How does human memory work—no, more, how can it work? How can our knowledge of this influence the recurrent neural networks and reinforcement learning algorithms that we implement? Those replete with input and forget gates, with memory replay?
“What is your earliest memory?”
He spoke of a prideful moment. He had just speared a fish, while running barefoot up and down the mountains.
She spoke of humiliation, when all the children in school ridiculed her.
I spoke of curiosity. My hand moved under a glass table—presumably in awe at the physics of transparency.
What our minds felt compelled to remember was at once most striking and least surprising to me. Each matched our different personalities spot on.
Colors of memory. Notes of memory. Flavors of memory. Tenses of memory.
People store representations of the world in different ways. Some see numbers in everything and remember them as such in their memory. Others, color. Still others, notes. Musical notes.
For me, it is tension. This, I feel subtly in different parts of my body. It is primarily skeletomuscular. I organize foreign languages this way. I recall new people and interactions this way. I remember precise placements of objects in rooms and punctuation in sentences this way. I have recited Shakespeare this way.
That is my short-term memory. My long-term memory is devoid of physical sensation or emotion. It represents events as is—without the anger, fear, disgust, without the joy, relief, curiosity that must have been present at the time.
I cannot fathom what other resources my system uses to store and compute data—and steeped in each method of resourcefulness, what it is uniquely primed to synthesize and/or forget.
Extension, Preview, Replay
Setting. First, I do not see faces. I sense presence, so I know who is there based on their distinct features. There is little color. It is silent. There is gait and there is language. There is punctuation. Literally, there is punctuation that is there to help represent how words are communicated. Words are transmitted, not spoken. Grammar is essential.
Content. When lucid, I generally extend a set of possible interactions into minutely different instances. Differences occur in the tone with which I said something, the body language I gave off, the punctuation I had replaced. I play them out, all of them, endlessly, sensing the interaction dynamic skeletomuscularly. I compare them. I am preparing for them.
This is close to visualizing your possible future interactions—except nothing appears to engage the visual cortex. I love these simulations.
Memory replay. Recent work in reinforcement learning show that algorithms, which replay perturbed version of policies that led to disastrous outcomes, edge autonomous agents towards performing safer exploration. I sense that I am using replay not only to remember past mistakes were a bad idea, but also to test, experience, and consider high-risk situations—with which I engage, in a bout of rich déjà vu, when I awake.
Source: Deep Learning on Medium