbic/reward circuitry in yellow, the perception/state-of-the-world in green, the analytical/synthetic (search) portion of thought is in red, and finally the action is taken in blue. This connects well with current theories regarding reinforcement learning.
Consider that there is no environmental stimulus such that the existence of a critic is required. Creativity and criticism arise out of human nature naturally. Then a model of creativity should be a model which begs the existence of a critic. Or rather more likely, the existence of a critic begs the existence of a creator. ~ The model of a creator above would also explain the emergence of a formal critic as one that takes in another’s work as their initial condition. In this way a critic proper should be understood to be themselves, creators. What makes critics unique is that through abundant exposure to many works of art, their reward centers are among the most well informed in the society. Reward is given based on predictability and surprise, so they are the hardest to surprise and are the most aware of new developments on which an artist may draw. They are therefore, most well situated to appreciate the work of a contemporary artist. They are also expected produce a derivative work of art themselves in the form of a written criticism. Therefore, the model above may be a suitable framework with which to model agents to develop emergent creativity and criticism. ~ In reality there also exists an audience not only creators and critics. However, if all people (and thus all agents in this ecosystem) are fundamentally similar in their cognitive build, why are some gifted to be creators or critics while others are not? Certainly every person has an imagination which can conceive of great beauty. This should be understood to be a situation in which the agent above is able to supply a description to the effector in the model above which is beautiful. Those who are gifted to be creators are able to take the description in their mind and take actions through the effector which generate an artwork. When that artwork is observed by the creator’s receptor or by the other agent’s receptors they are rewarded either egocentrically or exocentrically. An agent which attempts to create but does not experience any reward is not likely to continue to create. These agents that are not rewarded learn to consume thus becoming part of the audience. When the audiences learns to consume the work of creators they are learning to reward the creator. In this way, those which are capable of producing egocentrically or exocentrically rewarding creations become creators. Those which can’t become the audience. Interestingly this also predicts that those which will be critics did not learn to stop creating suggesting that they themselves were most likely mildly successful at creating.
In the beginning stages of the society, creators must create for an egocentric reward, however as time goes on the development of an audience could provide an exocentric reward which is large enough to overwhelm the egocentric. These creators which become devoted to the exocentric reward may become fan service which fail to subvert their audiences expectations and rather only succeed in meeting them. Meeting the audiences expectations spells death to a creator as the primary reward delivered to the audience comes from them being able to predict most of the creation but not being able to predict it fully which ultimately results in the powerful better than expected reward prediction error.
A reinforcement learning model built from this with agents that have different creative processes and different tastes based on past experience and predilection for surprise should offer interesting insights into how the various pieces interlock to bring about the various actors in the society (creator, critic, and audience). As well as why some creators who develop exocentric reward centers fail to continue to reward the audience and why creators who focus on the egocentric reward may die in obscurity as Van Gogh did.
Targeting BabyLM 2024 with a cobweb based implementation. The interface definition for the implementation is housed here.
I was thinking about a previous discussion on forgetting and hit upon a notion. Infinite memory with limited retrieval ability should be indistinguishable from limited memory. So, instance based systems with limited retrieval should exhibit something indistinguishable from forgetting. This would seem to apply both to cobweb and to in-context-learning.
The papers title will be: Forgetting Beyond Neural Networks
Show memory-based forgetting and retrieval based forgetting in both the neural and non-neural cases.
Good knowledge with appropriate retrieval algorithm with unexpected/non-optimal stopping (possibly from limited resources). Leads to catastrophic retrieval which may be indistinguishable from forgetting. Gives an account of the link between ADHD and risk taking behavior.
]]>From the Program Director, it seems we need to ensure that we connect with a pressing need in the DEB portion of biology. Probably the ES side. On the computer science of the proposal, it needs to be clear how the virtuous cycle will play out. Specificaly, how will the data force us to innovate in cs?
Satellite images are taken on the scale of days, creating a vast amount data. By merging that data with other GIS data (weather, topography, gps, and existing land use labels), we should be able to learn to do next frame prediction including how a given area of land changes given precipitation and nearby snow melt (pre-training task 1). In order to identify land that will be attractive in the coming season to animal populations, it’s necessary to consider the present land state (to estimate snow melt induced water levels or calorie content) as well as the predicted future state. As the temporal distance to the season in question decreases, the fidelity of the model’s prediction should increase. Such a model must be able to perform predictions at an arbitrary time scale.
One method to achieve this would be to perform continual next frame prediction until the season in question is reached. However, this is limited by the temporal context length possible in the model. Further, iterative single step predictions has not been a successful method for predictions at an arbitrary distance in other domains.
This creates an interesting and novel problem. The season in question is at a fixed point in the future but the temporal distance between the current frame and that point is a moving target. We believe a model can be trained to predict what a frame may look like on a given date based on the context provided without considering the temporal distance explicitly (pre-training task 2). As already mentioned, this should encourage the model to learn features that affect long term seasonal conditions whereas next frame prediction allows the model to learn to exploit short-term semantic connections.
Such a model will learn a representation that is appropriate for near and far term predictions (learning interesting semantic connections between terrain, region, weather, climate, and land use). This is interesting and innovative from a foundation model perspective and is necessary to understand and animal habitat preferences.
A model of this variety that merges large scale GIS data for prediction at arbitrarily prompted dates is not only of intellectual merit. This model is likely to have applicability to flood plain prediction, excavation/development effect prediction across multiple seasons (through counterfactual analysis), and use in climate change effect prediction.
]]>The proof sketch is fairly simple and elegant. However, it warrants more consideration. A universal function approximator should be necessarily able to approximate any feedforward neural network, including an arbitrary RNN for a given time step and hidden state. The UFA can be recursively given access to the previous output and may be given the same input which would have been provided to the RNN. Therefore, it seems that any architecture that is capable of universal function approximation, when coupled with recursion, should be Turing complete.
]]>next sentence masked token next token prediction
Image masking
CLIP - association of caption with image FLIP - association of caption with image with simultaneous image patch masking
Masked frame prediction
Successful pre-training tasks require models to exploit semantically related information to succeed. Further, the data typically exceeds simple label associations and tends toward high semantic complexity.
]]>We discussed multiple general ideas:
In some fields it seems a forgone conclusion that rhetoric can be faithfully translated to another language. This presupposes that the rhetoric is not dependent upon the language. If this is the case, then are there some rhetorical notions that cannot be translated between some languages? If all rhetorical notions can be translated between all languages, what is the representation of rhetoric beyond language?
An interesting related point is the fact that there are no words for the idea that physicists had about the nature of particles. This lead to a folk misunderstanding of particle-wave dualism. The meaning was, if things with particle-like momentum have wave-like interference then things understood as waves may have momentum. The resultant discovery was that each was actually an expression of a particle but that particles were more complicated than originally understood. However, this idea is not well translated into language and the folk understanding is often that light can be a particle or a wave.
]]>I acquired these pieces of advice from wonderful researchers who have shaped my philosophy of research.
Researchers often tilt at windmills. They will perform hundreds of experiments and gather tremendous data with no particular purpose. This is not science.
Start with the claims that you wish to make (this is your hypothesis). Then, devise a set of experiments that, if they go as you anticipate, will support the claims. This is the scientific method.
The point of a paper needs to be unmistakable. Often, a paper’s contributions will not be explicitly stated. Scientific writing is a form of rhetoric. To get a point across, you need to be consistent and careful in the claims you make so that others can understand how to learn from and build onto it.
STEM research can be very technical and dense. Frankly, it’s not terribly likely that even other researchers will read your work thoroughly unless the paper is of tremendous import. So, make the narrative easy to follow and tell the story in pictures.
This quote is not originally from Pat. But Pat repurposed it. His point was, done is better than perfect. This is tremendously important when performing research on a deadline. Get it done.
]]>The differing life experiences and personality may have led to the creation of a differently expressed but potentially ontologically same law. However, some notions of inspiration would suggest that the law would have been wholly unchanged.
If the law would have been changed in expression, then the nature of inspiration is not necessarily verbatim and verbal. There’s good reason to hold this view since some authors in the NT cover the same or related topics, yet they are different along vocabulary choices and even expressed ideas.
Alternatively, if inspiration is verbatim from God, then the entire accepted cannon should possess a cohesive rhetoric. As an example, John reports that Jesus said the work to gain heaven was to believe on the one God has sent, but Paul says that heaven can’t be gained by works so that no person can boast. The reality is, two different ideas are being conveyed using almost the same language. Rhetorically, this clearly results in confusion. However, assuming authorial agency beyond authorial intent resolves this. Each author is placed in a context to learn and develop, a situation to necessitate writing, and has experienced some divine intervention. These factors effect the decisions made by the author but leave them free to act as they see fit if they retain agency.
Should Biblical authors be seen as possessing agency? If they do not possess agency, then the text must also be considered to be absent of authorial intent likewise.
]]>