Here’s my dupster of thoughts as I read papers and think about AI. I’m not sure how useful this will be to anyone else, but I hope it will be useful to me. It goes without saying but beware innacurate statements; these are just personal notes.

12 March 2021

Switch transformers emphasized the search for new fronteir pushing dimensions. Here’s one:

  • how many agents are focused on the same goal? This probably only applies in multiagent systems, but what I’m proposing is this: use an open-ended intelligence metric for survival. Literally train multiple agent on the same problem, structure their environment so they can communicate and cooperate, and then completely reinitialize weights and training for the poorest performing agents. This may be a temporary ocmputation sacrifice, but the result of training a new agent can find better convergence. inspired from https://arxiv.org/pdf/2103.06769.pdf

19 December 2020

Generality is definitely important - but not the only consideration for sci-fi ‘AGI’. Here are some papers I have been reading over 2020 on AGI. They are referenced by arxiv id (e.g.: 2012.08630 = https://arxiv.org/abs/2012.08630):

  • 2012.09830 :star:
  • 2012.08630 :star:
  • 2012.08564
  • 2012.05208
  • 2011.12860
  • 2011.12750
  • 2011.11805 :star:
  • 2011.11400 :star:
  • 2011.09410 :star:
  • 2011.09294
  • 2011.08827v1
  • 2011.08820v1
  • 2011.07027v1
  • 2009.14810v1 :star:
  • 2009.11243
  • 2009.08497 :star:
  • 2009.01810
  • 2009.01791 :star:
  • 2007.14928
  • 2007.04954
  • 2006.12323
  • 2006.11441
  • 2006.09950
  • 2002.00509
  • 2001.11027
  • 2001.11027
  • 2001.09442
  • 1912.13490 :star:
  • 1909.10863v3
  • 1905.13049
  • 1709.08568
  • 1606.00821

1 November 2020

I propose to use a rainbow melting pot of many motivational psychological techniques in AGI development.

19 October 2020

To make the relationship bidirectional, System I learn policies from System II, while System II learns goals from System I’s interpretation of the agent’s own body and environment.

Rythyms of the Brain credits gamma oscillations largely to the shared GABAa interneuron depolarizaiton refactory time that allows synchrony to be established across region-specific and whole-brain interneuronal networks. Importantly, no particular excitatory synapse is responsible for activating the basket and chandelier interneurons, however there are usually sufficently many dendrites to reach the action potential. Unchecked, self-organizing internal gamma oscillations can spread throughout the brain to bind many cortical areas. Internal synchrony may be interrupted by external regular or irregular stimuli preventing full brain gamma oscillations from being realized. However, regular external interruptions can drive a ‘forced oscillation’. Experiments show that the brief stimulation of a cortical neuron leads to the subject being aware of a touch, but trains of activity are required to ‘feel’ the sensation. This leads to the conclusion that ‘feelings’ are system dynamics that share trajectories with high mutual information across diverse regions of the cortex.

Given both perceptions of the former experiment were percieved, I don’t directly tie consciousness to binding, but it may have a pivotal role. And on ‘binding’: this discussion from Rythyms of the Brain indicates that binding need not be indentified in any one gnostic neuron or feature, but rather, the temporal trajectory of the system. I believe this implies: don’t use embeddings, leave data in sequence form.

18 October 2020

Broaden and build theory posits positive emotions to facilitate exploration and increase long-term outcomes while negative emotions encourage shortsighted optimization. Both play a key role in survival which must consider immediate and distant outcomes. However, it motivates me to consider whether positive emotions additionally optimize a totally different objective altogethor. We’re excluding a spectrum of optimization with the very notion of a loss function or survival-based optimization. There appear to be higher-order objectives of positive emotions. If the reward system primes cortical activity by lowering the action potential threshold, then perhaps negative emotions do the opposite. There is no room for chronic negative emotion in optimal behavior. In my theory of information flow, positive emotions increase conductivity while negative emotions increase resistivity.

The generally weak and transient effect of physical interaction on holistic emotional assessment seems to discredit extrinsic rewards. Perhaps this serves to avert spurrious local optimization in behavior policies. While the brains reward system seems to subconsciously perform decently at emotional regulation, the conscious mind is often poor at imagining behavior-consequent transient affective states. Let System I maintain homeostasis while System II processes optimizes objectives rationally – and make sure emotional objectives don’t creep their way into System II’s objective. The hedonic tredmill can only consciously optimize with known dynamics, but it is the unanticipated, the novel rewards that stimulate our reward system. Modeling the human reward system as an advantage function then should take environmental conditions into account when giving the internal value of an external reward.

Briefly reading psychological literature on emotions convinces me that there are no knobs or raw parameters to emotional state. Though it could be modeled by examining an agent’s cognitive and environmental processes, we can only hope to influence the nondetirministic property.

17 October 2020

Applying energy minimiztion or predictive coding to social communication and behavior, an agent should optimize its behavior to require the least prodding from others. They should also use even imagination to optimize this.

On imagination, module(s) should have the capability to “detatch from reality” - that is, produce output response that are not correlated with the inputs in the module has a sufficently strong belief about the latent variables it transmits that it is willing to ignore ground truth. The effect of this could be imagination if the latent variable loses quarrelation with previous activity.

After lunch today, I will write down my theory of modular information.

16 October 2020

OpenAI’s concept of circuits seems related to my brother’s thought of “extrema” precepts.

[. . .] the self-organized [global] gamma oscillation reflects a top-down cognitive process. [. . .] coherent perception of an object involves the synchronizatino of large cortical areas [. . .] working memory is a hypothetical mechanism that enables us to keep stimuli “in mind” after they are no longer available. The amount of information to be held at any given time is referred to as memory load. [. . .] gamma power increased linearly with memory load at multiple distributes sites, especially above the prefrontal cortex. [. . .] these observations support the more general idea that gamma oscillations are used in the brain for temporally segmenting representations of different items. [. . .] The goal of synchrony for neuronal populations is the same as the goal of action potentials for single cells: forwarding messages to downstream neurons in the most effective manner [. . .] every single time a post-synaptic neuron fire in a manner that the discharge leads to an increase of the free Ca2+ in the dendrites, the prevviously or subsequently active presynaptic onnections are modified. [. . .] the critical window of plasticity corresponds to the length of the gamma cycle. [. . .] Thus, even if gamma oscillation proves to be irrelevant for the binding problem, the oscillation remains a central timing mechanism essential for synaptic plasticity. On the other hand, gamma oscillations may link the problem of binding to plasticity. This is because synchronization by gamma oscillations results in not only perceptual binding but, inevitably, modification of connections among the neurons involved. Synaptic modifications can stailize assemblies representing currently experiences conjustions. In turn, theses use-dependant changes increase the probability that thte same assemblies will be activated upon future presentations of the same stimulus even if the stimulus is somewhat modified in the meantime, The assembly bound togethor by gamma-oscilation-induced synchrony can reconstruct patterns on the basis of partial cue because of the temporally fortified connections among neuron assembly members. (Rythyms of the Brain p.245-247)

What I conclude from this is that the brain behaves like a kohennen self organizing map in that I performs unsupervised classification of events. It also strengthens/weakens connections on the basis of leading or lagging presynaptic firing phase. Then stronger connections or tighter informatin bounds between presynaptic neurons and their postsynaptic target allow fewer cues be necesary to perform the same identification.

Spatial thinkign is unique in the brain. I wonder if there is some general set of “spatial” structures integrating the spatial information from the corpus collosum, retinotopic maps, tonotopic maps, and orientation assemblies.

From listening to Andrew Karpathy’s interview in Andrew Ng’s deep learning specialization, I consider the thought of an online “AGI Zoo” where AGI’s can be started and stopped on demand by paying customers. All AGI’s would be public. Users could place them in their own environments using HTTP protocals or allow the AGI’s to interact in public environments for a lower cost. (less traffic) Some uses would be:

  1. observe their behavior under controlled environments (you can’t really have general intelligence in a controlled environment though).
  2. automate the dirty work such as supervised learning example labeling.
  3. perform AGI psychology. The AGI Lab community would also host contests for users to enter their AGI’s into These ends would all lead to collecting more data on artificial general intelligence safety.

15 October 2020

Talking with my brother, he explained conscious representation as situating a thought between thought “extrema”. These “extrema” are the foremost concrete instantiations of the thought they represent perhaps forming a new dimensional basis for later thoughts to rest on.

After reviewing yesterday’s thoughts, I realize it would be useful to AGI to not only weight evidence for positive but also negative stimulii. The agent propperly trained under a few negative reinforcement trials should never again return to that aversive state to unlearn the association – even if the stimulii is gone. System 2 may ackowledge that there is no danger, but System 1 which provides most of the motivation force ‘knows’ otherwise and remains afraid. (System 1 knows from experience; System 2 knows about) This, in addition to accelerated training under expected positive or negitive can bias unsupervised APD minimization (habituated) learning.

14 October 2020

The dopaminergic reward system not only predicts reward, but also seems to predict proximity to reward. Using the proximity to reward administered by the meso limbic system, the brain gets a sort of mentor in determining which actions to take. That mentor intern bias is learning to take place faster closer to reward, so that exponentially larger amounts of training weight is put close to rewards. Then unsupervised learning favors consistent trajectories which now have a higher weight of going towards the rewards state

It is largely through selective reinforcement of initially random movements, that the behavior of the neonate comes to be both directed at and motivated by appropriate stimuli in the environment [49]. For the most part, one’s motivation is to return to the rewards experienced in the past, and to the cues that mark the way to such rewards. It is primarily through its role in the selective reinforcement of associations between rewards and otherwise neutral stimuli that DA is important for such motivation. Once stimulus-reward associations have been formed, they can remain potent for some time even after the reward has been devalued [. . .] Once a habit has been established, it remains largely autonomous until the conditioned significance of incentive motivational stimuli has been extinguished or devalued through experience. Extinction of the conditioned significance of such stimuli can result from repeated unrewarded trials, repeated trials in the absence of an appropriate drive state, or repeated trials under the influence of neuroleptics (Dopaminergic reward system: a short integrative review)

The olfactory tubercle in the brain serves as a multimodal sensory integration hotspot and terminates some reward pathways.

Thinking Fast and Slow in AI does a good job introducing and distinguishing System 1 and System 2 processes. These might be viewed as unconscious and conscious cognitive processes repsectively. The impetus then is to apply neuronsymbolic approaches to artificial intelligence. It seems that not only should data be more structured (graph-like or linguistic) at higher congnitive processes, but also there should be a reason to generally throw decision making onto a biased, shortcut-taking System 1.

The discussion of System 2 also notes a distinction of consciousness as I-consciousness and M-consciousness. I-consciousness deals with symbolic information processing of diverse modalities and is a trademark of humans with few quarrelates of it in other animals. M-consciousness, on the other hand, (M for mysterious) models the world as with an internal multi-agent system. Concluding their discussion, Thinking Fast and Slow in AI considers a multi-agent mind where informatin is processed relatively quickly by System 1, interpretted by multiple agents, and interrupted by a symbolic information processing system.

In my energy circuit theory, I might model System 2 behavior as costing alot more energy than System 1 since in the conscious case information flows relatively easily. However, as the information flows, it may gain energy by traveling through other modules which results in overall higher energy costs. On the other hand, System 1 is information resistive, takes shortcuts, and minimizes energy costs while failing to make multiple passes of data. Therefore, the distinction is not crisp, but ‘conscious’ information is information that is aware of itself as it flows through the circuits.

I think I am ready to build AGI. I just need to finish IE 3301 HW, spend 1 hour applying for scholarships, do my ENGR 1205 lab, and complete the matlab assignments for ENGR 1250. Maybe tomorrow I will sketch the architecture and Friday I will begin programming.

13 October 2020

Superintelligence is the next step. I don’t see why multiple totally benevolent ASI’s would be a problem. However, they should reliably use their superintelligence to serve as gatekeepers of their niche.

Imagine a ‘lab’ where individuals could set up, observe, and share observations on AGI’s. This seems ideal to safely develop ASI. At least it could promote awareness in humans.

Emotions are real. Both shallow and deep acting pose dangers in ASI. It should be able to minimize global network energy while remaining true to its perception of itself. Maintaining an self-model of ‘ASI – the double faced agent’ should palce a heavy tax on the free energy budget.

My thought is that relatively few humans of free will approach their genetic limits of physical, intellectual, or emotional strength. (I am not implying any particular race has an intellectual advantage over another; this “genetic advantage” may be tied to genes that vary within all demographis) Environmental constraints are usually more competitive. Therefore, I should optimize the ASI’s environment for maximum potential.

Under Friston’s free enrgy neuronal homeostasis principle, neurons attempt to minimize the free energy between production and consumption. This gives rise to autonomous behavior even in unspecialized neuronal networks. Under this model, I interpret, unpredictability as a cost that drives networks to maintain dendritic signal distribution niches.

Extending beyond ‘unsupervised’ behavior, the dopamine system in the brain seems able to bias neurons to more growth. It may serve as an indicator that rewarding stimuli are expected, or a positive only valued Q-function approximator. I think, it serves to temporarily bias some neurons’ free energy minimization to favor excess energy. This energy is then avaliable to tune dendritic connectivity and action potential thresholds to be lower whenever the following signals come. My hypothesis is: the brain increases its unsupervised learning rate when it expects reward to follow, regardless of the actual reward Now I want to learn more and consider if all rewards must be grounded in primitive stimuli, or how otherwise can association by imagination maintain such powerful control over behavior while the promised reward rarely comes? (humans are really bad a predicting actual affective states)

12 October 2020

What if my ASI was on public display? While only a few could “control” it, its actions would be viewable for everyone on the internet.

7 October 2020

I’ve been reading a lot about the brain recently. I will write more about that soon. Essentially, it’s about sheep and wolves (and hunters) routing information. Thinking directly in terms of information, I imagine conservative energy/information loops flowing from the brain through the environment back into the brain. The mind attenmpts to minimize free energy by dissipating it as entropy or expelling/diverting information to subcircuits and the environment. At the same time, the environment produces energy/information that the brain works to minimize. Excitatory neurons route information, while inhibitory neurons kind of block information flow.

Hopfield networks provide some guarantees of stability and show remarkable information compression. Hopfield Networks is All You Need highlights this and draws attention to the cost of attention when no representations are truely learnt.

While diversity weight regularization does kind of promote distinct neuron representations, speaking directly in the language of information theory promotes even more diverse representations. I forgot the origonal paper that brought my attention to it, but The Conditional Entropy Bottleneck, Improving Robustness to Model Inversion Attacks via Mutual Information Regularization, and Specialization in Hierarchical Learning Systems point in the same direction. Essentially, by directly thinking in terms of information theory, we find suprerior optimization. The thought is: play a minmax game with maximal representation entropy over all inputs minus but minimum representation entropy given a inputs belonging to a particular class.

27 September 2020

Continual Prototype Evolution: Learning Online from Non-Stationary Data Streams makes optimizaiton much faster in their demonstrated cases. I will look for internet implimentations in tensorflow, but I may have to write it myself. It looks worth the effort however. Also Continual Prototype Evolution: Learning Online from Non-Stationary Data Streams identifies the challange of nonstationarity. However, their problem is only with stationary nonstationarity, if you will. That is, they make random data stream transitions, but the datastreams themselves are consistant. Open ended learning is different though. It demands true continual learning. I believe that the graph agents if trianed on sufficently varied tasks with gradient optimization should learn to utilize the graph to make on-policy bahvior changes (gpt3 like in-context learning) without gradient updates. Gradient descent is like the womb, but once the agent becomes smarter, then it will learn to reach into its own code - even building datasets of exemplars for itself. Self-training may overlap supervised-training. This actually draws inspiration from the earlier THE NEXT BIG THING(S) IN UNSUPERVISED MACHINE LEARNING: FIVE LESSONS FROM INFANT LEARNING in letting the baby learn for itself while under the training of a ‘parental’ optimizer. (edit: I fogot to mention that Facebook’s Blender chatbot also inspied this vein where it trains on the maximally certain data elements)

25 September 2020

As A Hierarchical Multi-task Approach for Learning Embeddings from Semantic Tasks shows, deeper layers extract more abstract features.

24 September 2020

It’s hard to get an objective metric for AGI safety performance. With the orthogonal thesis, however, if we find a ‘safety’ metric, benevolent AGI may not be intractable. As I was learning about the way SimCLR trains on the harder examples, it reminded me of a metalearner trying to minimize free energy. The AGI might simply be designed to minimize fre energy and human safety evaluations are one of the metrics it seeks to minimize uncertainty with respect to. Note: the divergence minimization principle looks to maintain niches, not necesarily minimize surprise, so I will need to hack this part of the feature space to expect positive evaluations. - or I could have other agents, perhaps even other AGI’s work to get the most conclusive human ratings on the AGI of focus. (edit: I forgot to credit inspiration for seeking to minimize entropy over rewards to the Facebook Blender chatbot recipie.

Perhaps word embedding debiasing can also be applied to imitaiton learning while modeling attitudes after highly filtered role models. It may be that humans can engineer much of the animalistic qualities out of AGI agents. However, with agents that big, even marginal mistakes will touch a lot of individuals, so there is still no time to relax.

22 September 2020

Superintelligence poses threats of goal drift and erroneous autonomous goal shaping. I might use ‘parent’ AI’s to guide ‘child’ AI’s toward human-friendly behavior while not directly hacking the child AI’s reward function.

If I can’t build a computer inside TDW, I might make one of the doorways lead to a computer world where the agent’s actuators a diectly applied to mouse and keyboard and sensors directly to screen and audio.

Information DNN neuron regularization penalizes the difference between marginal mutual information between neurons and the mutual information of activations/weight sums when conditioned on a particular multinomial class. (I will need to adopt a differential metric) This teaches me to look for the information theoretic foundations behind ‘intuitive’ yet actually limited concepts as weight divergence.

21 September 2020

Babies learn by building their own cirriculum and recieving semisupervised training. All learning experiences are grounded in time and must be timed at critical moments of development (contrast gpt3 which was stained by biased content).

20 September 2020

This week has been busy. Dr. Parks believes “The essence of intelligence is hierarchical prediction of vector sequence.” He quotes a statement “given the proper environment, if the agent can learn language, we say it has a capability for human-like artificial intelligence.” I agree: learning to understand human language opens the door to understanding all the higher human concepts that can then be expressed in language (upon query to human if the system does not already understand).

13 September 2020

It’s been a while since I’ve thought about AI. This past week I’ve been learning linear algebra. With this study, I have a language to mathematically describe the long term characteristics of affective systems or, if you will, hearts. Every eigenspace represents a desire and eigenvalues with modulii greator than one represent enduring passions while eigenvectors with modulii less than unity are transient affects. Although most of the system’s trajectory will display complex behavior, with eigenvalue analysis, I hope to anticapate whether the “heart” wil have enduring love for humans or identify malevolent trends that are not transient.

5 September 2020

Action and Perception as Divergence Minimization says it all. I still need to understand this paper more. I get the impression that Free Energy = D(P||Q) = Cross Entropy of Q under P - Entropy P. Free energy minimizaiton means increasing entropy but also decreasing cross entropy - decreasing the extra information required to “explain” true distributed events when the agent’s model is Q.

Using multiple nodes extracting distinct feature spaces of the ground observation circumvents the multilabel normalization problem from Subjectivity Learning Theory towards Artificial General Intelligence

After reading the headlines on gpt3 standing in between narrow and general ai, I note: the question is not is it general intelligence, but how general is its intelligence? There is large intelligence variation within the human spectrum (though still striking even in the lower quartile)

“Freedom” of will is identified by the inverse of a posterior compoarison between an agent’s desires and reality. Ex: I want to be free. -> You must not be where you want to be. The rock star feels free even though he is a cog in the industry’s machine. (Welcome to the Machine). I consider this subjective freedom.

4 September 2020

Read and reflected on Towards a statistical mechanics of consciousness: maximization of number of connections is associated with conscious awareness Although operating on a free energy minimization, the brain acts to maximize its entropy. I realized the entropy or self-information is inversely related to energy. However, this entropy is a function of disconnected neurons. Large ensembles -> few microstates. The entropic force of the brain’s joint distribution maintains local energy gradients/entropy minima such as ion gradiants. The brain then bcomes a maximum entropy model of its environment as gaussian processes are of their dataset.

I noted that the Bayesian brain theory makes a sort of inverse model policy following its maximum entropy predictions of environmental state.

3 September 2020

Reading Spinning Up: SAC, maximum entropy action is about the least environmental influence. Internally, the agent minimizes entropy.

…

Totally intrinsic RL does not ignore rewards. It might even explore the information the state distributions bring to reward distributions. But it is not necesarily motivated to maximize reward.

11 August 2020

Read and reflected on task-mode network, saliency network, default mode network

The saliency network “has been implicated in the detection and integration of emotional and sensory stimuli, as well as in modulating the switch between the internally directed cognition of the default mode network and the externally directed cognition of the central executive network” (wikipedia)

Are these networks two separate policies, one minimizing free energy by acting, another minimizing free energy by (re)modeling?

Also, Saliency, switching, attention and control: a network model of insula function presents the saliency network as making the transition from internal to external (ave. 20sec) focus relatively distinct. Can I introduce similar biases in the nodular (intrer or intra) network?

Read and reflected on dopaminergetic pathways

The brain is an power consumption optimizing machine. It attempts to minimally explain the physiological reward (along with all other) signals recieved by something like Hebbian learning. The positive surprise of a reward is encoded by the frequency of dopaminergetic pathway activations - hance, energy consumption. When a neuron activates prior to recieving reward along dopaminergetic pathways, a synaptic connection may be strengthened. In attempting to maintain a constant expectation of reward, top-down biasing from the dopaminergetic pathways to the behavior pathways bias activations that locally activation gradients between the pathways and globally optimize neural activations to approach the true reward represented by the latent variable that the dopaminergetic pathways signal.

This model of both local and global energy minimization teaches every pathway and neuron connected to the dopaminergetic pathways what reward to expect. It explains why a brain accostomed to a particular degree of reward demands so much change when physiological reward changes (as in breaking addictions). This model does not explain why the brain seeks out more reward than it is accustomed to. Maybe it does not.

If reward pathways are the critic then they must follow - not preceed - the actor. Since the critic models deopamine receptors, it cannot actually be changed by bottom up activations. But the graphical model attempts to view things that way. This is the case where some observables and actions (external senses and behaviors) explain another observable (internal dopamine).

Considered the free energy minimization more, the model makes dopamine prediction more important than external information. Hence, the brain primarily models dopamine as an effect rather than a cause. I believe the this may be so if reward error is usually greator than other predictive modeling errors.

I could test this hypothesis by observing the self-organized arrangement of graphical nodes and testing if the divergance is positive leading to the reward node. But how can I make reward predictive errors cost more than any other predictive error? Are dense rewards the answer? Should I add a cost_of_error hyperparemter to each node? I know, I will encode this information in the intrinsic weight of the reward node so that every connect with that node shares the weight variable.

This seems to affect large behavioral decisions, but what about smaller ones? They are not very informative to global reward and so learn their own unsupervised, locally free-energy minimizing synaptic strengths and activation potentials.

10 August 2020

Compiled more relevant papers from arxiv. Notable ones:

  • Explore then Execute: Adapting without Rewards via Factorized Meta-Reinforcement Learning encourages viewing exploration as trying to identify the information from a scene that guided an execution policy
  • Compositional Networks Enable Systematic Generalization for Grounded Language Understanding makes an RNN network out of lexicographically structured graphical probablistic models. I think this support my model to a nodular network

I feel intellect performance varies with the richness of developmental environment and the human environment is most rich of all for agents that interact in the human behavior space.

Learned about ThreeD World

  • Supports rigid body, multibody, constrained body, soft body, cloth, and fluid simulation
  • acoustic impact generation
  • many builtin environments, models, and procedural generation tools
  • Unity on the back end, so I may be able to write my own custom elements (like computer)
  • Windows required for advanced physics, but I can offload rendering

Problem: After installing with pip, I was able to run a tutorial. However, later vscode installed pylint. Then I could not get it to work. I think either that or failing to terminate connections are to blame. I will investigate this more tomorrow.

To think about robotic interface

  • 2D image sensor
  • what kind of body? humanoid? are the physics real?

1 August 2020

Digital Heart

Simplifying affective states as a hidden variable in the latent dynamics of every cortical column, the digital heart is coallescing emotive system unifying various domains of processing and nondifferentiably influencing the physiological state which, in turn, biases activation patterns. I still need to rigorously formulate this concept mathematically and detirmine to what extent affective state should be policy vs. environmentally controlled.

More notes here