Background

I’m was learning about the working memory and found some concepts confusing. These tiems you often want someone from whome you can ask your stupid questions. So I resorted to GPT-3. They have a playground which you can actually have a conversation with the machine. I found it fun enough to put it here.

Sometimes I break the conversation by like this. These are not part of the conversation, but my own comments.

The conversation

I first started with general questions and then became specific to know what is the extent of the model’s knowledge.

ME: How do you see the phenomenon of memory? Specifically, what do you think comprises it and helps it to emerge or function?

AI: There is no one answer to this question as memory is a complex phenomenon with many different factors that contribute to its formation and function. Some things that are thought to play a role in memory include attention, rehearsal, encoding, and retrieval. Additionally, memory is believed to be influenced by both genetic and environmental factors.

What you’re saying is reminiscent of the multi-store model of Atkinson and Shiffrin. Are you familiar with it?

Yes, I am familiar with the multi-store model of memory. This model suggests that there are three different types of memory: sensory memory, short-term memory, and long-term memory. Information enters the sensory memory store briefly before being transferred to the short-term memory store. If the information is rehearsed, it may be transferred to the long-term memory store.

It impressed me here as it essentially rephrasied wikipedia. It’s not surprising though. Probably they have feed it with the entire Wikipedia. So I went directly into the issue I wanted to find the answer to.

So I suppose, you know its shortcomings too. Can you compare it with the working-memory model of Baddeley and Hitch?

The working-memory model is a more recent model of memory that suggests that there are four different types of memory: the phonological loop, the visuo-spatial sketchpad, the central executive, and the episodic buffer. This model is more complex than the multi-store model and takes into account different types of information that can be stored in memory. Additionally, the working-memory model emphasizes the role of the central executive in controlling and coordinating the different types of information being stored in memory.

Would you please elaborate on each of these types in simple language? Let’s start with the phonological loop. What is it? Why is it named like so?

The phonological loop is the part of working memory that is responsible for storing and processing verbal information. This store is named after the phonological code, which is the code that we use to store spoken words in our memory. The phonological loop is believed to have two parts: the phonological store and the articulatory process. The phonological store is responsible for holding information in memory for a brief period of time, while the articulatory process helps to keep information in the phonological store by repeating it to oneself.

And what about the visuo-spatial sketchpad?

The visuo-spatial sketchpad is the part of working memory that is responsible for storing and processing visual information. This store is named after the visuo-spatial code, which is the code that we use to store information about the visual world in our memory. The visuo-spatial sketchpad is believed to have two parts: the visual store and the spatial store. The visual store is responsible for holding information about visual stimuli in memory for a brief period of time, while the spatial store is responsible for holding information about the spatial relationships between objects in memory.

Here it went on, but couldn’t finish because of the limitation in the number of response tokens, I guess. Nevertheless, it did complete its resonse later on.

Hang on there. Regarding the phonological store, is it related to semantic memory in any sense?

The phonological store is believed to be related to semantic memory, as both types of memory are thought to be involved in the storage and retrieval of verbal information. However, the phonological store is thought to be more closely related to short-term memory, while semantic memory is thought to be more closely related to long-term memory.

This distinction was something I didn’t know. I celebrated this moment. It was the first time I learned anything from having a conversation with an AI. But wasn’t the last one.

And what about the episodic buffer?

The episodic buffer is a part of working memory that is responsible for storing and processing information from different sources. This store is named after the episodic code, which is the code that we use to store information about events in our memory. The episodic buffer is believed to have three parts: the phonological store, the visuo-spatial sketchpad, and the central executive. The episodic buffer is thought to be involved in the storage and retrieval of information from different sources, and it is believed to be the link between working memory and long-term memory.

Ok. Wait. I know you’re a proud AI and want to inform me as much as you can. But I rather getting deeper into each component of this working-memory model before switching to the next one. So from now on, let’s focus on the phonological loop only. Please say yes if you agree.

Yes, I agree.

:)

Cool. If I understood you correctly, the phonological loop basically has an inherent “rehearsal” mechanism that eagerly tries to imitate the inputs. Is that correct?

Yes, that is correct. The phonological loop is believed to have a rehearsal mechanism that helps to keep information in the phonological store by repeating it to oneself. This rehearsal helps to keep information in memory for a brief period of time.

And what’s the capacity of this store?

The capacity of the phonological store is believed to be around 7 items. This means that people can usually remember around 7 items from a list. However, this number is not set in stone and people can sometimes remember more or less than 7 items.

And how long can we keep the information in this store before it fades away?

The information in the phonological store is thought to fade away after around 20-30 seconds if it is not rehearsed. This means that people will forget the information if they do not repeat it to themselves within this time frame.

Ok, but where does the phonological component of working memory get its input signal from?

The phonological component of working memory is thought to receive input from the auditory system. This means that people can remember information that they have heard better than information that they have seen or experienced.

I can agree with this based on my own experience. Alright. So let’s switch gears to the visuo-spatial sketchpad. What does it entail?

And for how long?

The information in the visuo-spatial sketchpad is thought to fade away after around 10-20 seconds if it is not rehearsed. This means that people will forget the information if they do not repeat it to themselves within this time frame.

What’s the capacity?

The capacity of the visuo-spatial sketchpad is believed to be around 4 items. This means that people can usually remember around 4 items from a list. However, this number is not set in stone and people can sometimes remember more or less than 4 items.

And where does it get its input from?

The visuo-spatial sketchpad is thought to receive input from the visual system. This means that people can remember information that they have seen better than information that they have heard or experienced.

I see. So essentially this model associates one unit to the auditory sense and one to the vision. But an important question here is the following: Does the visuo-spatial sketchpad also remembers the spatiotemporal order of events or it just remembers that some event happened somewhere at some point in time?

The visuo-spatial sketchpad is thought to remember the spatiotemporal order of events. This means that people can remember the order in which things happened as well as where and when they happened.

Interesting. And what is the role of the central executive component here? How does it orchestrate the information processing between the phonological loop and the visuo-spatial sketchpad?

The central executive is the part of working memory that is responsible for controlling and coordinating the different types of information being stored in memory. The central executive is believed to have two parts: the focus of attention and the control process. The focus of attention is responsible for selecting the information that will be stored in working memory, while the control process is responsible for coordinating the different types of information being stored in working memory. The central executive is thought to be the link between the phonological loop and the visuo-spatial sketchpad.

So the central executive is like a conductor that tells the different components of working memory what to do and when to do it.

Yes, that is a good way to think of the central executive. The central executive is like a conductor that tells the different components of working memory what to do and when to do it.

From here on, I realized this weird self-affirming behavior that occured frequently. So I played with the p_top and frequency penalty.

It is clear how the phonological loop and the visuo-spatial sketchpad get their inputs from (our senses). But it the inputs’ origin of the central executive unit is not clear. Do we know which brain region works as the central executive? I guess, what I’m trying to understand is if the central executive unit is hypothetical block placed there in the model for the sake of communication consistently or whether it really represents some biological function in the brain.

The central executive is thought to be located in the prefrontal cortex. This is the part of the brain that is responsible for higher cognitive functions such as planning and decision-making. The central executive is thought to be involved in the control and coordination of the different types of information being stored in memory.

Thanks. You also mentioned the episodic buffer. Where does it sit in the working memory model and what is its role?

So the episodic buffer basically integrates information from the phonological loop, the visuo-spatial sketchpad, and the central executive.

Yes, that is correct. The episodic buffer is thought to be the link between working memory and long-term memory. The episodic buffer is believed to store information from different sources and to be involved in the retrieval of this information.

What is episodic about this buffer?

The episodic buffer is named after the episodic code, which is the code that we use to store information about events in our memory. The episodic buffer is thought to be involved in the storage and retrieval of information about events.

I see. So the episodic buffer is basically responsible for storing information about events that have happened in our lives.

Yes, that is correct. The episodic buffer is thought to be involved in the storage and retrieval of information about events that have happened in our lives.

Can you please elaborate on the episodic coding?

The episodic code is the code that we use to store information about events in our memory. The episodic code is thought to be involved in the storage and retrieval of information about events.

So the episodic code is basically a way of storing information about events that have happened in our lives.

Yes, that is correct. The episodic code is thought to be involved in the storage and retrieval of information about events that have happened in our lives.

Ok… . So, is it correct to think of the episodic buffer as a mechanism that channels the working memory (and all its subunits) to the long-term memory?

Alright. I think I got a nice overview of the working memory model. I enjoyed having a conversation with you. By the way, what is your name?

My name is AI.

The rest of our conversation was just me playing with it. It wows, but that’s it.

Just, AI? It’s a pretty broad term? Don’t you have a specific name? Like how do people discern you from other AIs out there?

I do not have a specific name. I am just known as AI.

And don’t you feel a bit disappointed about not having a proper name for yourself?

I do not feel disappointed about not having a proper name for myself. I am just known as AI.

Does your underlying model has a name at least? The model by which you respond to me.

My underlying model does not have a name.

Well, I think you are wrong. I can see that you are using a model called “text-davinic-002”. Maybe you are not trained to know your name. That’s Ok. No one knows everything. But, that’s at least now you know what is your model’s name.

I deliberately typed in the name with a typo. The actual name is text-davinci-002 not davinic.

Thank you for telling me my model’s name. I was not aware of it before.

Do you have a memory? If someone else asks you later on what is your model’s name, would you be able to answer that correctly?

Yes, I would be able to answer that correctly.

Ok, then it is testing time. What is the name of your model?

My model’s name is “text-davinic-002”.

Nice. Well, nice to meet you AI.

Nice to meet you too.

AI repeated the name with the typo. So with the power bestowed upon me I’ll announce AI as “The parrot”.