chips copy our ability to learn with 4 rules
replicating synaptic plasticity in neuromorphic systems 👩🔬
welcome to introspection ft. harsehaj! ⭐️ i’m harsehaj — always up to something in social good x tech. scroll to the end for a daily roundup on unique opportunities. :)
chips copy our ability to learn with 4 rules 👩🔬
i wrote about neuromorphic chips exactly a month ago, and now i’m back to dig into how exactly neuromorphic chips learn like the human brain does. in other words, replicating synaptic plasticity in neuromorphic systems.
first things first, synaptic plasticity refers to the ability of a connection (synapse) between two neurons to strengthen or weaken over time in response to activity. this is learning, at a biological level. so, if you touch a hot stove and feel pain, the neurons responsible for “hot” and “pain” will reinforce their connection.
neuromorphic chips model this plasticity through hardware by updating synpases via biology-inspired rules instead of learning by calculating gradients (like in deep learning). the first rule, of course, is hebbian learning, summarized as “neurons that fire together, wire together.” simply put, if two neurons activate at the same time, their connection strengthens and associations are formed. liiike linking a song to memories associated with summer 2016.
now, we further hebbian learning with spike-timing dependent plasticity (stdp), which adds a significant twist: time matters. in biology, causality matters so the order of spikes tells us which signal likely caused the other. as a result, we have temporal pattern learning, sequence recognition, and real-time adaptation. for neuromorphic chips, this stdp piece can actually be hardcoded into silicon or reproduced through algorithms. this enables chips to learn by exposure to patterns — like us.
neuromorphic chips tend to also follow 2 more brain rules:
homeostatic plasticity (self-regulation system) → synaptic strengths are reduced for overactive neurons, and vice versa. this ensures the system is stable, and is crucial in preventing parts of neuromorphic chips from dominating each other or shutting down. think of it like auto-adjusting the brightness of your phone according to the surroundings.
neuromodulation (rewards system) → yes, timing matters, but so does importance. dopamine is like a reward signal for our brains, and it also modulates learning by strengthening connections to reward good outcomes. similarly, neuromodulation adds reward-driven learning in neuromorphic systems.
so (hebbian learning → stdp) + homeostatic plasticity + neuromodulation = neuromorphic chip that can learn like we do. :)
if you want to dig into how this is implemented more technically, i’d recommend a deep dive here.
today’s drops 🔍️
apply to pitch at new york tech week by may 25th
product internship @ omada health
$11,000 scholarship for pre-med students (high school, undergraduate & graduate)
i also write a bi-monthly update focused on my work and project updates if you’d like to check that out here.
want to put an opportunity on my blog? shoot me a line at harsehajx@gmail.com.
teehee,
harsehaj ✌️
PS. if you have a question/topic you think would be interesting for me to reflect on, don’t hesitate to comment or reply to my emails with any ideas you ever have. :)