This bit discusses working memory and decision making (Page 25):Hi FF,
This paragraph at the bottom of column 1 of page 15 discusses memory.
The original motivation for this model was to obtain a realistic description of associative memory occurring on top of spontaneous brain activity with the latter being a global attractor of the dynamics. In a later version in which neural clusters can overlap (modeling the fact that real neurons can code for more than one stimulus), this model is the closest biologically plausible analog so far of the influential Amari–Hopfield network (discussed below in Sec. VB3) and can store an extensive number of stimuli modeled as patterns of firing rates across neurons.
My neuroscientific layman's guess is that, in Akida, the model libraries and associated weights are equivalent to memory. Coding neurons for more than one stimulus is performed in Akida by selecting a different model from the library.
So my question is, can Akida be, to some extent, adapted to cortical neuron configuration by configuration of the data in the model, and to what extent is hardware modification required to enable cortical operation? Does this involve modification of the NPUs or the interconnexion matrix, or both?
The existence of this kind of metastable dynamics in a spiking network was first pointed out in Refs. 16 and 17 who noticed that, because of metastable activity, the network produces slow fluctuations in the neural activity, much slower than the time scales of the single neurons—the origin of such timescales is a long-standing problem in theoretical neuroscience.170–172 They also showed that, unlike the case of a homogeneous excitatory population, a stimulus will suppress trial-to-trial fluctuations, another widespread phenomenon in cortical circuits.173 Reference 22 also found that, in this clustered network, the dimensionality of the neural activity (Sec. VA3) is larger during ongoing metastable dynamics than when the network is externally stimulated.
I think what this is saying is that concentration suppresses idle thoughts, much like changing from an idling motor to engaging drive. Or, in the case of Akida, the model is loaded, but there is no input. There may be degrees of concentration, like thinking about doing something and actually doing something.
Reminds me of the time I was taking a young lady for a drive in a rather sporty automobile, and she mentioned that the engine was rather noisy. I explained that that was the acoustic tachometer ... you know that look you get ...
Actively holding information online for a brief period of time (seconds) is an important ability of the brain. This capability is a part of working memory (WM), which is used for tasks such as planning, organizing, movement preparation, and decision making.281–284 In contrast to long-term memory, which requires structural changes in neural circuits and in the connections between neurons, the mechanisms underlying working memory are believed to depend on persistent neuronal activity.285–287 In general, positive reverberation driven by recurrent synaptic excitation in interconnected neural clusters can work as the basic principle for generating persistent activity. Triggered by incoming signals, working memory circuits can sustain an elevated firing even after the inputs are withdrawn. As discussed in Sec. VB1 [see, e.g., Fig. 6(a)], dynamical models with local feedback excitation between principal neurons that are controlled by global feedback inhibition can exhibit multiple attractor states (each coding a particular memory item) that coexist with a background (resting) state. This is illustrated in Fig. 11 for a circuit model of WM with two excitatory populations.
In Akida 2000, LSTM will be a static equivalent of the dynamic "positive reverberation driven by recurrent synaptic excitation. ".
"In addition to robust maintenance of memory states, activity should be reset quickly when there is a novel sensory cue that needs to be stored. In other words, a working memory system should have the properties of robustness against fluctuations while being very sensitive to incoming stimuli. In recent theoretical works, this fundamental contradiction can be achieved by global inhibitory connections, where the system can exhibit structurally stable dynamics with fixed stimulus and qualitatively change its dynamics if the stimulus is changed."
I learnt a long time ago that academics proposing to provide a review of an entire area of science are having a lend and just trying to keep up their academic publication numbers which are apparently important in CV's for promotion in their world.
Poor old Peter van der Made would not get a job based on publications at the moment as he has been in stealth mode for years now.
I only posted the extract as being the easy read as I have searched for ages for someone prepared to encapsulate the outcome of understanding and creating a simulation of a cortical column. As for the maths I always assume that is correct and peer reviewed as I could not check it anyway.
The philosophy of whether we will ever understand the human brain using the human brain to derive an understanding and allowing for the variation in human brains and ongoing evolution is an intellectual quandary that I often mull over and cannot resolve. It does seem the more those engaged in this endeavour learn the less they know.
Personally I think that the human brain is a flawed model upon which to build useful artificial intelligence. The human brain makes mistakes constantly hence the saying "To err is human to forgive divine" or something like that. Being human not sure if I am misquoting but I will quote it nonetheless as I am using a flawed human brain to construct this response.
What use is an intelligent machine that perfectly mimics the human brain down to the making of inexplicable mistakes and the ego to keep going at it regardless.
In Autonomous EV's the push is in large part so that they can remove human drivers and eliminate accidents. Go figure that one out when you read about attempts to emulate the human brain to create truly autonomous vehicles. I am not Einstein but the logic breaks down in my opinion.
In short the artificial general intelligence that Peter van der Made is seeking must at some point divert from the position of fully replicating the human brain and as such a complete understanding of the human brain and how it works is unnecessary to this goal in my opinion but what would I know I am human after all and constantly right and wrong unless you ask my wife who will tell you I am constantly wrong unless I agree with her unless it comes to the law and investment and doing our tax planning.
My opinion only DYOR
FF
AKIDA BALLISTA
On-chip learning goes a way to meeting the requirement to store novel sensory cues.
...
Conclusion (Page 30)
One of the most relevant implications of the fact that cortical activity evolves as a sequence of discrete, metastable states is that transitions in neural activity are not just triggered by external events, such as a stimulus or a reward, but are instead spontaneously generated and may occur at anytime, including when the subject is idling and not engaged in a task. This goes against the notion that neural activity is just a “reaction” to external events or a static representation of incoming stimuli and is compatible with the presence of incessant “ongoing activity” observed in the cortex.
As FF asks, to what extent is it necessary to reproduce daydreaming?