Nand Chandravadia

ml phd @georgiatech | trying to understand neural networks | formerly @columbia @ucla | love music & cats

[Nand Chandravadia @notnandgate]

@slowcargofast A prior is built-in knowledge — an assumption baked in before any data arrives. Here, the brain's rhythm is that assumption: it expects speech to ebb and flow at a certain rate, so it's ready for speech before it even hears it

[Nand Chandravadia @notnandgate]

[1] Speech has a rhythm. So does your brain. We found that's not a coincidence — it's a built-in prior for recognizing speech 🧠🌊

Excited to share recent work with Nabil Imam: "Neural rhythms as priors of speech computations"

[1/6]

[Patterns, a Cell Press journal @Patterns_CP]

Online Now: Neural rhythms as priors of speech computations https://t.co/iakIEogKhw #datascience

[Nand Chandravadia @notnandgate]

For all the technical details:

Paper: https://t.co/xu67PK81Lv

[6/6]

[Nand Chandravadia @notnandgate]

When we gave those same networks urban sounds — car horns, jackhammers, sounds with no real rhythmic structure — the advantage just disappeared. The rhythm isn't a generic trick; it's a prior tuned to speech, with the brain's rhythms encoding knowledge of speech structure itself.

[5/6]

[Nand Chandravadia @notnandgate]

@curiouswavefn apparently, he tried to persuade Hinton that his work on neural networks was "nonsense"

[Nand Chandravadia @notnandgate]

@pmddomingos understanding principles of brain computation may lead to more efficient AI

[Nand Chandravadia @notnandgate]

this paper, "Biologically Plausible Speech Recognition with LSTM Neural Nets" was written in 2004...talk about being incredibly prescient @SchmidhuberAI

[Nand Chandravadia @notnandgate]

@rieszspieces actually, they span the lounge 🤓

[Nand Chandravadia @notnandgate]

accidentally called it frenchel duality. Ma faute!

[Nand Chandravadia @notnandgate]

just randomly stumbled into this: quantum neuroscience...hmm

[Nand Chandravadia @notnandgate]

@AcquiredFM 1978. the pre-backprop days...

[Nand Chandravadia @notnandgate]

@GrantStenger interesting. the first plot looks an action potential in a neuron (e.g., spike) and red trace looks like the raw neural trace (with the spikes)

[Nand Chandravadia @notnandgate]

@__paleologo "Simple Model of Spiking Neurons" (by E. Izhikevich). A relatively simple model, but captures a wide variety of qualitative (brain) dynamics

[Nand Chandravadia @notnandgate]

@moleculesrcool https://t.co/a3uX21pVu3 (plenty more)

[Nand Chandravadia @notnandgate]

@Aella_Girl Dvoretzky–Kiefer–Wolfowitz