Episodic Memory and Blockchain Technology: A Requirement for Cooperation between Humans and AI
It is believed that when the brain records past events, it stores them in the form of episodic memory. In this article, we will discuss episodic memory.
Episodic memory emphasizes storing memories in a consistent manner without contradictions and allows for forgetting. The reasons and mechanisms behind this will be explained in this article.
We will also touch upon guidelines when equipping AI with the ability to record events. For AI, using a mechanism different from human episodic recording is more efficient. However, maintaining consistency, as emphasized in human episodic recording, is crucial. Applying blockchain technology as a means to ensure robust consistency is a potential idea.
Let’s delve deeper into the main content below.
Requirements for Episodic Memory: Forgetting is Allowed, But Mistakes Should be Minimized
Every day, we store memories of events. Such time-associated information is called temporal information. Events like daily occurrences are stored in a part of the brain known as episodic memory.
We believe the events stored in our episodic memory to be true. We trust what we directly observe, feel, and think. If not, on what would we base our thoughts, live our lives, or communicate with others?
Therefore, a high level of reliability is demanded from our brain’s episodic memory. Our brains often forget various episodes. Forgetting is not the problem; the issue arises when a memory is wrong or morphs into something incorrect.
It’s unpredictable which parts will be forgotten. For instance, suppose you go out with two friends, A and B. If you forget the entire event, it’s not a major issue. But if you forget about B, your memory becomes about just being with A. That’s an incorrect memory.
Understanding Causality in Episodic Memory and Consistency Check
To prevent such mistakes, episodic memory utilizes two functions of brain information processing: grasping causality and checking for consistency.
Events have causes and outcomes. A ball hit breaks a window. You went out because person A invited you. A photo remains because B took a picture of you with A.
When recalling a memory, the brain checks if the causality remains consistent.
For example, if you forget that B also went on a trip with you, you’d feel uneasy seeing a photo with only A and you. Typically, if you don’t ask strangers to take photos during trips, it wouldn’t make sense unless another friend was there.
As a result, even if you can’t recall B, you realize the memory of the trip’s participants is vague.
Additionally, when recording an event in episodic memory, the brain ensures causal consistency.
For instance, suppose you interact with twin brothers A and B. You witness A perform task C and, immediately after, B perform task D. However, you soon notice a scar on B, which he shouldn’t have. You realize you confused A and B.
At this moment, you feel compelled to correct your memory of A performing task C and B task D, updating it to B performing C and A performing D.
As shown, when making an episodic memory, causality is used to ensure the memory doesn’t remain contradictory.
This way, our brain’s episodic memory ensures memories of events are correct, both when saving and recalling.
The Need for Consistency Checks
Let’s consider why the human brain, when writing and reading episodic memory, uses this mechanism based on causality for consistency checks. Likely, it’s because the neural net, an aggregation of nerve cells in the brain, is not suited for recording history.
This is evident in the nature of episodic memory to forget events. The structure of the neural net itself doesn’t operate to remember data chronologically, and its nature requires repeated experiences to retain memories long-term. While efficient for pattern learning, it poses challenges for episodic memory.
It might have been beneficial for a mechanism exclusively for episodic memory to have evolved, or perhaps the advantages of forgetting outweighed the drawbacks. Regardless, the brain has utilized the neural net to realize reliable episodic memory. I believe this reliability enhancement mechanism is the causality-based consistency check.
I theorize that episodic memory ensures history isn’t recorded or played back incorrectly to meet reliability requirements. This is crucial for survival, like remembering safe or dangerous places and where and when to find food. However, this nature to forget becomes a disadvantage.
Considering this, the reason for adopting a memory mechanism that might forget but doesn’t compromise on reliability is probably for communication with others. If two people experience the same event and one remembers while the other forgets, they can complement each other’s information. This fosters cooperation.
On the other hand, if two people experience the same event but recall contradictory details, it’s problematic. Conflicts arise, and they can’t support each other. This is inconvenient for a strategy that improves survival rates by maintaining a group. Therefore, episodic memory probably adopted a method that minimizes contradictions, even if it allows for some forgetting.
Requirements for AI’s Temporal Memory of Events
Let’s consider a scenario where conversational AIs become more integrated into human society beyond just being tools. In such a case, we would demand that AI have consistent memories, just as we expect from human episodic memory.
Humans, using neural networks that aren’t adept at remembering chronological history, are thought to have mechanisms to detect inconsistencies.
On the other hand, it might be better for AIs to remember events in chronological order, like standard computer logs, rather than trying to store episodes on a neural network.
However, from a reliability perspective, there are concerns different from humans. Humans have the risk of memory errors if they don’t check for inconsistencies because neural networks forget. On the contrary, since the brain is enclosed within the skull, there is no risk of memories being maliciously altered from outside.
For AI, unlike these human traits, there’s a possibility that their memory logs can be intentionally altered, either by tampering with the storage or through hacking over the internet. Therefore, for different reasons than humans, we need to enhance the reliability of AI’s memory logs. One could consider using blockchain technology. Another measure might be to erase memories if the device is tampered with or hacked.
Applying from Human Episodic Memory
If AI logs everything it perceives, experiences, and thinks, it would require a massive amount of memory. Envisioning a future where various AIs, from large to small, are used appropriately, we will need technology to conserve memory space.
There are common methods to reduce memory storage, such as data compression or deleting old data. In addition, ideas from human episodic memory might also offer insights into memory reduction.
Human episodic memory strongly retains impactful events while weakly storing mundane ones. Moreover, humans not only forget but also strategically don’t remember unnecessary details.
For instance, routine actions like washing hair with shampoo or locking the front door might not always get stored as episodic memories. Feeling uncertain if you’ve locked the door after a while and going back to check is due to this.
On the other hand, when something out of the ordinary happens, the brain stores the event in its episodic memory. Unexpected surprises or tension from novel experiences are considered signals to strengthen such memories. For example, during the routine action of locking a door, if you drop a handkerchief, you’ll likely remember it.
Furthermore, the memories we hold are often fragmented. With these fragments, the brain doesn’t need to remember easily deduced information. For instance, if you walk from point A through B to C, unless something unexpected happens, remembering starting at A and arriving at C is sufficient. Recalling every step isn’t necessary.
If someone asks whether you walked past point B, even if you don’t remember passing point B, you can confidently answer “Yes, I walked past it” by combining your memories.
Studying these techniques of human episodic memory should significantly contribute to reducing AI’s memory needs.
In Conclusion
From observations of human episodic memory, I’ve organized the requirements and mechanisms inherent to it. While there isn’t scientific backing for this, when considering the characteristics of how we personally remember things and the properties of neural networks, I believe I’ve managed to explain it reasonably well.
In the latter part, I delved into considerations regarding how AI could potentially store memories of events. Pattern recognition and natural language processing thrive because of neural networks that imitate the structure of human brain neurons. On the other hand, the way AI stores events doesn’t necessarily have to mimic the structure of human episodic memory. Taking into account the inherent characteristics of computers, which form the foundation of AI, designing it in a way that seems logical might be more appropriate.
Furthermore, the idea itself of giving AI long-term memory is something various researchers and engineers would likely come up with quickly. In fact, research and development appear to be ongoing in the form of AI agents. As such AIs become more integrated into society, it’s not just about having them chronologically store events; this article argues for the importance of ensuring the trustworthiness of these memories.
Of course, the development and actual operation of AI with long-term memory should be carried out carefully, checking trends in AI ethics and regulations. At the same time, it’s a fact that research and development for AI with long-term memory are actively underway.
As a result of such research and development, we must avoid the risk of allowing the spread of unreliable AIs that can be hacked or have their memories altered. From this perspective, having requirements for the long-term memory of AI seems essential.
