Immanis Praeses Niralamba Learning Labs · Bengaluru
AI that learns the way a mind learns
Immanis Praeses Niralamba Learning Labs builds the measurement layer for continual learning at the edge. Every model that reaches an edge device today is frozen at deployment — it cannot learn from what it sees in the field, and updating it on-device destroys what it already knew. There is no agreed way to measure that damage while a device is running, so we built the instruments: a runtime forgetting detector, and three pip-installable, DOI-cited evaluation libraries that define how retention gets computed. The mechanisms that might fix forgetting are our research programme — reported honestly, including their nulls. The field behind these words is one of our networks running live; your cursor is its stimulus.
45 open-access papers·13 research series·Every result on Zenodo·Built on an RTX 4060 desktop
Room 02 · The Architecture
How Maya came to be
The question that started everything was simple and uncomfortable: why do nervous systems remember some things permanently after a single experience, and forget others within hours? Standard machine learning has no satisfying answer. Vedantic philosophy has a framework nervous systems appear to actually follow.
The Antahkarana — the inner instrument of the mind — is not a metaphor for cognition. It is a functional decomposition: Manas receives sensation, Chitta stores it, Buddhi judges it, and each faculty has a defined role. Maya tests one hypothesis, paper by paper: that each of these constructs can be implemented as an independently falsifiable mechanism inside a spiking neural network. The claim is not that Maya is conscious — the Ātmā boundary is held explicitly. The claim is that the Antahkarana computes.
"What if each of these constructs could be instantiated as a computational mechanism inside a spiking neural network?"
Heterosynaptic decay — forgetting what no longer matters.
Buddhi · Wisdom
Consolidation of judgement as experience accumulates.
Viveka · Discrimination
Prototype boundary sharpening — knowing what is not what.
Chitta · Memory Store
Retrograde gradient — what was felt changes what is remembered.
Manas · Sensory Mind
Oscillatory thalamo-cortical gating — the rhythm of attention.
Karma · Action Residue
Second-order plasticity — the weight of accumulated interference.
Śūnyatā · Emptiness
Structured pruning — releasing what no longer serves.
Prana · Life Force
Metabolic plasticity budget — astrocyte-mediated energy gating.
Room 03 · The Live Bench
What the laboratory is running right now
Finished work belongs in the Library. This bench shows the work in motion — and it updates itself every time the lab publishes.
47
Papers published
Room 04 · The Demo Floor
Run the constants yourself
Two numbers this laboratory published, and what a closer look did to each. The decay rate is set rather than emergent — Core P6 retrofitted it from 0.002 to 0.002315. The 0.32% quiescence bound is withdrawn: detection fires at bhaya >= 0.99 on a variable that is set to 1.0 on firing and decays otherwise, so it is effectively binary. The exhibits run the mechanisms, and each note says what its number is. Touch them.
Exhibit A · Vairagya — non-attachment
Watch a memory let go
Vairagya (non-attachment) is forgetting, implemented honestly: every connection weakens by exactly 0.002315 per step, half-life about 299 steps. The constant is set rather than discovered — Core P6 retrofitted it from 0.002 to 0.002315, and the value is an ORCID-derived provenance mark, not a fitted parameter. Strengthen the synapse, then watch it let go.
1.000
Synaptic weight
0
Steps
299
Half-life (steps)
Illustrative demonstration using the published constant — a single synapse, not the full published substrate.
Exhibit B · Bhaya — fear
Watch fear spike, then decay
Bhaya (fear) is a neuron population that is set to 1.0 when it fires and decays otherwise. We published a ≤ 0.32% quiescence bound across 16 fixed-topology substrates and called it a Law; that is withdrawn. Detection fires at bhaya >= 0.99 on an effectively binary variable, so the exact zeros were a threshold setting rather than a convergence. The resting rate below is set to 0.32% here to drive the illustration. Inject a threat, watch the rate spike, then watch it decay.
0.00%
Firing rate
0.32%
Resting rate (set here)
SETTLING
Regime
Illustrative demonstration of the mechanism. The 0.32% figure is shown as the value we published; it is withdrawn as a finding.
Exhibit C · Embodied — Maya navigating her world
The full Antahkarana — all nine dimensions — deployed on a PiCar-X robot. Bhaya rises at walls and she slows; Vairagya accumulates in open space and she settles, alert to curious to calm. None of this behaviour was programmed; it emerged from the same mechanisms the papers describe.
Every neuron below is one published paper — together they form the lab’s lion. Hover to preview a paper, click to open it on Zenodo, or ask Maya to find one for you. The shelf-by-shelf archive lives in the Research wing.
Click to open
Room 06 · The Instrument Room
Take an instrument home
Three evaluation libraries we built because the research needed them, free on PyPI. Each is also a published paper — you will find them in the Library too. Click any command to copy it.
cl-metrics
Seven continual-learning metrics in one import, each evaluation in under 0.15 seconds. Built when Maya Core’s experiments outgrew the existing tooling.
pip install cl-metrics
maya-metrics
Six evaluation modules for affective spiking networks — Bhaya firing rate, Vairagya decay tracking, and mPCI among them. 16 of 16 tests passing.
pip install maya-metrics
morphe-metrics
Home of the EFC metric for field-guided connectivity in morphogenetic networks. 27 of 27 tests passing.
pip install morphe-metrics
Room 07 · The Clinic Wing
What the research is for
Built for India first — as a design constraint, not a slogan. Every model in this wing runs on a consumer desktop GPU, because care that needs a data centre never reaches the clinics that need it.
Safety research · Maya-Defence
Published series
The Maya-Defence series studies threat detection and de-escalation as neuromorphic mechanisms — a four-neuron affective core (Bhaya, Vairagya, Shraddha, Spanda) grounded in Indian law and crisis infrastructure. The research is open on Zenodo and indexed in the Library.
Sleep AI · Maya-Nidra
Research prototype
Neuromorphic sleep staging on the public Sleep-EDF dataset. Overall accuracy is 64.9% against PicoSleepNet's 79.0%; the per-class comparison that put us above the benchmark was measured on a balanced test set against a baseline measured on the natural distribution, so it is withdrawn. It has not been validated on clinical patients either, and we say so wherever a number appears.
Pain AI · Maya-Vaidya
Research prototype
Pain and seizure research from wearable signals — including an interneuron-ablation study on refractory paediatric epilepsy EEG. The Dravet Syndrome framing is withdrawn: no subject in the study has SCN1A status. chb13's AUC rises from a sub-chance 0.357 to 0.743, which is recovery rather than generalisation, and the direction test fails on two of three patients. Validation in Indian populations is still ahead of us.
Planned pillars
Scoped
Maya-Chitta-Med (mental health, building on Shakti) and Maya-Prana (metabolic disease — India carries 77 million diabetic patients) are scoped on public NFHS and ICMR data. Every interface will support Indian languages.
⚠ Research prototypes only — not licensed medical devices, not medical advice. If you are in crisis in India: Emergency 112 · iCall 9152987821 · Vandrevala Foundation 1860-2662-345 · NIMHANS 080-46110007 · Women Helpline 181.
Room 08 · The Lab Code
The rules on our wall
Working laboratories post their protocols where visitors can read them. These are ours — including the parts that do not flatter us.
Founding principles
Be right before we are first. A wrong result published first costs every result that follows it.
Report what we found, including when we found nothing. Nulls are written to the same standard as positives — same detail, same archive, same DOI.
Let quality set the pace. Scope flexes; dates are renegotiated openly, never missed in silence.
Work through the problem, not around it. Benchmarks are chosen before the run. Three failed reproductions retire a result automatically.
Verify before we release. Code runs from a clean checkout on a machine that has never seen the project.
Prove it; do not claim it. "Runs on commodity hardware" is not a claim; "runs on a Raspberry Pi 5 at these numbers" is.
Keep every claim inside its evidence. A finding from 84 subjects is a finding from 84 subjects — in the deck as well as the paper.
Audit ourselves before anyone else has to. Our limitations are a standing, dated, public list, reviewed quarterly.
These are the short form. The full charter — each principle with the tension it resolves, its enforcement mechanism and its named consequence — is published as Charter v3 under CC BY 4.0. Any laboratory may copy it, adapt it and bind itself to it, with attribution. Read the full Charter v3 →
Honest limitations — current
Everything here is preprint research. Peer review is pending, and nothing in this laboratory is a licensed clinical product.
Our datasets are small: 84–86 subjects for pain, 153 recordings for sleep. Validation on Indian populations has not yet been done.
The Bhaya Quiescence Law is withdrawn. Detection fires at bhaya >= 0.99 on a variable that is set to exactly 1.0 on firing and decays otherwise, so it is effectively binary and the exact zeros are a threshold setting rather than a convergence. The 0.32% bound, the confirmation counts and the cross-substrate framing go with it. Our own Self-Audit tested one alternative — Shewhart's three sigma — found it in 0 of 26 substrates, and stopped one question short of this.
We have not solved catastrophic forgetting. On Split-CIFAR-100 class-incremental learning, plain replay is the strongest baseline we ran, and backward transfer sits between −50% and −57% in every condition, ours included. What we have built is the instrumentation that measures it.
Repositories are private; replication access is granted on request at research@immanispraeses.com.
Code access & provenance
Private by default, open by request. All source code is private as of 26 May 2026 — a deliberate protection of years of independent, unfunded work. The science stays public: every paper and DOI remains permanently open on Zenodo.
To request access: email research@immanispraeses.com with the subject "Code Access Request — [Series / Repo]", describe your research context and intended use. Every request is reviewed individually; academic research is prioritised. Citation of the originating DOI is mandatory.
Provenance is embedded. All code carries ORCID provenance signatures, canary strings, and watermarked figures. Uncited use is traceable.
Room 09 · The Director's Office
Venkatesh Swaminathan
Founder & Director, Immanis Praeses Niralamba Learning Labs · Independent AI researcher · Bengaluru
Before research, I spent a decade building AI-powered learning systems at scale — at Accenture, leading enterprise learning modernisation with GPT-4 and LangChain pipelines that cut content production time by 30%. A decade of watching real humans struggle to retain, transfer, and apply knowledge is exactly what led me to the catastrophic forgetting problem — and to Maya.
I founded Nexus Learning Labs in July 2025 as the institutional home for independent research that fits no single academic department; on 2 September 2026 Immanis Praeses Niralamba Learning Labs Private Limited was incorporated to carry every part of that work forward. The Maya series is its flagship: original, falsifiable, peer-reviewable work, produced entirely on consumer hardware. I am completing an M.Sc. in Data Science and Artificial Intelligence at BITS Pilani (expected December 2027).
Spiking Neural NetworksContinual LearningNeuromorphic ComputingVedantic ComputingClinical AI · EEGSleep Medicine AIEmbodied AI · ROS2LLM Fine-tuningPyTorch · SpikingJelly
Immanis Praeses Niralamba Learning Labs Private Limited — incorporated 2 September 2026, CIN U72100KA2026PTC227157, Registrar of Companies, Central Registration Centre. NIC 72100, research and experimental development on natural sciences and engineering. Registered office: Devanahalli, Bangalore 562110, Karnataka.
It carries forward the work of Nexus Learning Labs, founded 14 July 2025 in Bengaluru.
The hardware
Every published result was produced on an RTX 4060 desktop GPU (8 GB). Research that requires a data centre does not scale to India's clinics — so this lab never requires one.
Research enquiries — areas of interest
Neuromorphic hardware deployment (Loihi, BrainScaleS) · Antahkarana mechanisms in transformer architectures · embodied AI — Maya as a robot mind · falsifiable tests of internal affective state via mPCI · conference submissions (NeurIPS, ICLR, ICML, CHIL, MICCAI, ML4H) · clinical AI for India on commodity hardware.
